XLV Mexican Conference on Biomedical Engineering: Proceedings of CNIB 2022, 6–8 October, Puerto Vallarta, México (IFMBE Proceedings, 86) 3031182553, 9783031182556
132 88 141MB
English Pages [902]
Preface
Distinguished Lecturer Message
Organization
Executive Committee
Conference Chair
Program Chair
Scientific Committee
Program Chair
Program Co-chair
Session Coordinators
Committee Assistant
Awards Committee
Student Competition
SOMIB Awards
Organizing Committee
Organizing Chair
International Guests
Professional Guests
Universities Guests
Workshops
Exposition Chair
Students Coordinator
Education Committee
Chair
Co-chair
Industry Committee
Chair
Innovation Committee
Chair
Logistics Committee
Chair
Graphic Design
Administrative Support
Public Relations Coordinator
Referees
Sponsoring Institutions
Contents
Artificial Intelligence and Data Science
Device for the Fall Detection in Older Adults Through Neural Networks
1 Introduction
2 Methodology
2.1 Selection of Electronic Components
2.2 General Procedure
2.3 Neural Network Architecture for Posture Detection
2.4 Device Design
2.5 Participants
3 Results
3.1 Sample Characterization
3.2 Device Fall Detection
3.3 Neural Network Operation
4 Discussion
5 Conclusion
References
Breast Cancer Detection Algorithm Using Ensemble Learning
1 Introduction
2 Methods
2.1 Database
2.2 Naïve Bayes Algorithm
2.3 Artificial Neural Network
2.4 Support-Vector Machine
2.5 Composite Algorithm
3 Results
4 Discussion
5 Conclusions
References
Graph Analysis of Functional Connectivity Rs-FMRI in Healthy and Epileptic Brain Using Visibility Algorithm
1 Introduction
2 Methodology
3 Results
3.1 Case Study
3.2 Seed Regions of Interest
4 Discussion
5 Conclusions
References
Imagined Speech Recognition in a Subject Independent Approach Using a Prototypical Network
1 Introduction
2 Methodology
2.1 KaraOne Dataset
2.2 Dual Tree Complex Wavelet Transform (DTCWT)
2.3 Prototypical Networks
3 Experimental Settings and Results
4 Conclusions
References
Design and Comparison of Artificial Intelligent Algorithms for Breast Cancer Classification
1 Introduction
2 Methodology
2.1 Database
2.2 Artificial Neural Network
2.3 Naive Bayes Algorithm
2.4 Support Vector Machine
3 Results and Discussions
4 Conclusion
References
Electrophysiological Signals Simulation with Machine Learning
1 Introduction
2 Materials and Methods
2.1 Data Acquisition and Processing
2.2 Data Reshape for Machine Learning Model
2.3 Machine Learning Model Design, Training and Testing
2.4 Model Utility
3 Results
4 Discussions
5 Conclusions
References
Quantification of a Lip and Palate Clefts Classification
1 Introduction
2 Methodology
3 Results
3.1 Level 1. Unilateral Primary and Secondary Clefts
3.2 Level 2. Bilateral Primary Clefts
3.3 Level 3. Labial-Palatal Cleft (Primary and Secondary Palates)
3.4 Validation of the Relevance Factor (ρ)
4 Conclusions
References
Artificial Intelligence Applied to Breast Cancer Classification
1 Introduction
1.1 Mammography as a Method of Breast Cancer Screening
1.2 Convolutional Neural Networks (CNN)
2 Materials and Methods
2.1 Digital Image Processing
2.2 Development Environment
2.3 Proposed Algorithm
2.4 Training Models
3 Results
4 Conclusions
References
Computational Chemistry as an Educational Tool in Health Sciences
1 Introduction
1.1 Importance of the Study
2 Methodology
3 Results
3.1 Structure Analysis Results
3.2 Pedagogical Results
4 Conclusions
4.1 Discussion
4.2 Author’s Comment
References
A Gene-Community Overview of Transcriptional Dynamics During Neurodevelopment
1 Introduction
2 Methods
2.1 Database
2.2 Community Neurodevelopment
3 Results and Discussion
3.1 Dataset
3.2 Gene Community Expression Dynamics
4 Conclusions
References
CNNs for ISCI Stage Recognition on Video Sequences
1 Introduction
1.1 Development of ICSI
1.2 Convolutional Neural Networks
2 Methods
3 Results
4 Discussion
5 Conclusions
References
Stacked Spatial and Temporal Deep Learning Methods for Identification of Parkinson's Disease Using Gait Signals
1 Introduction
2 Materials and Methods
2.1 Databases Description
2.2 Signal Processing: Gait Cycles
2.3 CLDNN
2.4 Classification
3 Results and Discussion
4 Conclusions
References
Diversity of Genotyping Chlamydia Trachomatis Serovars in Urogenital Samples from Mexican Patients: A Molecular and Bioinformatic Characterization
1 Introduction
2 Methodology
2.1 Ethics Statement
2.2 Research Population
2.3 Samples and DNA Collection
2.4 Sampling
2.5 Multiplex PCR Assay for the Diagnosis of STI
2.6 OmpA-Targeted PCR
2.7 Identification of CT Serovars
2.8 Statistical and Bioinformatic Analysis
3 Results
3.1 PCR Results
3.2 Prevalence and Coinfection of CT Serovars
3.3 Phylogenetic Analysis
4 Analysis
5 Conclusions
References
Detection of Breast Cancer in Mammography Using Deep Learning Models
1 Introduction
2 Materials and Methods
2.1 Dataset
2.2 Model and Hyper-parameter
3 Results and Discussion
4 Conclusion
References
Modeling and Simulation of Biological Systems
A Comparative Study on the Interaction of an Ototoxic and an Otoprotective with the Megalin Receptor Associated with Hearing Loss
1 Introduction
1.1 Aminoglycoside Antibiotics
1.2 The Megalin
2 Material and Methods
2.1 Preparation of the Structures
2.2 Molecular Docking Studies
2.3 Binding Free Energy (ΔGb) Determinations
3 Results
3.1 Docking Assays
3.2 Electrostatic Contributions to the ΔGb
4 Conclusions
References
Collagen/Plasma-Polymerized Pyrrole Interaction: Molecular Docking and Binding Energy Calculations
1 Introduction
1.1 The Role of Collagen in Heart Tissue
1.2 Plasma-Synthesized Polypyrrole (PPPy)
2 Metodology
2.1 Preparation of Structures
2.2 Molecular Docking Studies
2.3 Binding Free Energy Calculations
3 Results and Discussion
4 Conclusions
References
Implanted Pediatric Patient Early Audiometry
1 Introduction
2 Methodology
2.1 Population
2.2 Tests
2.3 Analysis
3 Results
4 Discussion
5 Conclusion
References
Thermal Performance of a Triple Slot Antenna Considering Temperature Dependence of Thermal and Electrical Conductivity, Blood Perfusion and Tissue Metabolism
1 Introduction
2 Methodology
2.1 Micro-coaxial Triple-Slot Antenna
2.2 Antenna Modeling: Finite Element Method
2.3 Parameters Evaluation
3 Results and Discussion
3.1 SWR and Areas of Ablated Tissue
3.2 Temperature Profiles
4 Conclusions
References
Modeling of the Interaction of Plasma-Polymerized Pyrrole with Immunoglobulin M (IgM) by Biocomputational Tools
1 Introduction
2 Methodology
2.1 Molecular Docking Studies
2.2 Binding Energy Calculations (∆Gb)
2.3 Molecular Dynamics
3 Results
4 Conclusion
References
Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis
1 Introduction
2 Nitrofuran Antibiotics
3 Computational Chemistry
4 Materials and Methods
5 Results and Discussion
6 Conclusion
References
Simulating the Ca2+-cAMP Crosstalk and Its Role in Pancreatic Cells
1 Introduction
2 Methods
3 Results
4 Discussion
5 Conclusions
References
Simulating the Loss of -cell Mass in a Human Pancreatic Islet: Structural and Functional Implications
1 Introduction
2 Methods
2.1 Reconstruction of Pancreatic Islets
2.2 Loss of -cells
2.3 Islet Connectivity Networks
2.4 Modelling the Electrical Behavior of the Human -cell
2.5 Computational Aspects
3 Results
4 Conclusions
References
Role of Endogenous Ca2+ Buffering and the Readily Releasable Pool on Fast Secretion in Auditory Inner Hair Cells
1 Introduction
2 Methods
3 Results and Discussion
4 Conclusions
References
Effects of Blood Flow on Insulin Concentration: A Modelling Study
1 Introduction
2 Methods
2.1 Bi-dimensional Slice of a Human Pancreatic Islet
2.2 Modeling Blood Flow in the Pancreatic Islet
2.3 Insulin Secretion from -cells and Diffusion Throughout the Islet Vasculature
2.4 Initial and Boundary Conditions
2.5 Computational Aspects
3 Results
4 Conclusions
References
Non-invasive Hypoglycemia Regulatory Patch with Glucagon Administration
1 Introduction
1.1 Background
1.2 Hypoglycemia in Diabetic and Non-diabetic Patients
2 Materials and Methods
2.1 System Description
2.2 Model Design and Flow Simulation
2.3 Passive Facilitated Diffusion
2.4 FICk’s Law
2.5 Physiology and Characteristics of the Skin
2.6 Glycerin
3 Results
3.1 Flow Analysis
3.2 Transdermal Patch
4 Conclusion
References
The Enzymatic Core of Snakes
1 Introduction
2 Methodology
2.1 RNA-seq Reads Acquisition
2.2 Read Quality Assessment and de novo Assembly
2.3 Identification of Transcript Sequences and Putative Coding for Enzymes
2.4 Enzymatic Intersection Analysis
3 Results and Discussion
4 Conclusions
References
Structural Analysis for Enzymatic Homology Determination in Terpene Cyclases
1 Introduction
2 Materials and Methods
3 Results
4 Discussion
5 Conclusions
References
Effect of Thermal Dependence of Tissue Properties on the Antenna Performance: A 3D Parametric Model
1 Introduction
2 Methodology
2.1 3D Antenna Modeling
2.2 Electromagnetic and Thermal Modeling
2.3 Parametric Study
3 Results and Discussions
4 Conclusion
References
Hepatic Cell Radial Flow Bioreactor Parametrization and Characterization as an Alternative Therapy to Liver Failure
1 Introduction
2 Methodology
2.1 Bioreactor
2.2 Model Development
3 Results and Discussion
3.1 Model Validation
3.2 Dynamic Analysis of the Bioreactor
4 Conclusions
References
Medical Physics and Nuclear Medicine
Gamma Radiation Detection Simulation System
1 Introduction
2 Materials and Methodology
2.1 Gamma Radiation Distance Detector Design
2.2 JAVA Interface Design
2.3 Validation
3 Results
4 Discussions
5 Conclusion
References
Development of an Alternative Radiochromic Film Digitizer for Clinical Dosimetry
1 Introduction
2 Materials and Methods
2.1 Experimental Densitometry System: Layout
2.2 Radiochromic Film: Sampling, Irradiation and Calibration
2.3 Flatbed Scanner
2.4 Data Analysis: Noise, Fitting Uncertainty and Total Uncertainty
3 Results and Discussion
4 Conclusion
References
Processing of Biomedical Signals
Decoding Imagined Speech of Daily Use Words from EEG Signals Using Binary Classification
1 Introduction
2 Materials and Methods
2.1 Acquisition
2.2 Preprocessing
2.3 Feature Extraction
2.4 Classification
3 Results
4 Conclusion
References
Nonlinearity of Electrohysterographic Signals is Diminished in Active Preterm Labor
1 Introduction
2 Methods
3 Results
4 Discussion
5 Conclusion
References
Trend of Concentration of Men and Women Elucidated by Analysis of EEG Signals Recorded During a Fast Game
1 Introduction
2 Methodology
2.1 Participants
2.2 Experiment Protocol
2.3 Pre-processing Step: Filtering
2.4 Characterization and Analysis of the Filtered Signals
3 Results and Discussion
3.1 Time Results
3.2 Frequency Results
4 Conclusions
References
Effects on Body Posture and Gait Caused by Different Weights in the Backpack of University Students
1 Introduction
2 Methods
2.1 Participants
2.2 Protocol
2.3 Video Analysis
2.4 Data Analysis
3 Results
4 Discussion
5 Conclusion
References
Multiscale-Multifractal Assessment of Heart Rate Variability in Shift Workers by Detrended Fluctuation Analysis
1 Introduction
2 Materials and Methods
2.1 Database
2.2 Multifractal-Multiscale Detrended Fluctuation Analysis
2.3 Statistical Analysis
3 Results and Discussion
4 Conclusions
References
EEG Connectivity Analysis in a Motor Imagery Task
1 Introduction
2 Materials and Methods
2.1 Database
2.2 Experimental Paradigm
2.3 Data Selection
2.4 Preprocessing
2.5 Adjacency Matrices Estimation
2.6 Graph Analysis
2.7 Statistical Analysis
3 Results and Discussion
4 Conclusion
References
Brain Mapping: Location of the Words Through EEG
1 Introduction
2 Methods
2.1 Participants
2.2 Stimuli
2.3 Segmentation and Relation of Words
2.4 EEG Recording
2.5 Signal Pre-processing and Artifact Removal
2.6 Signal Processing and Analysis
3 Results and Discussion
4 Conclusion
References
Processing of Biomedical Images
Artifacts Generated by the 3D Rotation of a Freely-Swimming Human Sperm in the Measurement of Intracellular Ca2+
1 Introduction
2 Materials and Methods
2.1 Sperm Preparations
2.2 Experimental Setup
2.3 Image Processing
3 Results
4 Conclusions
References
Morphological Temporal Analysis in Subjects with Alzheimer's Disease by Brain Graph Descriptors
1 Introduction
2 Methodology
2.1 T1–MRI Database
2.2 Automatic Brain Segmentation and Morphological Descriptors
2.3 Morphological Brain Networks
2.4 Graph Descriptors and Statistical Analysis
3 Results and Discussion
4 Conclusions
References
PET Image Reconstruction Using a GRU-Convolutional Network
1 Introduction
2 Theory
2.1 The Sinogram
2.2 Recurrent Networks
3 Methods
3.1 Proposed Method
3.2 Phantom Simulation
4 Results
5 Conclusions
References
Characterization of COVID-19 Diseased Lung Tissue Based on Texture Features
1 Introduction
2 Material and Methods
2.1 Medical Images Dataset
2.2 Lung Tissue Segmentation
2.3 Definition of Pulmonary Regions
2.4 GLCM and GLRLM Matrices Definition
2.5 Textural Features Estimation
2.6 Descriptive and Statistic Analysis
3 Results and Discussion
4 Conclusions
References
Glioblastoma Classification in Hyperspectral Images by Reflectance Calibration with Normalization Correction and Nonlinear Unmixing
1 Introduction
2 Method
2.1 Proposed Calibration
2.2 Nonlinear Unmixing
2.3 Classification Results
3 Results
3.1 Calibration Results
3.2 Classification Results
4 Conclusions
References
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein with the Phasor-FLIM Approach Signatures of Laurdan
1 Introduction
2 Methods
2.1 Materials
2.2 Cell Culture
2.3 Fluorescence Lifetime Imaging Microscopy
2.4 The Phasor Approach for Lifetime Data
2.5 Phasor Distributions Analyzed by the Three-Cursor Approach
2.6 Statistics
3 Results
3.1 Expression of mHtt with a PolyQ53 Increases Membrane Fluidity in Total Membranes
3.2 Expression of mHtt PolyQ53 Does not Increases DR Processes in Total Membranes
4 Conclusion
References
A Method for Automatic Monoplane Angiography Segmentation
1 Introduction
2 Materials and Methods
3 Results
4 Discussion
References
Lung Segmentation Algorithm and SVM Classification of COVID-19 in CT Images
1 Introduction
2 Materials and Methods
2.1 Lung Segmentation
2.2 Classification
3 Experimentation and Results
4 Conclusions
References
IOT in Health and Bioinstrumentation
Disinfection Method Based on UV-C Light Using the Internet of Things for Cleaning Hospital Areas (COVID-19)
1 Introduction
1.1 Paper Distribution
2 Materials and Methods
2.1 Raspberry Pi
2.2 Passive Infrared Sensor (PIR)
2.3 Smart Switch
2.4 UV-C Mobile Platform
2.5 Mobile Interface
3 Results
4 Conclusion
References
Development of Alpha Prototype of Handheld Device for Meibography
1 Introduction
2 Justification and Objectives
3 Methodology
3.1 Lens System (macro)
3.2 Infrared LED Illumination System
3.3 CPU Selection
3.4 Evaluation of Device Power Consumption
3.5 Evaluation of Image Quality
4 Results
4.1 Lens System (macro)
4.2 Illumination System
4.3 Graphic User Interface (GUI)
4.4 Evaluation of Device Power Consumption
4.5 Evaluation of Image Quality
5 Conclusions
A Bill of Materials
References
Prototype of a Pulse Oximeter Based on an Open-Source Platform with Wireless Design and Cloud Service
1 Introduction
1.1 Problem Statement
2 Related Work
2.1 Project Proposal
3 Methodology
3.1 Pulse Oximeter Design
3.2 Prototype Implementation
3.3 Prototype Evaluation
4 Results
4.1 Pulse Oximeter Prototype
4.2 Prototype Validation
4.3 Platform for Data Storage, Visualization and Analysis
5 Discussion
6 Conclusion
References
Wearable System for Measuring Vertical Ground Reaction Forces During the Gait Cycle
1 Introduction
2 Methodology
2.1 Design of Instrumented Insole
2.2 Manufacturing of the Instrumented Insole
2.3 Data Acquisition System (DAS)
3 Results
3.1 Characteristic Curve of the FSR Sensors
3.2 Vertical Ground Reaction Forces
4 Discussion
5 Conclusion
References
Design of a Pulse Oximeter with Altitude Measurement Bluetooth Communication and Android Application
1 Introduction
1.1 Post-COVID Sequels and Syndrome
1.2 Objective
2 Oximetry
2.1 Photoplethysmography (PPG)
2.2 Optimal Saturation Levels
3 Methods and Materials
3.1 Pulsatile Oximetry by the Reflective Method
3.2 Measurement of SpO2
3.3 Calibration
4 Design and Manufacture
4.1 Software Development
5 App in App Inventor
6 Results and Discussion
6.1 Oximeter Implementation and Prototype
7 Conclusions
References
Braille System Learning Introductory Device
1 Introduction
2 Materials and Methods
2.1 Programming
2.2 Electronic
2.3 Design
2.4 Testing
3 Results
4 Discussion
5 Conclusions
References
Signal to Noise Ratio and Current Consumption in LED-LED Photoplethysmography
1 Introduction
2 Materials and Methods
2.1 The LED-LED PPG System
2.2 Signal-to-Noise Ratio and Current Consumption
3 Results
4 Conclusions
References
Prototype for the Monitoring of Soda Lime in Anesthesia Machines Using Wi-Fi Alarm
1 Introduction
2 Materials and Methods
2.1 Programming and Design of the Web Page
2.2 Programming and Installation of the TCS-230 and MQ-7 Sensor Libraries
2.3 Programming the ESP8266-NodeMCU
2.4 Final Assembly of the Prototype
3 Results
4 Conclusion
References
Design and Implementation of a Smartphone-Based Digital Phonocardiograph with Wireless Transmission Capabilities
1 Introduction
2 Methodology
2.1 Hardware Design
2.2 Software Design
2.3 Pilot Tests of the Digital PCG
3 Results and Discussion
3.1 Hardware
3.2 Software
3.3 Device Tests
4 Conclusion
References
System for Detection of Neonatal Apnea
1 Introduction
2 Materials and Methods
2.1 Circuits
2.2 Program
3 Results and Discussion
4 Conclusions
References
The Road to Making “Exergames” More Widely Available
1 Introduction
2 Materials and Methods
2.1 Methodology
3 Results
4 Discussion
5 Conclusions
References
Design and Construction of Capacitive Coupling Electrostimulator to Induce Bone Tissue Regeneration
1 Introduction
2 Materials and Methods
2.1 Function Generator
2.2 Signal Conditioning
2.3 Current Circuit
2.4 Obtaining Current Values
3 Discussion
3.1 Test Protocol
4 Conclusions
References
Biosensors
Paper-Based Microanalytical Device for Colorimetric Detection of Stress in Human Saliva Sample
1 Introduction
2 Materials and Methods
2.1 Reagents
2.2 μ-PADs Design and Fabrication
2.3 Characterization of the Nanostructured Surface
3 Results and Discussion
3.1 Design and Colorimetric Detection
3.2 Characterization of the Nanostructured Surface
4 Conclusions
References
Construction of an Electrochemical Nanogenosensor for K-RAS Oncogene Detection
1 Introduction
2 Materials and Methods
2.1 Reagents
2.2 Modification of the Genosensor
2.3 Hybridization Assay and Selectivity of the Genosensor
2.4 Analytical Measurements
2.5 Atomic Force Microscopy Analysis
3 Results
4 Conclusion
References
Computational Study of a:SiC:H Thin Films Deposited on Interdigitated Microelectrodes Using Electrical Impedance Spectroscopy
1 Introduction
2 Design and Simulations
2.1 Building of the Model in COMSOL
2.2 Equivalent Circuit Model
3 Simulation Results and Discussion
3.1 Conductivity Effects
3.2 Thickness Effects
3.3 Surface Topology Effects
4 Conclusions
References
Development of Non-enzymatic Sensor for Uric Acid Detection Based on Gold Nanoparticles Electrodeposited on Laser-Induced Graphene Electrodes
1 Introduction
2 Materials and Methods
2.1 Reagents
2.2 Materials and Apparatus
2.3 Electrode Preparation
2.4 Electrochemical Characterization
2.5 UA Detection
3 Results
3.1 Characterization
3.2 Amperometry Detection
4 Conclusion
References
Unmodified Screen-Printed Electrodes-Based Sensor for Electrochemical Detection of Bisphenol A
1 Introduction
2 Materials and Methods
2.1 Chemicals
2.2 Screen-Printed Electrodes
2.3 Instrumentation
3 Results
3.1 Linear-Sweep Voltammetry Study
3.2 Cyclic Voltammetry Study
3.3 Square-Wave Voltammetry Study
3.4 Comparison Between Electrochemical Methods
4 Conclusions
References
Molecularly Imprinted Polymer Paper-Based Biosensor for Wireless Measurement of Sweat Glucose
1 Introduction
2 Materials and Methods
2.1 Chemicals and Materials
2.2 Molecularly Imprinted Polymer on Paper Strips
2.3 Paper-Based Glucose Sensor Fabrication
2.4 Measurement Device
3 Results
3.1 Sensor Characterization
3.2 Sensor Calibration
4 Conclusions
References
Bioimpedance and Micro-nanotechnologies
The Predictive Capacity of Bioelectrical Impedance Parameters at Frequencies of 5, 20, 50, 100, and 200 kHz to Identify Vector-Associated Febrile Syndromes in the Emergency Room of the Hospital Civil de Guadalajara
1 Introduction
2 Material and Methods
2.1 Study Population and Sample
2.2 Bioelectrical Impedance Analysis, Weight, and Height
2.3 Statistical Analysis
3 Results and Discussion
4 Conclusions
References
Application of Palladium Nanoparticles as a Contrast Agent for Electrical Bioimpedance Measurements on Biological Tissue
1 Introduction
2 Materials and Methods
2.1 Biological Tissues
2.2 Nanoparticles and Solutions
2.3 Equipment and Systems
2.4 Procedure
2.5 Statistical Analysis
3 Results
4 Conclusions
References
Bladder Volume Monitoring by Electrical Bioimpedance Technique. Calibration Mathematical Models
1 Introduccion
2 Materials and Methods
2.1 Bioimpedance Equipment
2.2 Ultrasound Equipment
2.3 Participants
2.4 Resistor-Capacitor Model
2.5 Data Acquisition
2.6 Data Analysis and Statistics
3 Analysis and Results
4 Conclusions
References
Mechanical Characterization of Patellar Tendon Strain by Electrical Impedance
1 Introduction
2 Materials and Methods
2.1 Bioimpedance Device
2.2 Ultrasound Device
2.3 Participants
2.4 Procedure
2.5 Statistical Analysis
3 Results
3.1 Difference Obtained Between States
3.2 Results of the Wilcoxon Signed-Rank Test
3.3 Spearman Correlation
4 Conclusions
References
Nanoparticles for Glioblastoma Treatment
1 Introduction
2 Methods
3 Results and Discussion
4 Conclusion
References
Bioimpedance Spectra in Final RT-PCR Products: A Sensitivity Threshold Analysis
1 Introduction
2 Methodology
2.1 Experimental Design
2.2 DNA Extraction and RT-PCR Procedure
2.3 Bioimpedance Measurements
3 Results
3.1 RT-PCR Products
3.2 Bioimpedance Spectra
4 Discussion
5 Conclusions
References
Biomaterials, Molecular, Cellular and Tissue Engineering
Design and Fabrication of a Radial Flow Bioreactor to Decellularize Muscular Arteries
1 Introduction
2 Methodology
2.1 System Description
2.2 Model Design and Flow Simulation
2.3 3D Printing
3 Results
3.1 Flow Analysis
3.2 Printed Model
4 Conclusions
References
Iron Carbide@Iron Oxide Core-Shell Nanoparticles Functionalization with L-Arginine Amphiphillic Bioconjugate
1 Introduction
1.1 Iron Nanoparticles
1.2 The Role of Surfactants in Metal Nanoparticles
1.3 Nanoparticle Bioconjugation
2 Materials and Methods
2.1 Oleic Acid Modified with L-Arginine
2.2 Surfactant Purification
2.3 Synthesis of Iron Nanoparticles
3 Results and Discussion
3.1 Prediction of Chemical Bonds of OAR Bioconjugation by Infrared
3.2 OAR Raman Spectroscopy
3.3 Morphological Analysis of OAR
3.4 OAR Size Analysis
3.5 ICIONPs Interaction with OAR
4 Conclusions
References
Evaluation of Hemolytic Behavior and Bioactive Properties of Natural Wollastonite and Synthetic Hydroxyapatites Produced by Two Sol-Gel Routes
1 Introduction
2 Materials and Methods
2.1 Wollastonite
2.2 Sol-Gel Synthesis
2.3 Sintering of Materials
2.4 In Vitro Hemolysis Assay
2.5 In Vitro Bioactivity Studies of the Sintered Bioceramics
3 Results
3.1 Bioactive Properties
3.2 Particle Size Distribution
3.3 In Vitro Hemolytic Properties
4 Conclusion
References
Microstructure and Mechanical Properties of Hydroxyapatite Nanofibers Synthesized Through the Microwave-Assisted Hydrothermal Method for Biomedical Applications
1 Introduction
2 Methodology
2.1 Synthesis of HAp by HAHM
2.2 Characterization of HAp Nanofibers
2.3 Preparation of the Composite Materials of HAp
2.4 Mechanical Compression of Composite Materials
3 Results
3.1 X-Ray Analysis
3.2 Electron Microscopy
3.3 High-Resolution Transmission Electron Microscopy
3.4 Mechanical Behavior Study
3.5 X-Ray Fluorescence Analysis
4 Analysis
5 Conclusions
References
Biological Pacemakers Obtained Through Cellular Differentiation for the Restoration of Sinoatrial Node Function. A Systematic Review
1 Introduction
2 Methods
3 Results
4 Conclusion
References
Evaluation of the Formation of an Ionic-Complementary Self-assembling Peptide Hydrogel for the Three-Dimensional Culture of Mammalian Cells in Vitro
1 Introduction
2 Materials and Methods
2.1 Formation of Peptide Hydrogel
2.2 Cell Culture
2.3 Three-Dimensional (3D) Cell Culture
2.4 Cell Viability
3 Results
3.1 Gelation of the F9 Peptide
3.2 Cell Viability
4 Conclusions
References
3D Bioprinting of Hydrogels Using Hydrophobic Sands and Calcium Chloride as Structural Support
1 Introduction
2 Materials and Methods
2.1 Materials
2.2 Hydrogel Preparation and Characterization
2.3 3D Bioprinting
2.4 Kinetic Sand Evaluation and Cross-Linking Procedure
3 Results
3.1 Hydrogel Properties and 3D Bioprinting
3.2 Rheological Properties of Kinetic Sand-CaCl2 Support Baths
3.3 Evaluation of Crosslinking Mechanism
4 Conclusions
References
Rehabilitation, Biomechanics and Biorobotics
Evaluation of Muscle Activity and Predisposition to Pain in Male Volleyball Players
1 Introduction
2 Materials and Methods
3 Results and Discussion
3.1 Abduction Gesture
3.2 Flexion Gesture
4 Conclusion
References
Hands-Free Walking Stick
1 Introduction
2 Analysis of the Weak Leg Problem
2.1 Effects on the Life of a Person with a Weak Leg
2.2 Load Analysis in One Step
3 Materials and Methods
3.1 Concept of the Device
3.2 Sections of the Device
4 Results and Discussion
5 Conclusion
References
Elbow Torque Estimation for Human-Robot Interaction Control
1 Introduction
2 Materials and Methods
2.1 Torque Estimation
2.2 Impedance Control with Bounded Actions
2.3 Human-Robot Interaction Task
3 Results
3.1 Experimental Torque Estimation
3.2 Performance of the Impedance Controller
4 Conclusions
References
Prototype of an Active Partial Hand Prosthesis for a Person with Symbrachydactyly
1 Introduction
2 Materials and Methods
2.1 Methodology
2.2 Customized Socket
2.3 Prosthetic Fingers Design
3 Results
4 Conclusions
References
Electronic System to Determine Proximal and Medial Phalanges Strength in a Hand Exoskeleton Robot
1 Introduction
1.1 Materials and Methods
1.2 Experiments
1.3 Results
1.4 Conclusions
References
Clinical Engineering and Education
Innovation and Control of Health Technology Management Procedures Applying Six Sigma Methodology
1 Introduction
2 Methodology
2.1 Analysis and Innovation of Health Technology Management Procedures
3 Results
3.1 Analysis of the Procedure of Attention to Work Orders
3.2 Analysis of the Procedure of Inventory Supplies Control
3.3 Innovation of the Procedure of Attention to Work Orders
3.4 Innovation of the Procedure of Supply Inventory Control
3.5 Health Technology Management Activities
4 Conclusions
References
Exploratory Data Analysis for Preventive and Corrective Maintenance for Medical Equipment in a General Hospital from the Health Institute of the State of Mexico
1 Introduction
2 Methodology
2.1 Medical Areas Selection
2.2 Data Collection
2.3 Data Processing
3 Results
3.1 Data Collection
3.2 Data Processing
4 Conclusions
References
Obsolescence Assessment Approach: Case of Mechanical Ventilators Under the Covid-19 Environment
1 Introduction
2 Methods
2.1 Risk Factors
2.2 Multicriteria Method
3 Results
4 Discussion
5 Conclusions
References
Application of the Quality Function Deployment Methodology for Quality Analysis in the Clinical Laboratory
1 Introduction
2 Methods
2.1 Study Design
2.2 Data Analysis
3 Results
3.1 Analysis of SERVQUAL Questionnaires
3.2 Identification of Patients' Requirements
3.3 Engineering Goals
3.4 Construction of the Planning Matrix
3.5 Engineering Features
3.6 Construction of the Design Matrix
4 Conclusion
References
Strategies Employed in the Reconfiguration of Healthcare Facilities During COVID-19 in OECD Countries
1 Introduction
2 Methods
2.1 Document Search
2.2 Reconfiguration Strategies
3 Results
3.1 General Strategies for Hospital Care of Patients with COVID-19
3.2 Healthcare Initiatives Focused on Medical Devices for COVID-19 Patients
3.3 Equipment Guidelines Provided by PAHO for Latin American Countries
4 Conclusions
References
CO2 Levels in the Naso-Buccal Area Due to the Use of Different Face Masks in Different Ventilation Conditions
1 Introduction
2 Materials and Methods
3 Results
4 Conclusions
References
Creation of a Needs Detection System for High Technology Medical Equipment or Medical Equipment that Require an Infrastructure Specification for the Mexico City’s Secretariat of Health
1 Introduction
2 Materials and Methods
2.1 Defining and Identifying the Needs of HTME and MEIS
2.2 Regulations and Evaluation Criteria
2.3 Development of the SIDENE
3 Results
3.1 General Information
3.2 Mem
3.3 Performance
3.4 Human Resource
3.5 Special Requirements
3.6 Services and Infrastructure
4 Conclusions
References
Use of Audiovisual Strategies as a Complementary Resource for Practical Courses in Biomedical Engineering
1 Introduction
2 Material and Methods
2.1 Development and Implementation of Audiovisual Strategies
2.2 Students Survey Satisfaction
3 Results and Discussion
3.1 Audiovisual Resources
3.2 Student Survey Satisfaction
3.3 Pros and Cons of both Modalities
4 Conclusions
References
Innovation of Technologies for Health
Mechanical Design and Additive Manufacturing for a Low-Cost Hybrid Dermatoscope
1 Introduction
1.1 Paper Distribution
2 Materials and Methods
2.1 Design Proposal
2.2 Static Analysis of the BiLux-TM Model
2.3 Construction and Instrumentation
3 Results
3.1 Construction and Instrumentation
3.2 Evaluation
4 Materials and Costs Analysis
5 Conclusion
References
Author Index
File loading please wait...
Citation preview
IFMBE Proceedings Citlalli Jessica Trujillo-Romero · Rafael Gonzalez-Landaeta · Christian Chapa-González · Guadalupe Dorantes-Méndez · Dora-Luz Flores · J. J. Agustin Flores Cuautle · Martha R. Ortiz-Posadas · Ricardo A. Salido Ruiz · Esmeralda Zuñiga-Aguilar Editors
Volume 86
XLV Mexican Conference on Biomedical Engineering Proceedings of CNIB 2022, 6–8 October, Puerto Vallarta, México
IFMBE Proceedings Volume 86
Series Editor Ratko Magjarevic, Faculty of Electrical Engineering and Computing, ZESOI, University of Zagreb, Zagreb, Croatia Associate Editors Piotr Ładyżyński, Warsaw, Poland Fatimah Ibrahim, Department of Biomedical Engineering, Faculty of Engineering, Universiti Malaya, Kuala Lumpur, Malaysia Igor Lackovic, Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia Emilio Sacristan Rock, Mexico DF, Mexico
The IFMBE Proceedings Book Series is an official publication of the International Federation for Medical and Biological Engineering (IFMBE). The series gathers the proceedings of various international conferences, which are either organized or endorsed by the Federation. Books published in this series report on cutting-edge findings and provide an informative survey on the most challenging topics and advances in the fields of medicine, biology, clinical engineering, and biophysics. The series aims at disseminating high-quality scientific information, encouraging both basic and applied research, and promoting world-wide collaboration between researchers and practitioners in the field of Medical and Biological Engineering. Topics include, but are not limited to: • • • • • •
Diagnostic Imaging, Image Processing, Biomedical Signal Processing Modeling and Simulation, Biomechanics Biomaterials, Cellular and Tissue Engineering Information and Communication in Medicine, Telemedicine and e-Health Instrumentation and Clinical Engineering Surgery, Minimal Invasive Interventions, Endoscopy and Image Guided Therapy • Audiology, Ophthalmology, Emergency and Dental Medicine Applications • Radiology, Radiation Oncology and Biological Effects of Radiation IFMBE proceedings are indexed by SCOPUS, EI Compendex, Japanese Science and Technology Agency (JST), SCImago. Proposals can be submitted by contacting the Springer responsible editor shown on the series webpage (see “Contacts”), or by getting in touch with the series editor Ratko Magjarevic.
More information about this series at https://link.springer.com/bookseries/7403
Citlalli Jessica Trujillo-Romero Rafael Gonzalez-Landaeta Christian Chapa-González Guadalupe Dorantes-Méndez Dora-Luz Flores J. J. Agustin Flores Cuautle Martha R. Ortiz-Posadas Ricardo A. Salido Ruiz Esmeralda Zuñiga-Aguilar
•
•
•
•
•
•
•
•
Editors
XLV Mexican Conference on Biomedical Engineering Proceedings of CNIB 2022, 6–8 October, Puerto Vallarta, México
123
Editors Citlalli Jessica Trujillo-Romero Instituto Nacional de Rehabilitación-LGII Mexico, Mexico Christian Chapa-González Instituto de Ingeniería y Tecnología Universidad Autónoma de Ciudad Juárez Ciudad Juárez, Mexico Dora-Luz Flores Universidad Autónoma de Baja California Ensenada, Mexico Martha R. Ortiz-Posadas Departamento de Ingeniería Eléctrica Universidad Autónoma Metropolitana Iztapalapa Mexico, Mexico
Rafael Gonzalez-Landaeta Departamento de Ingeniería Eléctrica y Computación Universidad Autónoma de Ciudad Juárez Ciudad Juárez, Mexico Guadalupe Dorantes-Méndez Facultad de Ciencias Campus Pedregal Universidad Autónoma de San Luis Potosí San Luis Potosí, Mexico J. J. Agustin Flores Cuautle Conacyt-Instituto Tecnológico de Orizaba Orizaba, Mexico Ricardo A. Salido Ruiz Departamento de Bioingeniería Traslacional Universidad de Guadalajara Guadalajara, Mexico
Esmeralda Zuñiga-Aguilar Universidad Autónoma de Ciudad Juárez Ciudad Juárez, Chihuahua, Mexico
ISSN 1680-0737 ISSN 1433-9277 (electronic) IFMBE Proceedings ISBN 978-3-031-18255-6 ISBN 978-3-031-18256-3 (eBook) https://doi.org/10.1007/978-3-031-18256-3 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The XLV Mexican Conference on Biomedical Engineering (CNIB2022), organized by the Mexican Society on Biomedical Engineering (SOMIB), was held on October 6–8, 2022, in Puerto Vallarta-México. The main goal of CNIB2022 was to promote the latest scientific, academic, and industrial advances in biomedical engineering and any other fields related to the medical and life sciences. The conference brought together leading researchers, academics, professionals, and representatives from the healthcare industry to discuss the latest trends in research, medical technology, regulation, innovation, needs of the healthcare industry, and potential solutions. We are proud to present in this book a selection of papers reporting the latest scientific and technological findings in the field of biomedical engineering. The content is divided into twelve chapters according to outstanding scientific research topics on biomedical engineering. The academic high quality of the content has been evaluated by a strict peer review process guided by the scientific committee and academic reviewers. The final acceptance rate was 58%; therefore, the content represents the effort of more than 600 participants including researchers, students, academics, reviewers, and scientific and organization committees. We are sure this book will provide our readers with a deeper insight into the latest contributions to the biomedical engineering field. Our greatest appreciation to the plenary and scientific sessions speakers, as well as to the representatives from the healthcare industry for making this conference a high-quality recognized event. We would like to thank the scientific committee as well as the organizing committee for their invaluable contribution, hard work, enthusiasm, and optimism along the stage before the conference. Moreover, we would like to thank all the authors, reviewers, and session chairs for their participation and invaluable contribution to the biomedical engineering field. The success of CNIB2022 would not have been possible without all of you. Rafael Gonzalez-Landaeta Co-chair Program CNIB 2022 Citlalli Jessica Trujillo-Romero Chair Program CNIB 2022 v
Distinguished Lecturer Message
The roots of medicine lie at the dawn of human experience on Earth and will forever be part of humanity. Dedicated and often gifted people are entrusted by tribe or society with the very special task of counseling people and helping them heal with a combination of empathy and knowledge. Over the millennia, medicine has evolved into a very complex web of activities that is far too diverse to be left to a single physician to master. Biomedical engineering is one of the cross-cutting approaches needed in modern medicine. The profession of biomedical engineering involves much the same vocation as that leading to other medical professions. You are welcome to read the results of research produced for the XLV Mexican Conference on Biomedical Engineering (CNIB2022) that took place in Puerto Vallarta, Jalisco, by Mexican research groups, authors from most Latin America Countries and beyond. The recent COVID-19 pandemic has brought living difficulties, suffering, and death to humanity. But the consequences could have been much worse—e.g., the influenza pandemic at the beginning of the twentieth century—if there had been no science/technology and international cooperation. This congress is precisely about science/technology shared in a big international event, where students and scientists from different countries and from the Mexican states participate in more than 100 activities among lectures, posters and plenary talks. Given the uncertain and challenging global situation on our planet, the spirit of Puerto Vallarta is related to three aspects: First, what was shared during the Congress is largely the result of collaborative efforts of Mexican, Latin American, and transnational researchers, with the Mexican Society of Biomedical Engineering (SOMIB) as the host organization; second, unlike bellicose behaviors in other regions, the strong determination of Latin America to act along concepts of Democracy and Federalism evolved from original Indian tribal tradition, western philosophy, and religious ethics, to be shared globally; third, the importance of Biomedical Engineering—or medical engineering—which made our species
vii
viii
Distinguished Lecturer Message
“Human-technological” rather than just biologically “Homo Sapiens.” As a species, we should be proud to be able to make our life better on Earth, thanks to medical engineering. Enjoy reading the following pages! Franco Simini Distinguished Lecturer CNIB2022
Organization
CNIB 2022 was organized by the Mexican Society of Biomedical Engineering (SOMIB)
Executive Committee Conference Chair Francisco Javier Aceves Aldrete
Universidad de Guadalajara, México
Program Chair Citlalli Jessica Trujillo-Romero
Instituto Nacional de Rehabilitación-LGII, México
Scientific Committee Program Chair Citlalli Jessica Trujillo-Romero
Instituto Nacional de Rehabilitación-LGII, México
Program Co-chair Rafael Gonzalez-Landaeta
Universidad Autónoma de Ciudad Juárez, México
Session Coordinators Guadalupe Dorantes-Méndez
Universidad Autónoma de San Luis Potosí, México
ix
x
Dora-Luz Flores Martha R. Ortiz-Posadas Ricardo A. Salido Ruiz Esmeralda Zuñiga-Aguilar
Organization
Universidad México Universidad Universidad Universidad México
Autónoma de Baja California, Autónoma Metropolitana, México de Guadalajara, México Autónoma de Ciudad Juárez,
Committee Assistant Angel Balam Benítez-Mata
University of California, Irvine
Awards Committee Student Competition Christian Chapa González J. J. Agustin Flores Cuautle
Universidad Autónoma de Ciudad Juárez, México Instituto Técnologico de Orizaba, México
SOMIB Awards Ana Luz Portillo
Universidad Autónoma de Ciudad Juárez, México
Organizing Committee Organizing Chair Francisco Javier Aceves Aldrete
Universidad de Guadalajara, México
International Guests Elliot Alejandro Vernet Saavedra
Consejo Regional de Ingeniería Biomédica para América Latina
Professional Guests Verónica Castillo
Universidad Olmeca, México
Universities Guests Gabriela Sámano Lira
Universidad Autónoma de Chihuahua, México
Workshops Montserrat Ramírez Nava
Sociedad Mexicana de Ingeniería Biomédica, México
Organization
xi
Exposition Chair Ricardo Gomez Ballardo
Clúster de Ingeniería Biomédica del Estado de Jalisco, México
Students Coordinator Daryana Martínez
Universidad Autónoma de Chihuahua, México
Education Committee Chair Eduardo Méndez Palos
Universidad de Guadalajara, México
Co-chair Mariana Tarín
Universidad Estatal de Sonora, México
Industry Committee Chair Christopher Bricio
Gas Latam México, México
Innovation Committee Chair Luis Fernandez
Tecnología en Ingeniería Clínica, México
Logistics Committee Chair Zayra Resendiz
Corporativo Zaynic, México
Graphic Design Leon Pacheco
Sociedad Mexicana de Ingeniería Biomédica, México
Administrative Support Nicolle Ramírez
Sociedad Mexicana de Ingeniería Biomédica, México
Public Relations Coordinator Sandra Sanchez
Sociedad Mexicana de Ingeniería Biomédica, México
xii
Organization
Referees Abraham Ulises Chávez Ramírez Adeodato Israel Botello Arredondo Adolfo Flores Saiffe Farias Adriana Cristina Pliego Carrillo Aida Jiménez González Aldo Rodrigo Mejia Rodríguez Alejandra Ancira Alejandra Guillén Mandujano Alejandro Aganza Torres Alejandro Miranda Cid Alejandro Morales Alejandro Santos Diaz Alma Aide Sánchez Ramírez
Alvaro Anzueto Rios
Amanda Carrillo Castillo Ana Bertha Pimentel Aguilar Ana Laura García Martínez Aurora Espinoza Valdez Balam Benítez-Mata Braniff de la Torre Berenice Maldonado-Fregoso Bersain Alexander Reyes Carlos Alberto Martínez Pérez Carlos Alberto Pereyda Pierre
Centro de Investigación y Desarrollo Tecnológico en Electroquímica, México Instituto Tecnológico y de Estudios Superiores de Monterrey, México Instituto Tecnológico y de Estudios Superiores de Monterrey, México Universidad Autónoma del Estado de México, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma del Estado de México, México Universidad Autónoma Metropolitana-Iztapalapa, México Universidad Autónoma de San Luis Potosí, México Universidad Politécnica del Valle de México, México Universidad de Guadalajara, México Tecnológico de Monterrey-CDMX, México Unidad Profesional Interdisciplinaria en Ingeniería y Tecnologías Avanzadas IPN, México Unidad Profesional Interdisciplinaria en Ingeniería y Tecnologías Avanzadas IPN, México Universidad Autónoma de Ciudad Juárez, México Instituto Nacional de Enfermedades Respiratorias, México Universidad Politécnica de Chiapas, México Universidad de Guadalajara, México University of California Irvine, USA Universidad de Guadalajara, México Universidad de Colima, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Ciudad Juárez, México Instituto Tecnológico de Hermosillo, México
Organization
Carlos E. Cabrera Ramos Carmen Toro-Castillo César Antonio González Díaz Christian Chapa-González Christian Cruz Sosa Citlalli Jessica Trujillo-Romero Claudia Ivette Ledesma Ramírez Claudia Haydée González de la Rosa Daniel Hernández-Gordillo Daniel Librado Martínez Vázquez Daniel U. Campos Delgado Dante Magdaleno Moncayo David Cervantes Vásquez Dayanira Paniagua Meza Diana Araiza Diomar Enrique Rodríguez Obregón Dora-Luz Flores Eden Morales-Narváez Eduardo Gerardo Mendizabal-Ruiz Eduardo Murillo-Bracamontes Edson Francisco Estrada Meneses Emilio Sacristán Rock Erik Bojorges Valdez Erika Guadalupe Meraz Tena Esmeralda Zúñiga Aguilar
xiii
Centro de Investigación Científica y de Educación Superior de Ensenada, México Universidad de Guadalajara, México Instituto Politécnico Nacional, México Universidad Autónoma de Ciudad Juárez, México Escuela Superior de Ingeniería Mecánica y Eléctrica-Zacatenco, México Instituto Nacional de Rehabilitación-LGII, México Universidad Autónoma del Estado de México, México Universidad Autónoma Metropolitana, México Instituto Mexicano del Seguro Social, México Universidad Autónoma Metropolitana-Lerma, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Baja California, México Universidad Autónoma de Baja California, México Universidad Autónoma de Baja California, México CBIOKS, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Baja California, México Centro de Investigaciones en Óptica A. C., México Universidad de Guadalajara, México Universidad Autónoma de México, México Universidad Autónoma de Ciudad Juárez, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad Iberoamericana, México Universidad Autónoma de Ciudad Juárez, México Universidad Autónoma de Ciudad Juárez, México
xiv
Eunice Vargas Viveros Everardo Gutiérrez López Everardo Inzunza-González Fabiola Margarita Martínez Licona Fabiola Reveca Gómez Velázquez Fausto David Cortes Rojas Filiberto Rivera Torres Flavio Ernesto Trujillo Zamudio Francisco Alvarado Rodríguez Francisco Heredia Francisco Javier Álvarez Padilla Franco Simini Gemima Lara Hernández Gerardo Romo Cárdenas Griselda Quiroz Guadalupe Dorantes Méndez Guillermina Guerrero Mora Guillermo Paredes Gutiérrez Gustavo Adolfo Alonso Silverio Héctor Alejandro Galván Espinoza Héctor Alfaro Hugo Abraham Vélez Pérez Humiko Yahaira Hernández Acosta Imelda Olivas Armendáriz Inés Alejandro Cruz Guerrero
Organization
Universidad Autónoma de Baja California, México Universidad Autónoma de Baja California, México Universidad Autónoma de Baja California, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad de Guadalajara, México Centro de Investigación y de Estudios Avanzados del IPN, México Universidad Nacional Autónoma de México, México Hospital Regional de Alta Especialidad de Oaxaca, México Universidad Autónoma de Guadalajara, México Universidad Autónoma de Yucatán, México Universidad de Guadalajara, México Universidad de la República, Uruguay Instituto Tecnológico de Orizaba, México Universidad Autónoma de Baja California, México Universidad Autónoma de Nuevo León, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Baja California, México Universidad Autónoma de Guerrero, México Instituto Nacional de Cancerología, México Universidad Universidad Universidad México Universidad México Universidad México
de Guadalajara, México de Guadalajara, México Politécnica del Valle de México, Autónoma de Ciudad Juárez, Autónoma de San Luis Potosí,
Organization
Isela Bonilla Gutiérrez Israel Román-Godínez Ivonne Bazán Trujillo Jacinto Villegas Juan Manuel Jaeson Santos Calla Choque Jaime Fabian Vázquez de la Rosa Javier Castro Carmona Javier Flavio Vigueras Gómez Javier Mauricio Antelis Ortíz Jesús Emilio Camporredondo Saucedo Jesús Gómez-Correa Jorge Alberto Pérez León Jorge Alberto Roacho Pérez Jorge Luis Pérez González
Jorge Rodríguez Arce José José José José
Alfredo Soto Álvarez Alfonso Cruz Ramos Ambrosio Bastián Ángel Pecina Sánchez
José David Díaz Román José De Jesús Agustín Flores Cuautle José Francisco Rodríguez Arellano José Javier Reyes Lagos José Joaquín Azpiroz Leehan José Luis Herrera Celis
xv
Universidad Autónoma de San Luis Potosí, México Universidad de Guadalajara, México Universidad Autónoma de Aguascalientes, México Universidad Autónoma del Estado de México, México University of California San Diego, USA Universidad Nacional Autónoma de México, México Universidad Autónoma de Ciudad Juárez, México Universidad Autónoma de San Luis Potosí, México Instituto Tecnológico y de Estudios Superiores de Monterrey, México Universidad Autónoma de Coahuila, México Centro de Investigación Científica y de Educación Superior de Ensenada, México Universidad Autónoma de Ciudad Juárez, México Universidad Autónoma de Nuevo León, México Instituto de Investigaciones en Matemáticas Aplicadas y en SistemasIIMAS Mérida, México Universidad Autónoma del Estado de México, México Universidad de Guanajuato, México Instituto Jalisciense de Cancerología, México Universidad La SALLE, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Ciudad Juárez, México Instituto Tecnológico de Orizaba, México Rinku Research Group, México Universidad Autónoma del Estado de México, México Universidad Autónoma Metropolitana Iztapalapa, México Centro de Investigación y Desarrollo Tecnológico en ElectroquímicaCIDETEQ, México
xvi
José Luis Ortiz Simón José Manuel Valencia Moreno José Marco Balleza Ordaz José Martin Luna Rivera José Manuel Mejía Muñoz Juan Carlos García López Juan Miguel Colores Vargas Juan Odin Ramírez Fernández Julieta García Porres Karla Karina Gómez Lizárraga Karla Patricia Godínez Macías Laura Guadalupe Castruita Avila Laura Mercedes Santiago Fuentes Lidia Núñez Carrera Lizeth Ávila Gutiérrez Luis Carlos Pérez Ruíz Luis Jiménez Ángeles Luz María Alonso-Valerdi Marcelo Romero Huertas Marco Octavio Mendoza Gutiérrez Marcos David Moya Bencomo María de la Luz Mota González María Del Rocío Ortiz Pedroza María de Montserrat Godínez García
Organization
Instituto Tecnológico de Nuevo Laredo, México Universidad Autónoma de Baja California, México Universidad Autónoma de Guanajuato, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Ciudad Juárez, México Centro de Investigación y de Estudios Avanzados del IPN, México Universidad Autónoma de Baja California, México Universidad Autónoma de Coahuila, México NOVARTIS, México Universidad Nacional Autónoma de México, México University of California San Diego, USA Universidad Autónoma de Coahuila, México Universidad Autónoma del Estado de México, México Instituto Nacional de Rehabilitación-LGII, México Instituto Nacional de Geriatría, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad Nacional Autónoma de México, México Instituto Tecnológico y de Estudios Superiores de Monterrey, México Universidad Autónoma del Estado de México, México Universidad Autónoma de San Luis Potosí, México Instituto Tecnológico y de Estudios Superiores de Monterrey, México Universidad Autónoma de Ciudad Juárez, México Universidad Autónoma Metropolitana Iztapalapa, México Vicepresidenta Consejo de Ingenieros Biomédicos, México
Organization
Mariana Álvarez Carvajal Martha Refugio Ortiz Posadas Martin Oswaldo Méndez García Miguel Alejandro Díaz Hernández Miguel Ángel López Guerrero Miguel Ángel Peña Castillo Miguel Ángel San Pablo Juárez Mónica Vázquez Hernández Nallely Patricia Jiménez Mancilla Nayda Patricia Arias Duque Nohra Elsy Beltrán Vargas Norma Alicia Barboza Tello Norma Castañeda Villa Norma Ramírez Nelly Gordillo Castillo Omar Mendoza Montoya Omar Paredes Omar Piña Ramírez Oscar Fernando Avilés Sánchez Oscar Yáñez Suárez Otniel Portillo Rodríguez Pablo Antonio Stack Sánchez Pablo Samuel Luna Lozano Paola Andrea Niño Suarez Pedro Bertemes Filho Rafael Bayareh Mancilla
xvii
Universidad Autónoma del Estado de México, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad Autónoma de San Luis Potosí, México Universidad Autónoma de Baja California, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad de las Américas Puebla, México Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, México Instituto Nacional de Investigaciones Nucleares, México Universidad de Sucre, Colombia Universidad Autónoma Metropolitana, México Universidad Autónoma de Baja California, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad de Guadalajara, México Universidad Autónoma de Ciudad Juárez, México Instituto Tecnológico y de Estudios Superiores de Monterrey, México, México Universidad de Guadalajara, México Instituto Nacional de Perinatología, México Universidad Militar Nueva Granada, México Universidad Autónoma Metropolitana Iztapalapa, México Universidad Autónoma del Estado de México, México Universidad de Alberta, Canadá Universidad Veracruzana, México Instituto Politécnico Nacional ESIME Azcapotzalco, México Universidade do Estado de Santa Catarina, Brasil Centro de Investigación y de Estudios Avanzados del IPN, México
xviii
Rafael Eliecer González-Landaeta Raquel Martínez Valdez Rebeca Romo-Vázquez Ricardo Estrada Meza Ricardo Perea Jacobo Ricardo Rodríguez Vera Rigoberto Martínez-Méndez Roberto Carlos Carrillo Torres Roberto Giovanni Ramírez Chavarría Sandra Balderas Santos Adriana Martel Estrada Saraí Esmeralda Favela Camacho Sergio Eduardo Sánchez Hernández Sergio Rivera Tello Sergio Sánchez-Manzo Solange Ivette Rivera Manrique Stewart Santos Sulema Torres-Ramos Svetlana Kashina Tomás Zamudio López Valeria del Carmen Silva Acosta Vianney Muñoz-Jiménez Yizel Becerril Alarcón Zaira Pineda Rico
Organization
Universidad Autónoma de Ciudad Juárez, México Universidad Politécnica de Chiapas, México Universidad de Guadalajara, México CBIOKS, México Universidad Autónoma de Baja California, México Instituto Nacional de Enfermedades Respiratorias, México Universidad Autónoma del Estado de México, México Universidad de Sonora, México Universidad México Universidad Universidad México Universidad México Universidad
Nacional Autónoma de México, de Guadalajara, México Autónoma de Ciudad Juárez, Autónoma de Ciudad Juárez, de Guadalajara, México
Universidad de Guadalajara, México Mayers Memorial Hospital District, USA Universidad La SALLE, México Universidad de Guadalajara, México Universidad de Guadalajara, México Universidad Autónoma de Guanajuato, México BAZAFI, México Baxter, México Universidad Autónoma del Estado de México, México Universidad de la Salud del Estado de México, México Universidad Autónoma de San Luis Potosí, México
Organization
Sponsoring Institutions Mexican Society of Biomedical Engineering (SOMIB)
xix
Contents
Artificial Intelligence and Data Science Device for the Fall Detection in Older Adults Through Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Carolina Arana Cohuo, Luz Andrea Hernández Ocón, Diana Marilú Domínguez Lizama, Diego Alejandro González Bautista, Sahyan Mutt Ruiz, and Rutilio Nava Martínez Breast Cancer Detection Algorithm Using Ensemble Learning . . . . . . . Sophia Sandoval Torres, Ana Paola Romero Espinoza, Grisel Jhovana Castro Valles, and Carlos Eduardo Cañedo Figueroa Graph Analysis of Functional Connectivity Rs-FMRI in Healthy and Epileptic Brain Using Visibility Algorithm . . . . . . . . . . . . . . . . . . . Rosa Victoria Villa Padilla, Katya Rodríguez Vázquez, Mónica Vázquez Hernández, Bayron Alexander Sandoval Bonilla, and Josafat Jonathan Sánchez Dueñas Imagined Speech Recognition in a Subject Independent Approach Using a Prototypical Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Alan Hernandez-Galvan, Graciela Ramirez-Alonso, Javier Camarillo-Cisneros, Gabriela Samano-Lira, and Juan Ramirez-Quintana Design and Comparison of Artificial Intelligent Algorithms for Breast Cancer Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Karen Valdez Hernández, Jhovana Cano Villalobos, Ana Castro Reyes, Andrea Gutiérrez Jurado, Sofia Moreno Terrones, Carlos Eduardo Cañedo Figueroa, Abimael Guzmán Pando, and Gabriela Sámano Lira Electrophysiological Signals Simulation with Machine Learning . . . . . . Mario Axel López Aguiñaga, Arturo Valdivia González, and Laura Paulina Osuna Carrasco
3
14
27
37
46
55
xxi
xxii
Contents
Quantification of a Lip and Palate Clefts Classification . . . . . . . . . . . . . Beatriz Gutiérrez-Sánchez, José Maya-Behar, and Martha Ortiz-Posadas
73
Artificial Intelligence Applied to Breast Cancer Classification . . . . . . . . Samara Acosta-Jiménez, Javier Camarillo-Cisneros, Abimael Guzmán-Pando, Susana Aideé González-Chávez, Jorge Issac Galván-Tejada, Graciela Ramírez-Alonso, César Francisco Pacheco-Tena, and Rosa Elena Ochoa-Albiztegui
83
Computational Chemistry as an Educational Tool in Health Sciences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Alexica Celine Márquez-Barreto, Celia María Quiñones-Flores, Graciela Ramírez-Alonso, Gabriela Sámano-Lira, and Javier Camarillo-Cisneros
94
A Gene-Community Overview of Transcriptional Dynamics During Neurodevelopment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 Gustavo Guzmán, Elsa Magaña-Cuevas, Juan Serna-Grilló, Omar Paredes, Hugo Vélez-Pérez, Rebeca Romo-Vázquez, and Jose Alejandro Morales CNNs for ISCI Stage Recognition on Video Sequences . . . . . . . . . . . . . 111 Gabriela Aguirre-Espericueta and Gerardo Mendizabal-Ruiz Stacked Spatial and Temporal Deep Learning Methods for Identification of Parkinson’s Disease Using Gait Signals . . . . . . . . . 119 Brenda Guadalupe Muñoz-Mata, Guadalupe Dorantes-Méndez, and Omar Piña-Ramírez Diversity of Genotyping Chlamydia Trachomatis Serovars in Urogenital Samples from Mexican Patients: A Molecular and Bioinformatic Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Fabiola Hernández-Rosas, Socorro Mariana García-González, Shumeyker Susmith Franco-González, Ana Paola Salgado-Álvarez, and Mercedes Piedad de León-Bautista Detection of Breast Cancer in Mammography Using Deep Learning Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 Ricardo Perea-Jacobo, Guillermo Paredes-Gutierrez, Miguel-Angel Guerrero-Chevannier, Dora-Luz Flores, and Raquel Muñiz-Salazar Modeling and Simulation of Biological Systems A Comparative Study on the Interaction of an Ototoxic and an Otoprotective with the Megalin Receptor Associated with Hearing Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 Gerardo David Hernández Cornejo, Iris Natzielly Serratos Álvarez, César Millán-Pacheco, Jonathan Osiris Vicente-Escobar, and Norma Castañeda-Villa
Contents
xxiii
Collagen/Plasma-Polymerized Pyrrole Interaction: Molecular Docking and Binding Energy Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Teresa Gómez-Quintero, Iris Serratos-Alvarez, Rafael Godínez, and Roberto Olayo Implanted Pediatric Patient Early Audiometry . . . . . . . . . . . . . . . . . . . 162 Juan Manuel Cornejo Cruz, Agar Karina Quintana López, and Ma. del Pilar Granados Trejo Thermal Performance of a Triple Slot Antenna Considering Temperature Dependence of Thermal and Electrical Conductivity, Blood Perfusion and Tissue Metabolism . . . . . . . . . . . . . . . . . . . . . . . . . 170 Dalia Braverman-Jaiven and Citlalli Jessica Trujillo-Romero Modeling of the Interaction of Plasma-Polymerized Pyrrole with Immunoglobulin M (IgM) by Biocomputational Tools . . . . . . . . . . . . . . 179 Esteban Rafael Ramírez Perez, Iris Natzielly Serratos, César Millán-Pacheco, Salvador Tello-Solís, and Roberto Olayo-Valles Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Ana Paola Leyva-Aizpuru, Yoshua Alberto Quezada-García, Graciela Ramirez-Alonso, Luis Carlos Hinojos-Gallardo, and Javier Camarillo-Cisneros Simulating the Ca2+-cAMP Crosstalk and Its Role in Pancreatic Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 Hugo Enrique Romero-Campos, Geneviève Dupont, and Virginia González-Vélez Simulating the Loss of b-cell Mass in a Human Pancreatic Islet: Structural and Functional Implications . . . . . . . . . . . . . . . . . . . . . . . . . 204 Sergio Ruiz-Santiago, José Rafael Godínez-Fernández, and Gerardo Jorge Félix-Martínez Role of Endogenous Ca2 þ Buffering and the Readily Releasable Pool on Fast Secretion in Auditory Inner Hair Cells . . . . . . . . . . . . . . . 212 Crystal Azucena Valverde-Alonzo, Gerardo Jorge Félix-Martínez, Virginia González-Velez, and Amparo Gil Effects of Blood Flow on Insulin Concentration: A Modelling Study . . . 219 Diego Alejandro Flores-Santillán, José Rafael Godínez-Fernández, and Gerardo Jorge Félix-Martínez Non-invasive Hypoglycemia Regulatory Patch with Glucagon Administration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 Jennifer Monserrat Gonzalez-Martinez, Jesús Emilio Méndez-Sánchez, Odin Ramirez-Fernandez, Ivan Cipriano Urbano, and Emilio Camporredondo
xxiv
Contents
The Enzymatic Core of Snakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234 Leonardo Juárez-Zucco, Victor Alvarado-Aparicio, Teresa Romero-Gutiérrez, and Ernesto Borrayo Structural Analysis for Enzymatic Homology Determination in Terpene Cyclases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242 Enrique Farfán-Ugalde, Cindy V. Flores Hernandez, Elsa Magaña-Cuevas, Omar Paredes, and J. Alejandro Morales Effect of Thermal Dependence of Tissue Properties on the Antenna Performance: A 3D Parametric Model . . . . . . . . . . . . . . . . . . . . . . . . . . 250 Gustavo Gutiérrez-Miranda and Citlalli Jessica Trujillo-Romero Hepatic Cell Radial Flow Bioreactor Parametrization and Characterization as an Alternative Therapy to Liver Failure . . . . . 259 Hector Adrian Ramirez-Nuñez, Odin Ramirez-Fernandez, Emilio Camporredondo, and Omar Anaya-Reza Medical Physics and Nuclear Medicine Gamma Radiation Detection Simulation System . . . . . . . . . . . . . . . . . . 271 Ana Cristina Torres-Alamilla, Anna Moreno-Mina, Eglaín Constantino-Cortés, and Diana Paulina Martínez-Cancino Development of an Alternative Radiochromic Film Digitizer for Clinical Dosimetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 Gerardo Jiménez-Aviles, Miguel Camacho-López, Olivia García-Garduño, and Keila Isaac-Olivé Processing of Biomedical Signals Decoding Imagined Speech of Daily Use Words from EEG Signals Using Binary Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 Marianna Gutiérrez-Zermeño, Edgar Aguilera-Rodríguez, Emilio Barajas-González, Israel Román-Godínez, Sulema Torres-Ramos, and Ricardo A. Salido-Ruiz Nonlinearity of Electrohysterographic Signals is Diminished in Active Preterm Labor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 José Rodrigo Zamudio-De Hoyos, Diego Vázquez-Flores, Adriana Cristina Pliego-Carrillo, Claudia Ivette Ledesma-Ramírez, Hugo Mendieta-Zerón, and José Javier Reyes-Lagos Trend of Concentration of Men and Women Elucidated by Analysis of EEG Signals Recorded During a Fast Game . . . . . . . . . . . . . . . . . . . 308 María Guadalupe Márquez Acá, Lucila Iraís Castelán León, Lorenzo Armando Matamoros García, and Alina Santillán Guzmán
Contents
xxv
Effects on Body Posture and Gait Caused by Different Weights in the Backpack of University Students . . . . . . . . . . . . . . . . . . . . . . . . . 316 Evelin Daniela Ramírez Ponce, Karla Arenas-Valerio, and Yajaira Zepeda-García Multiscale-Multifractal Assessment of Heart Rate Variability in Shift Workers by Detrended Fluctuation Analysis . . . . . . . . . . . . . . . 324 Raquel Delgado-Aranda, Guadalupe Dorantes-Méndez, Martín Oswaldo Méndez, Anna Maria Bianchi, and Juha Kortelainen EEG Connectivity Analysis in a Motor Imagery Task . . . . . . . . . . . . . . 332 César Covantes-Osuna, Omar Paredes, Diana Yaneli De la Mora, Hugo Vélez-Pérez, and Rebeca Romo-Vázquez Brain Mapping: Location of the Words Through EEG . . . . . . . . . . . . . 342 Omar Cano-Garcia, María Hernández-Rizo, Lorena López-Medina, and J. Alejandro Morales Processing of Biomedical Images Artifacts Generated by the 3D Rotation of a Freely-Swimming Human Sperm in the Measurement of Intracellular Ca2+ . . . . . . . . . . . . . . . . . . 355 Andrés Bribiesca-Sánchez, Fernando Montoya, Ana Laura González-Cota, Paul Hernández-Herrera, Alberto Darszon, and Gabriel Corkidi Morphological Temporal Analysis in Subjects with Alzheimer’s Disease by Brain Graph Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 Laura Gonzalez–Meza, Jesus Siqueiros–Garcia, Nidiyare Hevia–Montiel, José Javier Reyes–Lagos, and Jorge Perez–Gonzalez PET Image Reconstruction Using a GRU-Convolutional Network . . . . . 371 Jose Mejia, Boris Mederos, Leticia Ortega-Máynez, Nelly Gordillo, and Lidia Hortencia Rascón-Madrigal Characterization of COVID-19 Diseased Lung Tissue Based on Texture Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382 Jesús Gibrán Delgado-Alejandre, Diomar Enrique Rodríguez-Obregón, Alejandro Santos-Díaz, and Aldo Rodrigo Mejía-Rodríguez Glioblastoma Classification in Hyperspectral Images by Reflectance Calibration with Normalization Correction and Nonlinear Unmixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 Inés Alejandro Cruz-Guerrero, Juan Nicolas Mendoza-Chavarría, and Daniel Ulises Campos-Delgado Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein with the Phasor-FLIM Approach Signatures of Laurdan . . . . . 403 Balam Benítez-Mata, Francesco Palomba, Zhiqun Tan, Leslie Thompson, and Michelle Digman
xxvi
Contents
A Method for Automatic Monoplane Angiography Segmentation . . . . . 414 Héctor Emanuel Martín Alcala, Francisco Javier Alvarez Padilla, and Gerardo Mendizabal Ruiz Lung Segmentation Algorithm and SVM Classification of COVID-19 in CT Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 Luis Eduardo Gaeta-Ledesma and Francisco Javier Alvarez-Padilla IOT in Health and Bioinstrumentation Disinfection Method Based on UV-C Light Using the Internet of Things for Cleaning Hospital Areas (COVID-19) . . . . . . . . . . . . . . . . 437 Stephanie Carolina Juárez-García, Misael Sánchez-Magos, Iván Matehuala-Morán, Christi Torres-Vargas, Francisco Muñoz del Ángel, Ricardo Bautista Mercado, Juan Jesús Mejía Fernández, and Fanny Alvarado Development of Alpha Prototype of Handheld Device for Meibography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448 Héctor Retana, Erik Bojorges, and Everardo Quintela Prototype of a Pulse Oximeter Based on an Open-Source Platform with Wireless Design and Cloud Service . . . . . . . . . . . . . . . . . . . . . . . . 459 Martín Aarón Sánchez Barajas, Daniel Cuevas González, Roberto López Avitia, Marco Antonio Reyna, Juan Pablo García-Vázquez, and Néstor Alexander Zermeño Campos Wearable System for Measuring Vertical Ground Reaction Forces During the Gait Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468 David Alvarado-Rivera, Paola Andrea Niño-Suárez, and Leonel German Corona-Ramírez Design of a Pulse Oximeter with Altitude Measurement Bluetooth Communication and Android Application . . . . . . . . . . . . . . . . . . . . . . . 477 Carlos Adrián Cruz Malvaéz, Aurey Galván Lobato, and Manuel Ortínez Benavides Braille System Learning Introductory Device . . . . . . . . . . . . . . . . . . . . . 493 Karla Córdova-Reyes, Rodolfo López-Villarreal, Jonatan Oliva-Rodríguez, Olivia Sánchez-Barrios, and Diana Martínez-Cancino Signal to Noise Ratio and Current Consumption in LED-LED Photoplethysmography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502 Aurora Osorio, Angel Sauceda-Carvajal, and Rafael Gonzalez-Landaeta Prototype for the Monitoring of Soda Lime in Anesthesia Machines Using Wi-Fi Alarm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510 Morelia Vásquez-Quiroz, Belem Mendoza-Muñoz, José Vázquez, and Diana Paulina Martínez-Cancino
Contents
xxvii
Design and Implementation of a Smartphone-Based Digital Phonocardiograph with Wireless Transmission Capabilities . . . . . . . . . 518 Alexis Raciel Ibarra-Garnica and Bersaín Alexander Reyes System for Detection of Neonatal Apnea . . . . . . . . . . . . . . . . . . . . . . . . 530 Lizbeth Diaz Guerra, Rogelio Manuel Higuera González, and Tania Jetzabel Contreras Uribe The Road to Making “Exergames” More Widely Available . . . . . . . . . . 536 Brenda Nicole Gómez-Ávila, Alan Javier Escobedo-Núñez, Esmeralda del Socorro Orozco-Díaz, and Ricardo Antonio Salido-Ruiz Design and Construction of Capacitive Coupling Electrostimulator to Induce Bone Tissue Regeneration . . . . . . . . . . . . . . . . . . . . . . . . . . . 549 Romina Fontes Ruiz and María Flores Sánchez Biosensors Paper-Based Microanalytical Device for Colorimetric Detection of Stress in Human Saliva Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567 Paulina Hernández-Garcés and Nikola Batina Construction of an Electrochemical Nanogenosensor for K-RAS Oncogene Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576 Norma Andrea Chagoya Pio, Nikola Batina, and Luis Fernando Garcia-Melo Computational Study of a:SiC:H Thin Films Deposited on Interdigitated Microelectrodes Using Electrical Impedance Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585 José Herrera-Celis, Diana Jiménez-Rivas, Claudia Reyes-Betanzo, Emilia Méndez-Aguilar, Francisco Cuevas-Muñiz, and Goldie Oza Development of Non-enzymatic Sensor for Uric Acid Detection Based on Gold Nanoparticles Electrodeposited on Laser-Induced Graphene Electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594 Héctor David Hernández, Eider Aparicio-Martinez, Rocío Berenice Dominguez, and Juan Manuel Gutiérrez Unmodified Screen-Printed Electrodes-Based Sensor for Electrochemical Detection of Bisphenol A . . . . . . . . . . . . . . . . . . . . 603 María J. Hernández-Gordillo, Bryan E. Alvarez-Serna, and Roberto G. Ramírez-Chavarría Molecularly Imprinted Polymer Paper-Based Biosensor for Wireless Measurement of Sweat Glucose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611 Bryan E. Alvarez-Serna, Ain-ek Balderas-Zempoaltecaltl, and Roberto G. Ramírez-Chavarría
xxviii
Contents
Bioimpedance and Micro-nanotechnologies The Predictive Capacity of Bioelectrical Impedance Parameters at Frequencies of 5, 20, 50, 100, and 200 kHz to Identify Vector-Associated Febrile Syndromes in the Emergency Room of the Hospital Civil de Guadalajara . . . . . . . . . . . . . . . . . . . . . . . . . . . 621 Jennifer Vargas López, Rocio Bojórquez Pérez, Esteban González Díaz, Gabriela del Carmen López Armas, and José Cruz Ramos Application of Palladium Nanoparticles as a Contrast Agent for Electrical Bioimpedance Measurements on Biological Tissue . . . . . . 630 Andrea Monserrat del Rayo Cervantes Guerrero, Sofía Terán Sánchez, José Marco Balleza Ordaz, María del Rosario Galindo González, Francisco Miguel Vargas Luna, and Svetlana Kashina Bladder Volume Monitoring by Electrical Bioimpedance Technique. Calibration Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638 Jasiel Jaimes Lopez, Mariana Herrera Mosqueda, and Jose Marco Balleza Ordaz Mechanical Characterization of Patellar Tendon Strain by Electrical Impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648 Ximena Marbán Guerrero and José Marco Balleza Ordaz Nanoparticles for Glioblastoma Treatment . . . . . . . . . . . . . . . . . . . . . . . 656 Karen Janeth Guerra Sánchez, Nelly Gordillo Castillo, Saraí Esmeralda Favela Camacho, and Christian Chapa González Bioimpedance Spectra in Final RT-PCR Products: A Sensitivity Threshold Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665 Karla Lizeth Padilla García, Modesto Gómez López, Jennifer Viridiana Sánchez Camacho, Claudia Mariana Andrade Torres, Nadia Mabel Pérez Vielma, and César Antonio González Díaz Biomaterials, Molecular, Cellular and Tissue Engineering Design and Fabrication of a Radial Flow Bioreactor to Decellularize Muscular Arteries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677 Odin Ramírez-Fernández, Esmeralda Zuñiga-Aguilar, Laura Castruita, Emilio Camporredondo, David Giraldo-Gomez, David Abad-Contreras, and María Cristina Piña-Barba Iron Carbide@Iron Oxide Core-Shell Nanoparticles Functionalization with L-Arginine Amphiphillic Bioconjugate . . . . . . . . . . . . . . . . . . . . . . 684 Paul Zavala Rivera, Jesús Armando Lucero Acuña, Patricia Guerrero Germán, Aaron de Jesús Rosas Durazo, Lizbeth Alcantara Bastida, and Anya Isabel Argüelles Pesqueira
Contents
xxix
Evaluation of Hemolytic Behavior and Bioactive Properties of Natural Wollastonite and Synthetic Hydroxyapatites Produced by Two Sol-Gel Routes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694 Luis Alberto Núñez Rodríguez, Martín Antonio Encinas Romero, Dora Alicia Cortés Hernández, Jesus Leobardo Valenzuela García, Agustín Gómez Álvarez, and Diana Meza Figueroa Microstructure and Mechanical Properties of Hydroxyapatite Nanofibers Synthesized Through the Microwave-Assisted Hydrothermal Method for Biomedical Applications . . . . . . . . . . . . . . . . 705 Kevin Martínez-Arellano, Fabiola Hernández Rosas, and José Rafael Alanís-Gómez Biological Pacemakers Obtained Through Cellular Differentiation for the Restoration of Sinoatrial Node Function. A Systematic Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714 Julia Aidee Magallanes Marrufo, Victor Gómez Flores, Dora Luz Flores Gutiérrez, Rafael Eliecer González Landaeta, and Christian Chapa González Evaluation of the Formation of an Ionic-Complementary Self-assembling Peptide Hydrogel for the Three-Dimensional Culture of Mammalian Cells in Vitro . . . . . . . . . . . . . . . . . . . . . . . . . . . 721 Brandhon Francisco Flores-Ibarra and Luis Alberto Castillo-Díaz 3D Bioprinting of Hydrogels Using Hydrophobic Sands and Calcium Chloride as Structural Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 729 Mónica Pamela Montes-Ballardo, Jessica Marlene Medina-Lizárraga, Mariana S. Flores-Jiménez, and Rita Q. Fuentes-Aguilar Rehabilitation, Biomechanics and Biorobotics Evaluation of Muscle Activity and Predisposition to Pain in Male Volleyball Players . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 741 Mateo Gomez Arbelaez, Isabel C. Soto, and Elizabeth Pareja Hands-Free Walking Stick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 749 Juan Carlos Colin-Ortega and Alexa García-Aguilar Elbow Torque Estimation for Human-Robot Interaction Control . . . . . 760 Víctor Iván Ramírez-Vera, Marco Octavio Mendoza-Gutiérrez, and Isela Bonilla-Gutiérrez Prototype of an Active Partial Hand Prosthesis for a Person with Symbrachydactyly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770 Osmar Jassiel Machuca-Herrada, Ricardo Tapia-Herrera, and Manuel Arias-Montiel
xxx
Contents
Electronic System to Determine Proximal and Medial Phalanges Strength in a Hand Exoskeleton Robot . . . . . . . . . . . . . . . . . . . . . . . . . 781 Denisse German-Alonso, Miguel Hernández-Ramos, José de Jesus Agustín Flores Cuautle, Ofelia Landeta-Escamilla, Juan Manuel Jacinto-Villegas, Gerardo Aguila-Rodriguez, and Oscar Osvaldo Sandoval-Gonzalez Clinical Engineering and Education Innovation and Control of Health Technology Management Procedures Applying Six Sigma Methodology . . . . . . . . . . . . . . . . . . . . 793 Y. J. Navarro-Arcos, A. B. Aguilar-Pimentel, and M. R. Ortiz-Posadas Exploratory Data Analysis for Preventive and Corrective Maintenance for Medical Equipment in a General Hospital from the Health Institute of the State of Mexico . . . . . . . . . . . . . . . . . . . . . . . 805 D. N. Astivia-Chávez and M. R. Ortiz-Posadas Obsolescence Assessment Approach: Case of Mechanical Ventilators Under the Covid-19 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816 Rafael de Jesus Jimenez-Maturano and Fabiola Martinez-Licona Application of the Quality Function Deployment Methodology for Quality Analysis in the Clinical Laboratory . . . . . . . . . . . . . . . . . . . 826 Pablo Alexis Alejo-Vilchis and José Javier Reyes-Lagos Strategies Employed in the Reconfiguration of Healthcare Facilities During COVID-19 in OECD Countries . . . . . . . . . . . . . . . . . . . . . . . . . 836 Vanesa Cano, Nelly Gordillo-Castillo, and Ana Luz Portillo CO2 Levels in the Naso-Buccal Area Due to the Use of Different Face Masks in Different Ventilation Conditions . . . . . . . . . . . . . . . . . . . . . . . 843 Stephanie Saenz, Angel Sauceda-Carvajal, Nelly Gordillo-Castillo, Christian Chapa, and Rafael Gonzalez-Landaeta Creation of a Needs Detection System for High Technology Medical Equipment or Medical Equipment that Require an Infrastructure Specification for the Mexico City’s Secretariat of Health . . . . . . . . . . . . 851 Claudia Patricia Quiroz-Flores, José Antonio Lobaco-Montes de Oca, and Alfonso Hernández-Rico Use of Audiovisual Strategies as a Complementary Resource for Practical Courses in Biomedical Engineering . . . . . . . . . . . . . . . . . . 860 Jorge Luis Rodríguez-Medina, Guadalupe Dorantes-Méndez, and Aldo Rodrigo Mejía-Rodríguez
Contents
xxxi
Innovation of Technologies for Health Mechanical Design and Additive Manufacturing for a Low-Cost Hybrid Dermatoscope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 873 José Alberto Rodríguez-Mayrén, José Ricardo Cano-García, Maximiliano Zamora-Vega, Iván Matehuala-Morán, María Monserrat Díaz-Hernández, Lizeth Machado-Jaimes, Ruben Fuentes-Alvarez, Judith Guadalupe Dominguez Cherit, and Mariel Alfaro-Ponce Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885
Artificial Intelligence and Data Science
Device for the Fall Detection in Older Adults Through Neural Networks Carolina Arana Cohuo(B) , Luz Andrea Hernández Ocón, Diana Marilú Domínguez Lizama, Diego Alejandro González Bautista, Sahyan Mutt Ruiz, and Rutilio Nava Martínez Universidad Modelo, 97305 Mérida, Yucatán, Mexico [email protected]
Abstract. For an elderly adult, a fall can be fatality by the fact that his bones are fragile and propense to fracture. The present device is intended to detect a change of posture prior to a fall by using a sensor composed of an accelerometer and a triaxial gyroscope. The signal processing used multiple classification neural networks with hyper-parameters of five hidden layers and 500 interactions, training and validation showed an accuracy of 83% respectively. In the results, the device detects fall postures efficiently and with a minimum margin of error. Keywords: Fall detection · Falls in older adults · Neural networks
1 Introduction Hip fractures are debilitating and the leading cause of injury death among people age 65 and older. More than 350,000 hip fractures occur each year in the United States alone, at a total cost of 16 billion dollars [1]. Elderly people tend to fall because of balance disorders, low blood pressure, sudden changes in posture, visual disorders or some disease that affects their joints. In Mexico, falls represent 30% of the causes of death in people over 65 years of age [2], and a single fall can generate several consequences, among which are inflammation, pain in the area of the shock, minor skin injuries such as: scratches or bruises, and in the case of a higher degree injury, fractures and loss of consciousness. In Yucatan, a cross-sectional study that included patients older than 60 years diagnosed with hip fracture between November 2016 and April 2019 was conducted. The characteristics of the patients and the type of fracture, including anatomical location and type of trauma, were analyzed. It was observed that the most frequent type of fracture was extracapsular and the most prevalent subtype was trans-trochanteric [3]. The authors Bet, Castro and Ponti [4] show a compilation of works on older adult fall detection, the research papers come from important base dates: PubMed, IEEExplorer and Scopus. One of the technologies mentioned was computer vision, even though the review does not focus on this theme, other articles report good results in detecting falls with computer vision [5, 6]. Computer vision is a powerful tool for image processing. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 3–13, 2023. https://doi.org/10.1007/978-3-031-18256-3_1
4
C. Arana Cohuo et al.
Another alternative for a acquire data to analyze gait, balance and falls are sensors: the accelerometer and gyroscope. While the accelerometer quantifies the acceleration of a body, the gyroscope provides direction information of the same. Again, in review [4] show that, from 2010 to 2019, there was 29 projects using the accelerometer; most using smart phones, for example: the research projects of Handelzalts et al. [7] and Nguyen, Zhou, Mirza and Naeem [8] used wearable and smartphone, respectively. In the works of Nari et al. [9], and Jefiza et al. [10], show the use of Arduino with the MPU-6050 sensor, so they also use neural networks (NN): signal vector magnitude (SVM) y back-propagation neural network (BPNN), respectively. In both jobs, the Arduino acquires the information from the sensor. Nari et al. sends the information via Bluetooth to a computer that processes the information with SVM; while Jefiza stores the information on an SD and then processes the information on a computer with BPNN. Obtaining good results. Now, there is no device that uses Arduino with a multiple classification NN for posture assessment of an older adult, that is, that the Arduino processes the information without a computer. So that the objective of this project is to design a device for older adults, with the module MPU6050 (accelerometer) to detect a change of posture prior to a fall using multiple classification NN inside the Arduino.
2 Methodology 2.1 Selection of Electronic Components The device is mainly composed of the MPU6050 accelerometer and gyroscope, which has a system that combines a three-axis gyroscope and a three-axis accelerometer [11], as can be seen in Fig. 1.
Fig. 1. Sensor MPU6050 [10].
The gyroscope was not considered for signal capture. Only the accelerometer was used to detect movements in real time. The other device is the Arduino ONE R3 board, the device used for the processing and interpretation of the acquired signals. 2.2 General Procedure The entire process would be explained in three main blocks, see Fig. 2:
Signal acquisition
Device for the Fall Detection in Older Adults
5
Circuit construction (Arduino ONE y MPU6050)
Arduino-Python serial communication
Arduino with NN
CNN setup and training
Stored posture signal records (5 posture)
Three input layers (X, Y and Z axis)
Five hidden layers and 500 interactions (epochs)
Activation functions, loss function and backpropagation
Parameters and hyperparameters obtained
Values copied to the Arduino
Operation of the NN in Arduino
Fig. 2. General procedure
The first block consists of acquiring the accelerometer signals, it must be remembered that the gyroscope was not considered for this research work, and all the signals of each position must be stored in the computer. For this reason, all the falls were simulated, that is, the postures and movements when person falls were replicated. Stored poses last 10–15 s, which is like 15 simulations. All simulations were performed by D.M. Domínguez-Lizama y D.A. González-Bautista. The companions do not present motor or postural difficulties. The second block is about the configuration of the NN. The setup was done in Python with the TensorFlow library, the parameters and hyperparameters used can be read in Neural network architecture for posture detection. Finally, the results given by the training are copied to the Arduino, with this, the Arduino One has a functional NN. It should be noted that the Arduino does not need to be connected to the computer for the processing and interpretation of the signals, this means that the device is totally independent from the computer or smartphone.
6
C. Arana Cohuo et al.
2.3 Neural Network Architecture for Posture Detection A NN is a system of connections to simple, hierarchically organized elements. It is a tool capable of separating the classes involved [12], in order to achieve interaction with real-world objects. The NN is configured with three main parts: input layers (X, Y and Z axis), units representing input signals; the hidden layers, in this work there were five; and the output layer, which has five options, see Fig. 3.
Fig. 3. Front fall in multiple classification NN
The NN used is multiple classification, because there are five options, which belongs to the convolution neural network (CNN). Each sample can belong to one of C classes. The CNN will have C output neurons that can be gathered in a vector s (Scores). The target (ground truth) vector t will be a one-hot vector with a positive class and C − 1 negative classes. This task is treated as a single classification problem of samples in one of C classes, see Fig. 4.
Device for the Fall Detection in Older Adults
C=3
7
Samples
Labels (t) [0,0,1]
[1,0,0]
[0,1,0]
Fig. 4. Representation of a CNN multiple classification with three outputs
Subsequently, activation functions were programmed: the rectified linear unit (ReLu) transfer function, see Eq. 1, and SoftMax, see Eq. 2. The vectors coming out from CNNs (s) before the loss computation. if s ≤ 0, 0 y(s) = (1) if s > 0, s Negative values are not important in processing and so are set to 0. But positive values after convolution should be passed to the next layer. The SoftMax function cannot be applied independently to each si, since it depends on all elements of s. For a given class si, the SoftMax function can be computed as: esi y(s)i = C
j=1 e
sj
(2)
where sj are the scores inferred by the net for each class in C. Note that the SoftMax activation for a class si depends on all the scores in s. The next step of the training was the configuration of the loss function: multiple cross-entropy. In this case, it is intended for use with multi-class classification where the target values are in the set {0, 1, 2, 3, …, n}, where each class is assigned a unique integer value, see Eq. 3. 1 n (3) pi ∗log yi L(θ ) = − i=1 n where pi is the actual label, and yi is the classifier’s output. The cross-entropy loss is the negative of the first, multiplied by the logarithm of the second. Also, m is the number of examples, so the total loss is the average loss over all the examples. Finally, the backpropagation. The adaptive moment estimation (ADAM) optimizer was used, see Eq. 4 and 5. The reason was to minimize the error and obtain exact weights. m = β1 ∗ m + (1 − β1 ) ∗ W
(4)
v = β2 ∗ v + (1 − β2 ) ∗ W 2
(5)
where m and v are moving averages, W is gradient on current mini-batch, and betas, new introduced hyper-parameters of the algorithm. The vectors of moving averages are initialized with zeros at the first iteration.
8
C. Arana Cohuo et al.
Once the NN was trained and with an accuracy of over 90% in the validation of the training, the results and values were transcribed to the Arduino. 2.4 Device Design The design of the device was elaborated in the 3D Builder program taking a vest as a model, where the sensor was in the middle of the body, for this a compartment in the waist was included in the interior Fig. 5.
Fig. 5. Vest model with MPU6050 sensor compartment
2.5 Participants 15 people greater than or equal to 65 years of age participated voluntarily, with the informed consent of written. None have mental, motor or posture problems. The evaluations strictly complied with the Declaration of Helsinki and Nuremberg. Similarly, the device is non-invasive.
3 Results 3.1 Sample Characterization Table 1 shows the characteristics of the participants.
Device for the Fall Detection in Older Adults
9
Table 1. Descriptive results of the sample. *w = w = woman, m = man; cm = centimeters; kg = kilogram; average (standard deviation). Descriptive results Age
Sex (w/m)
Weight (kg)
Height (cm)
72 (5.6)
(8/7)
67.8(15.5)
156.6 (5.3)
3.2 Device Fall Detection The Fig. 8 show the vest with the MPU-6050 sensor and Arduino. The electronic circuits are not observable, nor do they bother the user. The device does not need to be connected to the computer. The reason is that the Arduino has the NN integrated.
Fig. 6. 65-year-old participant.
3.3 Neural Network Operation Training and validation showed an accuracy of 83%. Therefore, the proposed device can detect the inclination of the user and, for the time being, displaying the results on the computer, see Fig. 6. This means that the device detects when a person falls.
10
C. Arana Cohuo et al.
Fig. 7. A) Results in a normal posture; B) Results in rear fall; C) Results in right lateral fall; D) Results in left lateral fall and E) Results in front fall.
This is also observed in the results obtained in the loss function. It must be mentioned that the loss function is a method to evaluate how efficient the algorithm that models the data is, in the same way they are the main source assessment in modern machine learning. The loss function tells us how far away it is at a given moment, what the network offers us as output and the result that we consider to be the correct or desired one. In other words: if the deviation in the predicted value from the expected value by our model is large, then the loss function returns the highest number; if the deviation is small and much closer to the expected value, it generates a lower number. The model shows that the loss decreases as the epochs (interactions) increase, see Fig. 7. The precision of the device in posture detection was evaluated: if the device detects the correct posture change, then it scored a point; otherwise, it has zero points, that is, it does not detect the change of posture. For this, it was tested on different people regardless of gender, weight and height, see Fig. 8. There was a general configuration, this means that the device was not calibrated per person (Fig. 9).
Device for the Fall Detection in Older Adults
11
Fig. 8. Loss tests results.
Fig. 9. Percentage of precision of falls detection.
4 Discussion The objective of this research work is to design a device for posture detection focused on older adults. For this, a NN, a branch of artificial intelligence, was incorporated to classify the MPU6050 signals. The device has a minimal margin of error in precision for fall detection. Noting that the sensor must be located on the torso and adjusted, otherwise it gives erroneous readings. Low sensitivity is detected along the Z axis, in other words from back to front. For these reasons, it was considered to add the signals of the gyroscope to the NN, with this it could differentiate with greater precision the changes of postures.
12
C. Arana Cohuo et al.
It is important to mention that the sensitivity of the posture change detection system can vary depending on the body of the user biomechanics, thus making it convenient to carry out a personalized calibration prior to using the device to certify its good performance. Currently there are works that detect falls, but they use professional video cameras and artificial vision [5, 6]. These works obtained satisfactory results. Even with nonprofessional cameras, the authors used a video game camera (Kinect) [13]. Artificial vision is a branch of Artificial Intelligence widely used in various areas. The advantage of the device, compared to machine vision, is that it does not require a camera and a large device for data interpretation such as laptop or desktop computers. The downside is that the Arduino has computational limits far below computers. However, a functional NN was developed with the capabilities of the Arduino. In the work of Gupta et al. [14] and Bravo-Sánchez et al. [15], accelerometers were used. The first for the detection of falls of older adults and the second for the detection of the approach of the hand to the face. The difference with these works is the signal processing, that is, the NN involves a learning process that manages to identify noise or other artifacts, as well as adaptability. Compared to jobs [9, 10], the device is independent. The processing and identification of the posture is done by the Arduino. It is worth mentioning that, this is the first step in the development of a device that cushions the fall of older adults. Once it detects the change in posture, the system will inflate air bags to protect the neck, waist and head. Jefiza et al. [10] show good results. Adding more data to training could increase precision; so also, more people simulating movement, there would be greater variability of possible postures. According to the interactions between the hidden layers, Jefiza et al. [10], used 11 hidden layers and 30 thousand interactions. It was thought that many interactions and hidden layers would lead to overtraining of the NN, for which five layers and 500 interactions will be maintained. Maybe increase interactions to 3–5 thousand. Observing good results, as mentioned above, it is planned to add some air bags to the device to cushion the falls of older adults, for this reason a vest was used, in addition to a GSM system for sending messages. IoT was considered but it is not viable, the reason is because a home network is needed and its use is planned, both at home and on the street. Likewise, the incorporation of a system to reduce the impact of the fall detected with the developed prototype is contemplated.
5 Conclusion A device was developed in the form of a vest capable of detecting the change of posture of the user, through the use of the MPU6050 sensor, NN and Arduino. Once the corresponding tests were performed, it is concluded that the developed prototype demonstrates high accuracy, proving that it is an effective way to detect and monitor the position and inclination of the user during a loss of balance prior to a fall and display the results on a computer. These results were obtained by signal processing, i.e., NN. Furthermore, the proposed device is a more economical alternative for monitoring adults at risk of injury due to serious falls.
Device for the Fall Detection in Older Adults
13
Informed Consent. The authors obtained the informed consent of the subjects mentioned in this paper Declaration of Competing Interest. We declare that we have no conflict of interest.
References 1. Cummings, S.R., Rubin, S.M., Black, D.: The future of hip fractures in the United States. Numbers, costs, and potential effects of postmenopausal estrogen. Clin. Orthop. Relat. Res., 163–169 (1990) 2. de Alejo-Plaín, P., Roque-Pérez, L., Plaín, P.C.: Las caídas, causa de accidente en el adulto mayor. ECIMED Editorial Ciencias Médicas 59(276), 1–6 (2020) 3. Dzul-Hernández, J., García-Durán, A., Mendez, N., Argaez-Manzanero, A.: Frecuencia y tipo de fracturas de cadera en adultos mayores del Hospital General O’Horán entre 2015 y 2019 (2019) 4. Bet, P., Castro, P.C., Potin, M.A.: Fall detection and fall risk assessment in older person using wearable. Int. J. Med. Inform. 130, 1–11 (2019) 5. Chen, Y., Du, R., Luo, K., Xiao, Y.: Fall detection system based on real-time pose estimation and SVM. In: 2021 IEEE 2nd International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE 2021), pp. 1–4 (2021) 6. Bundele, M., Sharma, H., Gupta, M., Sisodia, P.S.: An elderly fall detection system using depth images. In: 5th IEEE International Conference on Recent Advances and Innovations in Engineering- ICRAIE 2020, pp. 1–4 (2020) 7. Handelzalts, S., Alexander, N.B., Mastruserio, N., Nyquist, L.V., Strasburg, D.M., Ojeda, L.V.: Detection of real-world trips in at-fall risk community dwelling older adults using wearable sensors. Front. Med. 7, 514 (2020) 8. Nguyen, H., Zhou, F., Mirza, F., Naeem, M.A.: Fall detection using smartphones to enhance safety and security of older adults at home. In: 2018 Eleventh International Conference on Mobile Computing and Ubiquitous Network (ICMU) (2018) 9. Nari, M.I., Suprapto, S.S., Kusumah, I.H., Adiprawita, W.: A simple design of wearable device for fall detection with accelerometer and gyroscope. In: 2016 International Symposium on Electronics and Smart Devices (ISESD), pp. 88–91 (2016) 10. Jefiza, A., Pramunanto, E., Boedinoegroho, H., Purmono, M.H.: Fall detection based on accelerometer and gyroscope using back propagation. In: 2017 4th International Conference on Electrical Engineering, pp. 1–6 (2017) 11. InvenSense Inc.: MPU-6000 and MPU-6050 Product Specification Revision 3.4. Sunnyvale (2013) 12. Martínez, L.A., Goddard, C.J.: Definición de una red neuronal para clasificación por medio de un programa. Revista Mexicana de Ingeniería Biomédica 22(1), 4–11 (2001) 13. Stone, E.E., Skubic, M.: Fall detection in homes of older adults using the Microsoft Kinect. IEEE J. Biomed. Health Inform. 19(1), 290–301 (2015) 14. Gupta, A., Srivastava, R., Gupta, H., Kumar, B.: IoT based fall detection monitoring and alarm system for elderly. In: 7th Uttar Pradesh Section International Conference on Electrical, Electronics and Computer Engineering (UPCON) (2020) 15. Bravo-Sánchez, G., Camas-May, R.E., Carrillo-Fierros, M.A., Cervantes-Rodríguez, R.F., Álvarez-Cervera, M.M.: Pulsera inteligente: Watch Ur Health. Memorias del 44 Congreso Nacional de Ingeniería Biomédica (2021)
Breast Cancer Detection Algorithm Using Ensemble Learning Sophia Sandoval Torres1 , Ana Paola Romero Espinoza1 , Grisel Jhovana Castro Valles1 , and Carlos Eduardo Cañedo Figueroa2(B) 1 Coordinación de ingenierías, Universidad La Salle Chihuahua, Prolongación, Lomas de
Majalca 11201, Labor de Terrazas, 31020 Chihuahua, Chihuahua, Mexico 2 Facultad de Medicina y Ciencias Biomédicas, Universidad Autónoma de Chihuahua, Circuito
Universitario S/N, 31125 Chihuahua, Chihuahua, Mexico [email protected]
Abstract. There are certain parameters in the human body that may be indicators of the presence of breast cancer, these can be assessed with different algorithms such as the Support-Vector Machine (SVM), the Naïve Bayes Algorithm (BA) and Artificial Neural Networks (ANN) to determine whether the laboratory tests are positive or not. Machine Learning (ML) has gained more uses across fields as it proposes a cost-effective classifier with versatility to be developed for any type of application, such as early breast cancer detection. This paper shows an ensemble of the algorithms previously mentioned that, based on a database consisting of 8 characteristics, can provide a high accuracy result. The highest F1 Score obtained was 78.261% from the BA, followed by the ANN’s score of 77.273% and a 72.34% from the SVM, resulting in a compositive F1 Score of 80.851%. All the data used on this article was trained using supervised machine learning techniques and variables of interest for breast cancer proliferation. Keywords: Breast cancer · Classifier · Ensemble learning
1 Introduction Breast cancer (BC) is defined as the abnormal growth of cells in the breast that can be felt as a lump. It can start in different parts of the breast, such as the lobules, ducts, and connective tissue, and can also metastasize to other parts of the body. This disease is most common amongst women, but men are also capable of developing breast cancer [1]. Cancer itself doesn’t have a particular set of causes, however, having high levels of certain substances in the body may be an indicator of its presence. Women with a higher Body Mass Index (BMI) than recommended tend to have elevated levels of insulin, a hormone that stimulates the proliferation of certain human breast cancer cells when in elevated levels, since it decreases levels of high-density lipoproteins that increase BC risk. [2]; glucose may also play a role in the proliferation of cancerous cells [3] as glucose levels rise when either leptin or Homeostatic Model Assessment of Insulin Resistance (HOMA-IR) increase [4]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 14–26, 2023. https://doi.org/10.1007/978-3-031-18256-3_2
Breast Cancer Detection Algorithm Using Ensemble Learning
15
On the other hand, having a resistance to insulin that exceeds normal limits can also be a risk factor, since insulin levels would also rise as a mechanism to beat this resistance and incorporate glucose into the cell [5]. At the same time, HOMA and adiponectin levels are inversely related and could be indicators of increased mortality in BC patients [6]. Adipokines are a group of substances that regulate numerous physiological functions, two of its most notorious proteins are leptin and resistin. Both have been shown to be a key factor in endocrine, paracrine, and autocrine evolution of breast cancer and are the direct link between obesity and BC, due to the microenvironment they create for tumor cells [7]. Also, Monocyte Chemoattractant Protein-1 (MCP-1) has been shown to provide favorable conditions for proliferation of tumorous cells, as its expression in malignant cells has a direct relation with tumor associated macrophages accumulation in the tumor area [8]. BC is often easily diagnosed, however if the lump is not of a considerable size, diagnosis may be delayed, complicating any treatment option available. As with any other disease, early treatment is key, which makes it of great importance to optimize current detection methods. Machine Learning algorithms can develop a significant role in detection, confirmation, and follow-up tests for BC, improving patients’ life chances and, therefore, recovery times. The efficiency of the algorithms mentioned on this paper has been a subject of study in BC detection by many authors, Chaurasia, Vikas and Pal, Saurabh obtained and F1 Score of 98% using the BA [9]. However, the SVM is one of the preferred models for this type of application, as it provides high precision classifications, as shown in the paper done by Raweh, Abeer et al. where F1 Scores for both the BA and SVM, were 92.2% and 94.4%, respectively [10]. Other techniques have also been used, such as the Sequential Least Squares Programming Method based on the SVM, K-Nearest Neighbors algorithm (KNN), Decision Tree (DT) and Logistic Regression (LR) for BC detection; Gupta, Madhuri and Gupta, Bharat were able to develop an ensemble algorithm with an F1 Score of 95% [11]. On the other hand, mammogram images can also be processed by ML algorithms to classify it as benign or malignant. Research done by Chand, Satish and Yadavendra, where five different algorithms were used to obtain an ensemble learning model resulted in an F1 Score of 81% for BC tumor classification [12]. While Sun, Xianhe et al. used a Regional Convolutional Neural Network with an F1 Score of 73.6% [13]. The application of machine learning to obtain a better, faster, and more precise cancer diagnosis has already been put to test by various authors; Kourou, Exarchos, et al. provide a similar focus on their paper, where based on already existing studies using the SVM, ANN and BA, the authors were able to conclude that a multidisciplinary technique including feature selection computerized algorithms provides a promising tool not only for diagnosis, but for prognosis as well [14].
16
S. Sandoval Torres et al.
2 Methods 2.1 Database The database used was obtained from the UCI Machine Learning Repository and contains 116 data rows. The characteristics listed on it were age, BMI, Glucose, Insulin, HOMA, Leptin, Adiponectin, Resistin and MCP.1, aside from a classification column, where healthy patients were identified with the number 1, and patients with BC with the number 2 [15]. The age column was not considered for the development of the algorithms, as no correlation between age and the presence of BC was found when analyzing the data. All the remaining characteristics were considered for the Naïve Bayes Algorithm and ANN, while for the SVM only resistin and leptin values were used, as they provided more relevant and consistent information. Aside from this data selection, several rows were deleted to have symmetric classes of 52 data sets each, from which only 30 were used for training and development of algorithms. 2.2 Naïve Bayes Algorithm This algorithm’s purpose is to determine whether a sample vector belongs to a class or another considering the values on each characteristic [16]. The first step consists of calculating the probability of it being positive or negative (Probneg and Probpos , respectively), using Eqs. (1) and (2). Where tneg corresponds to the probability of the sample belonging to the negative class, and tpos , to the positive class; both of which had values of 50%. tneg Probneg = tneg + tpos
(1)
tpos Probpos = tneg + tpos
(2)
The mean (x) and variance (σ 2 ) values were calculated for all characteristics on both classes. Then, with the Eq. (3) the probability of it belonging to one of the classes based on each characteristic was calculated.
1 Prob(Pos|C1, . . . , C8) = √ 2π σ 2
−(data−μ)2 2σ 2
(3)
The relation between all probabilities of each class was calculated using Eq. (4) for both positive and negative classes Pr neg,pos = Probneg,pos ∗ Prob(neg,pos|(C1,..,C8))
(4)
Then, the evidence was obtained adding both results given by the previous equation, which makes for Eq. (5). And the final equation’s (6) result indicates the final probability
Breast Cancer Detection Algorithm Using Ensemble Learning
17
of classification of the sample, meaning that this equation should be done for positive and negative Pr values. Evidence = Pr neg + Pr pos Ps =
Pr neg,pos Evidence
(5) (6)
The greater number is the one corresponding to the prediction class following the rule: Pneg > Ppos orPpos > Pneg
2.3 Artificial Neural Network Artificial Neural Networks are a subset of machine learning which are inspired by the human brain and aim to mimic signaling between biological neurons. They consist of an input layer, an indefinite number of hidden layers and an output layer, and its action is triggered by a hard limit function [17]. The ANN designed contains 8 inputs, 1 hidden layer with 20 neurons and 2 outputs (Fig. 1).
Fig. 1. ANN design with 8 inputs, one hidden layer of 20 neurons and two possible outputs.
The network’s training was done following the Levenberg-Marquardt backpropagation algorithm with the hyperparameters obtained during testing: • • • •
Learning rate: 0.01 Number of epochs: 200,000 Minimum error: 1e−59 Validation check: 1000
We selected these parameters so that the neural network would search for the minimum error, while avoiding over-fit. It should be noted that our values were experimental and that the starting values at the synaptic points were all random with each run. We chose the network that showed the lowest error, however, the minimum error was never reached, being the closest one obtained in iteration 62, after which the error began to rise, which can be seen in Fig. 3.
18
S. Sandoval Torres et al.
Before training 30 rows from each class were selected, where 90% correspond to training values, 5% to validation and the remaining 5% to testing. Obtaining the following parameters for the already trained ANN, as shown on Fig. 2
Fig. 2. Training, validation, test and global confusion matrixes obtained with a 100% accuracy.
Fig. 3. Training state. The final value of gradient coefficient was 9.2355e-24, near the ideal error of 0.
Breast Cancer Detection Algorithm Using Ensemble Learning
19
2.4 Support-Vector Machine A SVM is a supervised machine learning algorithm used for classification, where each data item is plotted and then separated into categories by a hyper-plane, usually defined as a line [18]. First, the two most significant categories information-wise were selected (resistin and leptin) and the distances between each vector from the negative class and each vector from the positive class were calculated to find the two smallest ones, and therefore find the shared vector, these are the support vectors: • S1 = [14.09,7.64,1] • S2 = [15.1248,9.1539,1] • S3 = [14.9037,8.2049,1] Where S1 and S2 correspond to the negative class and S3 to the positive one. Alphas were then calculated with equation system (7). α1 S1 ∗ S1 + α2 S2 ∗ S1 + α3 S3 ∗ S1 = −1 α1 S1 ∗ S2 + α2 S2 ∗ S2 + α3 S3 ∗ S2 = −1
(7)
α1 S1 ∗ S3 + α2 S2 ∗ S3 + α3 S3 ∗ S3 = +1 Then, Wx, Wy and Wb values were calculated with Eqs. (8), (9) and (10). ∝1 ∗ S1(1) + ∝2 ∗ S2(1) + ∝3 ∗ S3(1) = Wx
(8)
∝1 ∗ S1(2) + ∝2 ∗ S2(2) + ∝3 ∗ S3(2) = Wy
(9)
∝1 ∗ S1(3) + ∝2 ∗ S2(3) + ∝3 ∗ S3(3) = Wb
(10)
Finally, the resulting equation is as follows (11): Wx(x) + Wy(y) + Wb = 0 where: Wx = 4.677570008 Wy = −3.197271579 Wb = −42.47980654 The hyper-plane was plotted to divide both classes (Fig. 4).
(11)
20
S. Sandoval Torres et al.
Fig. 4. (+) indicates data from the negative class, while (*) represents data from the positive class.
2.5 Composite Algorithm A more complex algorithm was created using the outputs of all three previously mentioned algorithms, with the objective of determining if a data set was negative or positive based on the results given by the BA, the ANN and SVM. If two of the three results point towards the positive class, the patient is likely a BC patient, on the contrary, if at least two results are negative, the person is probably healthy [19] (Table 1 and Fig. 5). Table 1. Composite algorithm’s possible results. Bayes
ANN
SVM
Result
P_pos
A_pos
SVM_pos
Positive
P_neg
A_pos
SVM_pos
Positive
P_pos
A_neg
SVM_pos
Positive
P_pos
A_pos
SVM_neg
Positive
P_neg
A_neg
SVM_pos
Negative
P_pos
A_neg
SVM_neg
Negative
P_neg
A_pos
SVM_neg
Negative
P_neg
A_neg
SVM_neg
Negative
Breast Cancer Detection Algorithm Using Ensemble Learning
21
Fig. 5. Representative flowchart of the ensemble learning algorithm: data is processed by each algorithm separately, obtaining a positive or negative result. The three results are compared using an “if” condition. If two or more results fall in the positive category, the result will be positive. On the contrary, if two or more results are negative, the result will be negative.
3 Results In Fig. 6 the results obtained for the BA are shown, with a precision of 81.81%, accuracy of 77.27% and recall of 75%, which gives an F1 Score of 78.261%.
Fig. 6. Bayes algorithm confusion matrix, 18 true negatives and 16 true positives obtained.
22
S. Sandoval Torres et al.
Fig. 7. Bayes algorithm values graph
Figure 8 displays the confusion matrix for the ANN, where precision, accuracy and recall were all 77.27%, giving an F1 Score of 77.27% as well.
Fig. 8. Artificial Neural Network confusion matrix, with 17 true positives and negatives.
Fig. 9. Artificial Neural Network values graph
Breast Cancer Detection Algorithm Using Ensemble Learning
23
Data obtained from the SVM is shown on Fig. 10, with precision of 77.27%, accuracy of 70.45% and recall of 68%, resulting in an F1 Score of 72.34%.
Fig. 10. Support-Vector Machine confusion matrix, where true positives and negatives where 17 and 14, respectively.
Fig. 11. Support-Vector Machine values graph
For the composite algorithm the values for precision, accuracy and recall were 86.36%, 79.54% and 76%, respectively, for a composite F1 Score of 80.851%.
Fig. 12. Composite confusion matrix. A higher accuracy was achieved, where 19 data sets were classified as true negatives and 16 as true positives.
24
S. Sandoval Torres et al.
Fig. 13. Composite Algorithm values graph
As shown on Figs. 6, 7, 8, 9, 10, 11, 12 and 13, each algorithm has relatively low values of precision, accuracy, recall and F1 Score on their own, whereas the composite algorithm increases all of these metrics by the use of the three models working as a whole to process the same data sets.
4 Discussion The analysis of the metrics has shown that the accuracy values are practically identical between SVM and ANN, with 77.27%, while the accuracy of BA was higher by 4.54%. This means that the composite algorithm is the most accurate, with 86.36%. On the other hand, the accuracy of each algorithm showed similar variation, as the values were 77.27%, 70.45% and 77.27% for BA, SVM and ANN, respectively, for a composite accuracy of 79.54%. Recall was low for SVM with 68%, while BA had 75% and ANN 77.27%, giving a composite recall of 76%. All the metrics obtained conclude that the composite algorithm is the most reliable as all the results in it are higher than in the tests of singular algorithms, similar to the work [11], in which the F1 scores of the algorithms operating separately were lower than that obtained by the composite of all of them, with 95%. Table 2 shows the test metrics of various algorithms from the aforementioned paper. Table 2. Metrics obtained for all the algorithms used by Gupta, Madhuri and Gupta Bharat on their paper on BC prediction using SLSQP.
Breast Cancer Detection Algorithm Using Ensemble Learning
25
It is important to note that even though the purpose of this ensemble is to support medical professionals in diagnosis, it can’t operate on its own and its results should be confirmed by a specialist.
5 Conclusions Ensemble learning provides a different approach to the diagnosis of certain diseases and could represent more reliability than using a single algorithm alone. The F1 score obtained in this work could be greatly improved by using a larger database to train these algorithms. The results obtained were obtained by analyzing only one database [15], so it is intended to perform a set of tests using the same methodology with other databases, in order to complement the results obtained. At the same time, we intend to conduct a study collaborating with local hospitals to test our development in order to make improvements in the data classification process. Likewise, it should be noted that our algorithms work with characteristics, although, in comparison with other works it is in a medium ranking, if compared with works that work with image processing, our algorithm shows an advantage, since the metrics obtained are above itself, it does not require such a specialized computer equipment to run the algorithm.
References 1. Centers for Disease Control and Prevention. https://www.cdc.gov/cancer/breast/basic_info/ what-is-breast-cancer.htm. Accessed 26 May 2022 2. Rose, D., Vona-Davis, L.: The cellular and molecular mechanisms by which insulin influences breast cancer risk and progression. Endocr. Relat. Cancer 19, 228–233 (2012) 3. National Library of Medicine. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6567675/. Accessed 26 May 2022 4. Patrìcio, M., Pereira J. et al.: Using Resistin, glucose, age and BMI to predict the presence of breast cancer, 2–8 (2018) 5. Zacharzewski C., Tibolla, M., et al.: Obesidad y resistencia a la insulina como factores de riesgo en el cáncer de mama, pp. 5–6 (2016) 6. Duggan, C., Irwin, M., et al.: Associations of insulin resistance and adiponectin with mortality in women with breast cancer, pp. 4–5 (2011) 7. National Library of Medicine. https://pubmed.ncbi.nlm.nih.gov/31637624/#:~:text=Adipok ines%20exert%20independent%20and%20joint,dysfunction%20characterized%20by%20c hronic%20inflammation. Accessed 26 May 2022 8. Saji, H., Koike, M., et al.: Significant correlation of monocyte chemoattractant protein-1 expression with neovascularization and progression of breast carcinoma, pp. 2–4 (2001) 9. Chaurasia V., and Pal, S,: Performance analysis of data mining algorithms for diagnosis and prediction of heart and breast cancer disease, pp. 11–14 (2014) 10. Raweh, A., Nassef, M. and Badr, A.: A hybridized feature selection and extraction approach for enhancing cancer prediction based on DNA methylation, pp. 11–12 (2017) 11. Gupta, M. and Gupta, B.: An ensemble model for breast cancer prediction using Sequential Least Squares Programming Method (SLSQP), pp. 1–3 (2018)
26
S. Sandoval Torres et al.
12. Yadavedra and Chand, S.: A comparative study of breast cancer tumor classification by classical machine learning methods and deep learning method, pp. 7–10 (2020) 13. Sun, X., Cai, D., et al.: Efficient mitosis detection in breast cancer histology images by RCNN, pp. 3–4 (2019) 14. Kourou, K., Exarchos, T., Karamouzis, M., Fotiadis, D.: Machine learning applications in cancer prognosis and prediction. Comput. Struct. Biotechnol. J. 13, 1–10 (2014) 15. UCI Machine Learning Repository. https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+ Coimbra. Accessed 01 May 2022 16. Java T Point. https://www.javatpoint.com/machine-learning-naive-bayes-classifier. Accessed 27 May 2022 17. IBM Cloud Education. https://www.ibm.com/cloud/learn/neural-networks. Accessed 27 May 2022 18. Towards Data Science. https://towardsdatascience.com/support-vector-machine-introduct ion-to-machine-learning-algorithms-934a444fca47. Accessed 27 May 2022 19. García H, and Cañedo C.: Diseño de algoritmo compuesto por Machine Learning y un modelo probabilístico para la detección de diabetes
Graph Analysis of Functional Connectivity Rs-FMRI in Healthy and Epileptic Brain Using Visibility Algorithm Rosa Victoria Villa Padilla1(B) , Katya Rodr´ıguez V´azquez1 , M´ onica V´ azquez Hern´andez1 , Bayron Alexander Sandoval Bonilla2 , and Josafat Jonathan S´ anchez Due˜ nas3 1
2
Instituto de Investigaciones en Matem´ aticas Aplicadas y en Sistemas de la UNAM, Coyoac´ an, 04510 Cd. M´exico, Mexico victoria [email protected] Departamento de Neurocirug´ıa, CMN Siglo XXI del Instituto Mexicano del Seguro Social, Cuauht´emoc, 06720 Cd. M´exico, Mexico 3 Laboratorio de Cognici´ on y Acci´ on comparada, Facultad de Psicolog´ıa UNAM, Coyoac´ an, 04510 Cd. M´exico, Mexico
Abstract. In order to analyze brain functional connectivity (FC) in humans, this work focuses on resting state functional magnetic resonance imaging (Rs-fMRI) studies to measure functional connectivity within default-mode network (DMN). The objective of the present study is to compare functional connectivity networks in healthy subjects with a drug-refractory epilepsy patient. Since the FC has a dynamic nature, evidence of variations over time and pathology are explored. The FC graph in Rs was obtained using the visibility algorithm. Based on the resulting graph, it was possible to determine how each voxel is connected to others and it is also possible to identify the hubs in healthy subjects and compare them with the DMN and how connectivity networks are altered in epilepsy patients.
Keywords: Functional connectivity Graph · Epilepsy · BOLD
1
· Rs-fMRI · Visibility algorithm ·
Introduction
Functional magnetic resonance imaging (fMRI) utilizes changes in blood oxygen level-dependent (BOLD) signals to estimate neural activity patterns. This work focuses on spontaneous modulations in BOLD signals during “resting state” or periods in which brain functioning depends on no voluntary activity and is given no stimulus. The temporally correlated spatial patterns are well-known as functional activity maps [1,2]. In BOLD, when two or more events occur at the same time, answers are added [3]. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 27–36, 2023. https://doi.org/10.1007/978-3-031-18256-3_3
28
R. V. Villa Padilla et al.
The default mode network (DMN) are subregions that have some kind of interaction between those, particularly the medial prefrontal cortex (MPFC), posterior cingulate cortex (PCC), and left and right temporal-parietal junction (left- and right-TPJ). These regions are activate when the brain is at rest, and decrease when engaged in a goal-directed task. In healthy subjects, the DMN are the same but some of these networks are modified due to some pathologies [3,5–7]. The mathematical perspective treats the connectome as a graph of interactions among brain areas. Nodes in the graph are abstract representations of brain areas and edges represent pairwise relationships between nodes [4]. In this work, the visibility algorithm was used to transform time series into a graph [8]. This work was focused on an axial plane section in coordinates cortex Z = 26 to show the DMN, as it is shown in Fig. 1.
Fig. 1. Axial plane xy in z = 26. Functional connectivity map of the DMN.
Epilepsy is a chronic neurologic disorder which involves sudden, temporary, bursts of electrical activity in the brain that can cause involuntary changes in body movement or function, sensation, behavior or awareness. It has a prevalence of approximately 1% in the world’s population and approximately one-third of people with epilepsy do not respond adequately to antiepileptic drugs, hence the importance of creating tools that can help neurologists to improve the diagnosis [9,10].
2
Methodology
In this work, the database of multiband multi-echo imaging of simultaneous oxygenation and flow time series for resting state connectivity is used and described
Graph Analysis of Functional Connectivity Rs-FMRI
29
in [6]. This database has seven healthy adult subjects (four male, three female, mean age = 35.0 ± 13.6 years, age range 23–58 years) in resting state. Other works have shown that in healthy individuals in resting state fMRI networks, are generally reproducible across subjects. Therefore, preprocessing of resting-state fMRI data was performed using Matlab (MathWorks Inc.) and the toolboxes SPM12 and CONN. For each subject acquired, 200 images were first corrected for the differences in image acquisition time between slices, and then structural and functional realigned and head-motion correction. Head motion parameters were computed by estimating translation in each direction and angular rotation on each axis for volumes each one of 200 images. The realigned functional volumes were then spatially normalized to fit to the template created by the Montreal Neurological Institute (MNI) using the normalization parameters estimated by T1 structural image (voxel size [3, 3, 3]). Then, the datasets were smoothed using Gaussian kernel (FWHM = 6 mm). Model connectivity depends on neurophysiological time series. In order to evaluate the performance of the proposed algorithm in this paper to obtain resting state functional connectivity networks, two algorithms were used in data processing: 1) the visibility algorithm that is a model free method, this means it is a method that searches for a general pattern of connectivity without any pre-assumption, and 2) the model-dependent based on the activation DMN map, where the predefined ROI is called a seed [7]. Resting-state networks were extracted using the visibility algorithm [8]. Changes in the BOLD signal are interpreted as a time-series at each voxel and then voxel by voxel were transformed into time series. Thus, each signal is rectified and all values that form the signal are added, this procedure is repeated through the matrix for each voxel until obtaining a matrix, where the regions with more functional activity are visibles. The visibility algorithms are a family of geometric criteria which define different ways of mapping an ordered series into a graph of N nodes. The visibility algorithm takes two arbitrary data values (ta , ya ) and (tb , yb ) as input, the algorithm transform these points in nodes and consequently will become two connected nodes of the associated graph, if any other data (tc , yc ) placed between them fulfills the constraint. This is shown in Eq. 1: yc < yb + (ya − yb )
tb − tc tb − ta
(1)
The variant of the algorithm used to acquire input data from left to right for each row is called Horizontal Visibility Graph (HVG), whose topological properties have been shown to be analytically tractable.
30
R. V. Villa Padilla et al.
Each signal time is mapped to a node (at the same order that was acquired), if one can draw a straight line between a node “a” and node “b” to joining these points and does not intersect any intermediate node, then these are connected. The summarize of this network can be shown as an adjacency matrix. Brain regions are equated with nodes and their correlations detected by functional connectivity are showed by the edges between nodes, the node’s degree is defined by the number of connections between regions. Functional networks with the quantity of edges connected to a single vertex were then estimated.
3
Results
The nodes with highest adjacency are called hubs; one of these was chosen to show the links between it and other nodes with relevant values of adjacency. The nodes with higher adjacency are located inside the DMN regions, demonstrating that using the visibility algorithm to obtain DMN in healthy patients without the need of activation maps is feasible. The images shown in Figs. 2 and 3 correspond to a healthy subject 4 taken from database [11]. In Fig. 2, a hub (node) is shown in emerald green color (inside the circle and indicated with the arrow) and blue voxels represent other nodes which functional activity correlates with the hub. In Fig. 2, the hub in emerald color is considered the vertex and the regions with high correlation to it are shown as edges. After applying the visibility algorithm to subject 4 taken from [11], the visibility graph was obtained. The density of links of complete network is high; the complete brain network has more than 9000 nodes, and also is complex, shown in a 2D graph. Therefore, in Fig. 3(a), the links are represented in yellow lines, in (b) the node and their links are plotted on DMN which is shown in red color, (c) the complete network of only one node with MNI coordinates (4, −90) is observed as well as all its connections. The nodes represent areas with brain activity and the links indicate what areas synchronized with the activity node selected. The brain networks shown are undirected networks. In Fig. 3(c), the graph is shown and the nodes which corresponded to DMN are indicated in a green circle, these nodes are the nodes with more adjacency into the graph. The algorithm of visibility has been showed to be able to obtain the DMN in healthy subjects as is shown in Fig. 3 and Table 1, and the last can indicate which voxels are linked with others as is shown in Figs. 2, 3 and 4. In visibility algorithm case, the analysis was done voxel by voxel and in the CONN case, the analysis was performed using a region of interest (ROI).
Graph Analysis of Functional Connectivity Rs-FMRI
31
Fig. 2. a) In a reddish scale showing the areas with functional activity. The zones with more activity are associated with yellow color and the zones with less activity are associated with red color, the zones in blue color are voxels that are connected with the voxel in emerald green. These connections belong to subnetwork that was obtained when a voxel with more adjacency is chosen from the graph obtained with the visibility algorithm. b) The same information but now in grayscale, the DMN is shown with areas in red color, the blue voxels are connected with the emerald green voxel as is show in Fig. 3.
Fig. 3. The map of connections obtained with subject 4 who is a healthy person. The functional connectivity subnetwork obtained using the visibility algorithm. The emerald green voxel that was chosen as a voxel of study has connections with the voxels in blue color. (a) The lines in yellow were made to illustrate the connections between regions; (b) The DMN is show in red color and the voxels in emerald green and the voxels in blue are inside of DMN regions; (c) the nodes which correspond to DMN and are also the nodes with more adjacency, are indicated inside a green circle.
32
3.1
R. V. Villa Padilla et al.
Case Study
The functional connectivity has the ability to detect functional brain changes. The default mode network had shown to be disrupted to identify some diseases like autism, spectrum disorders, schizophrenia, Alzheimer’s disease, attention deficit, hyperactivity disorder and epilepsy [5,12–16]. In Fig. 4, the lack of functional connectivity means the brain activity associated with the default mode network decreases because the brain starts to focus on a task. The electrical activity associate with the seizure alters functional connectivity at different time scales. In support, the notion that resting state networks reconfigure themselves in a demand specific task or behavior, it means the brain has a flexible reconfiguration. The subnetwork topology for a woman with refractory epilepsy is shown in Fig. 4, the voxel chosen to compare health and epilepsy subject is located in posterior cingulate cortex. Deserving attention in the case of Fig. 4, the activity is lower compared with healthy subjects and in this case, there are low functional connections in the medial prefrontal cortex or lateral parietal cortex. It is clearly observed that the case study has less connectivity between brain’s regions associated with resting state and the hypothesis is the connections decrease because the brain starts to focus on a task (seizure). In this case, the node with higher adjacency is outside the DMN regions and the values of adjacency matrix are lower in epilepsy patient than in health subjects. However, the node was chosen and their links are displayed. The DMN regions are indicated with green circles and labels. In the case of epilepsy patient nodes with more adjacency are outside the DMN regions; on the contrary, in the case of healthy subjects, the nodes with more adjacency always are inside DMN regions.
Fig. 4. The functional connectivity subnetwork obtained from a patient with epilepsy. (a) The functional connectivity subnetwork obtained using the visibility algorithm, (b) The functional connectivity subnetwork over DMN (c) the DMN regions are shown in green.
Graph Analysis of Functional Connectivity Rs-FMRI
3.2
33
Seed Regions of Interest
The regions of interest (ROI) expressed in MNI coordinates were used to compare the seed-based functional connectivity those defined as DMN regions: medial prefrontal cortex (0, 52, −6), left lateral parietal ( −48, −62, 36), posterior cingulate cortex (0, −53, 26), and right lateral parietal (46, −62, 32) [7]. In Fig. 5, it is possible to observe that the images obtained with CONN are same or similar to those obtained with the visibility algorithm shown in Figs. 2 and 3, in both images the regions associated with DMN are clearly recognizable and match with seed regions.
Fig. 5. Healthy subject functional connectivity map of the DMN.
34
R. V. Villa Padilla et al.
In Fig. 6, it can observe that the correlation of the seed medial prefrontal cortex and the map activation obtained with CONN is low due to the brain starts to focus on a task and the functional network lost connections therefore it seems not exist activity in this region.
Fig. 6. Women with epilepsy functional connectivity map of the DMN.
The DMN nodes coordinates are shown in Table 1 that summarize the comparison between the coordinate (x,y) calculated with the visibility algorithm and the toolbox CONN, when z = 26.
Graph Analysis of Functional Connectivity Rs-FMRI
35
Table 1. MNI Coordinates of all subjects under study. Coordinates (x, y) in z = 26 Subject
MPC
Healthy
CONN
Visibility algorithm
LP(R) CONN
2
(11, 35)
(14, 66)
3
(9, 59)
(−4, 54)
4
(11, 54)
(10, 66)
(−44, −77) (−42, −78) (−11, −68) (−16, −72) (48, −66) (58, −48)
5
(7, 60)
(12, 68)
(−42, −76) (−42, −80) (4, −62)
(6, −60)
(48, −72) (42, −74)
6
(−3, 56) (0, 60)
(−40, −78) (−42, −78) (16, −69)
(4, −80)
(61, −56) (62, −52)
7
(0, 41)
(−4, 42)
Epileptic (−6, 44) (6, 62)
PCC Visibility algorithm
CONN
LP(R) Visibility algorithm
CONN
Visibility algorithm
(−37, −73) (−38, −72) (10, −61)
(10, −62)
(50, −67) (50, −66)
(−42, −75) (−48, −72) (−9, −61)
(2, −54)
(49, −68) (50, −66)
(−42, −76) (−38, −80) (2, −54)
(0, −56)
(52, −66) (54, −66)
(−46, −68) (−52, −66) (6, −65)
(6, −62)
(55, −67) (54, −64)
As seen from Table 1, the DMN coordinates of health subjects are similar; therefore the comparison of results is acceptable.
4
Discussion
The major advantage of the visibility algorithm is that it is a model free method. It is a method that searches for a general pattern of connectivity without any pre-assumption. The spatial resolution of the visibility algorithm is better than with seed analysis, because the seed analysis depends on ROI size that is always bigger than one voxel. The goal of this study was to determine the feasibility of using the visibility algorithm to detect resting-state connectivity. The resting state functional connectivity in epilepsy patients compared with the DMN in healthy subjects, diminished as a result of the activity starting to focus on the seizure propagation. The visibility algorithm could be used to do a first classification between health and epilepsy subjects, because the connection maps show a clear differentiation between both groups. It could be used by neurologists as a tool to help them in the diagnosis. Also, it could help to evaluate the progress of a treatment used to reduce seizures in epilepsy patients.
5
Conclusions
The visibility algorithm is a novel, simple and easy-to-apply technique that was not previously used for the study of the human brain. DMN was robustly detected using a visibility algorithm. Graphical representations of connectomes contain a wealth of information about brain functional connectivity which can be described using a variety of nodes. Studying functional connectivity in brain patient networks can provide new insights about how the organization of the human brain changes dynamically in neurological pathologies and how brain areas anatomically separated can have the same activation patterns and how these patterns can change under the pathology effect.
36
R. V. Villa Padilla et al.
Acknowledgments. Rosa Villa thanks the financial support of CONACyT.
References 1. Paz Guti´errez, J., Salgado, P., G´ omez Llata, S.: Utilidad de la t´ecnica bold de resonancia magn´etica funcional en los tumores intracraneales de pacientes prequir´ urgicos. Archivos de Neurociencias (Mex) 12(3), 152–161 (2007) 2. Ogawa, S., Lee, T.M., Nayak, A.S., Glynn, P.: Oxygenationsensitive contras in magnetic resonance imaging of rodent brain at high magnetic fields. Magn. Reson. Med. 14, 68–78 (1990) 3. Craddock, R.C., et al.: Imaging human connectomes at the macroscale. Nat. Methods 10(6), 524–539 (2013) 4. Vergun, S., et al.: Classification and extraction of resting state networks using healthy and epilepsy fMRI data. Front. Neurosci. 10, 440 (2016). https://doi.org/ 10.3389/fnins.2016.00440 5. Guye, M., Bettus, G., Bartolomei, F., Cozzone, P.J.: Graph theoretical analysis of structural and functional connectivity MRI in normal and pathological brain networks. Magn. Reson. Mater. Phys. Biol. Med. 23, 409–421 (2010). https://doi. org/10.1007/s10334-010-0205-z 6. Raichle, M.E., MacLeod, A.M.: A default mode of brain function. Proce. Natl. Acad. Sci. 98(2), 676–682 (2001). https://doi.org/10.1073/pnas.98.2.676 7. Ghaemmaghami, P.: Functional connectivity in default mode network during resting state: an evaluation of the effects of data pre-processing. Master’s thesis, University of Trento, Italy (2013). https://doi.org/10.48550/arXiv.1603.01077 8. Lacasa, L., Luque, B., Ballesteros, F., Luque, J., Nu˜ no, J.C.: From time series to complex networks: the visibility graph. Proc. Natl. Acad. Sci. 105(13), 4972–4975 (2008) 9. Definition of Epilepsy 2014. International League Against Epilepsy. https://www. ilae.org/guidelines/definition-and-classification/definition-of-epilepsy-2014 10. Vilches,J.L.M.: Epilepsia Refractaria: Conceptos fundamentales y aspectos cl´ınicos. Revista Chilena de Epilepsia. A˜ no 19(1) (2019). ISSN 0717-5337 11. Data base of multiband multi-echo imaging of simultaneous oxygenation and flow timeseries for resting state connectivity (2017). https://openneuro.org/datasets/ ds000216/versions/00001 12. Gonen, Ofer M.: Resting-state functional MRI of the default mode network in epilepsy. Epilepsy Behav. 111, 107308 (2020) 13. Qin, L., Jiang, W.: Alterations Functional connectivity in temporal lobe epilepsy and their relationships with cognitive function: a longitudinal resting-state fMRI study. Front. Neurol. 11, 625 (2020). https://doi.org/10.3389/fneur.2020.00625 14. Abela, E., Rummel, C., Hauf, M., Weisstanner, C., Schindler, K., Wiest, R.: Neuroimaging of epilepsy: lesions, networks, oscillations. Clini. Neuroradiol. 24(1), 5–15 (2014). https://doi.org/10.1007/s00062-014-0284-8 15. Rubinov, M., Sporns, O.: Complex network measures of brain connectivity: uses and interpretations. Neuroimage 52(3), 1059–1069 (2010) 16. Wang, S., Tepfer, L.J., Taren, A.A.: Functional parcellation of the default mode network: a large-scale meta-analysis. Sci. Rep. 10, 16096 (2020). https://doi.org/ 10.1038/s41598-020-72317-8
Imagined Speech Recognition in a Subject Independent Approach Using a Prototypical Network Alan Hernandez-Galvan1 , Graciela Ramirez-Alonso2(B) , Javier Camarillo-Cisneros1 , Gabriela Samano-Lira1 , and Juan Ramirez-Quintana3 1 Laboratorio de F´ısica Qu´ımica Computacional, Facultad de Medicina y Ciencias Biom´edicas, Universidad Aut´ onoma de Chihuahua, 31125 Chihuahua, Mexico 2 Facultad de Ingenier´ıa, Universidad Aut´ onoma de Chihuahua, 31125 Chihuahua, Mexico [email protected] 3 Laboratorio PVR, IT. Chihuahua/Tecnol´ ogico Nacional de M´exico, 31200 Chihuahua, Mexico
Abstract. Brain-computer interface (BCI) systems have gained significant interest given the different biomedical applications in which they can be used to help disabled individuals to communicate or control external devices. Imagined speech is related to BCI systems controlled only by thinking about a vowel, phoneme, or word without any physical movement. In this paper, a Prototypical Network approach, named Proto-imEEG, is presented for the automatic classification of seven imagined phonemic/syllabic prompts and four imagined words by analyzing EEG data of the KaraOne dataset. The Prototypical Network is selected because of its ability to learn from a few samples, a common issue in EEG data. The embedding function of our Prototypical Network is based on a 1D-convolutional layer and bidirectional recurrent networks. The average classification accuracy achieved by Proto-imEEG is 92.04% and 96% by using a Long-Short Term Memory or a Gated Recurrent Unit, respectively, with an average inference time of 0.2 s. These results demonstrate superior performance to state-of-the-art methods in classifying the eleven classes of the KaraOne dataset. As far as we know, this is the first time that a Prototypical Network approach is used in imagined speech classification tasks. Keywords: Imagined speech dataset · EEG data
1
· Prototypical networks · KaraOne
Introduction
Language constitutes an essential aspect of communication involving verbal or visual information [4]. Brain-Computer Interface (BCI) is a technology developed to establish communication between the human brain and computers that c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 37–45, 2023. https://doi.org/10.1007/978-3-031-18256-3_4
38
A. Hernandez-Galvan et al.
does not require a muscle intervention to issue commands [1]. Different areas such as medical, robotics, telepathy systems, entertainment, and military are highly interested in developing accurate BCI applications. In this work, we consider imagined speech tasks where the BCI is used to establish a communication path that connects subject and the computer by analyzing the brain activity without using physical movements of the subject. This kind of application is mainly intended to help patients affected by the locked-in syndrome in medical therapy, including people with amyotrophic lateral sclerosis (ALS), spinal cord injury, stroke, and other serious neuromuscular diseases or injuries [7]. For these patients, the ability to communicate with others and the possibility to control external devices on their own would improve their quality of life, increasing their independence and self-esteem. A common approach used to acquire the input signals that the BCI will process is by using non-invasive electroencephalography (EEG). There are different open-access datasets used to compare the strategies that the research community has developed in imagined speech tasks. DaSalla [3] is a popular dataset that considers the signals of the imagined pronunciation of two vowels and a control state. Nguyen [8] proposed a dataset composed of two long words, three vowels, and three short words. Nieto presents in [9] a database that considers four imagined Spanish words. KaraOne is one of the most complete databases that includes the imagined speech of seven phonemic/syllabic prompts and four words from 14 subjects [14]. Some recent publications that have used the KaraOne dataset are presented below. Bakhshali et al. [2] analyzed with the k-NN algorithm the Riemannian distance between the correntropy spectral density matrices from EEG channels. The reported results show 90.25% of average accuracy for the recognition of four English words in a pair-way manner. The KaraOne dataset was partitioned in 80% for training data and the rest for testing. Datta and Boulgouris [4] also use the KaraOne dataset but only reports results related to the recognition of the grammatical class verb or noun. The authors considered spatial, temporal, and spectral properties of the EEG data in three different Convolutional Neural Networks (CNN) that subsequently combined them at the classification stage. The EEG signals were separated according to the region where imagined speech is processed. Their proposal was trained in a subject-independent manner with nine subjects and tested with a different one; this strategy is defined as Leave-One-Subject-Out (LSO), achieving an average accuracy of 82.2%. Mini, Thomas, and Gopikakumari [7] extracted features of the EEG signals based on the Linear Predictive Cepstral Coefficients (LPCC), Sequency Mapped Real Transform and Mel Frequency Cepstral Coefficients (MFCC), in conjunction with Principal Component Analysis. Statistical parameters of the previous algorithms were processed with a multi-layer perceptron model achieving an average accuracy of 77.37% in a multiclass classification scheme considering the seven phonemic/syllabic prompts and four words.
Imagined Speech Recognition
39
Rusnac and Grigore [11] proposed a method based on the Mel-Cepstral Coefficients (MFCC) and a CNN to identify the seven types of phonemic/syllabic prompts, achieving an average accuracy of 24.19%. Even when the previous publications used the same dataset, they were validated differently. For example, some of them performed a five binary classification task, whereas others performed a multi-class classification, binary word classification, word and verb classification, or only syllables classification. Therefore, we consider that classifying the seven phonemic/syllabic prompts and four words in a subject-independent manner is the most challenging task but, at the same time, is the closest evaluation to the reality if the BCI application is intended to be used in medical applications. This article, presents a Prototypical Network approach to classify the EEG signals of the KaraOne dataset related to imagined speech tasks. Our proposal, named Proto-imEEG, is based on Prototypical Networks that are frequently used when few data are available for training [6], a common issue when working with medical data. As far as we know, this is the first time than a Prototypical Network approach is implemented in the classification of imagined speech even when they are widely used in different classification tasks [5]. As a pre-processing stage, the Dual Tree Complex Wavelet Transform (DTCWT) is implemented to extract the β frequency band of the EEG data that will be analyzed by ProtoimEEG.
2 2.1
Methodology KaraOne Dataset
The KaraOne dataset combines three modalities of data: EEG, face tracking, and audio [14]. Our proposal only uses the EEG signals recorded with a 64 channels Neuroscan Quick-cap considering the 10–20 system in the electrode placement. The EEG data were recorded using the SynAmps RT amplifier sampled at 1 kHz. In the acquisition stage, four woman and eight men (27.4± 5 years) imagined and vocalized seven phonemic-syllabic prompts (/iy/, /uw/, /piy/, /tiy/, /diy/, /m/, /n/) and 4 words (pat, pot, knew and gnaw ), one at a time. Our experiments only use the EEG data of all imagined tasks, and similar to [14], we excluded the data of four subjects because of problems in the acquisition process. The number of trials for subject is different, but each of them has at least 12 trials per class. In our implementation, 10 of the 64 channels of imagined speech were selected based on [14]: T7, FT8, CP3, CP5, C5, CP3, P3, CP5, C3, and CP1. 2.2
Dual Tree Complex Wavelet Transform (DTCWT)
The Dual Tree Complex Wavelet Transform (DTCWT) calculates the complex transform of a signal using two separate Discrete Wavelet Transform (DWT) decomposition. Two key advantages of DTCWT over DWT are shift-invariance
40
A. Hernandez-Galvan et al.
Fig. 1. EEG decomposition produced with DTCWT to select the β frequency band considering the /uw/ phoneme, subject MM05, channel 0.
and directional selectivity. These advantages are traduced in that the square magnitude of a given complex wavelet coefficient gives as result an accurate measure of spectral energy at a particular location in space, scale, and orientation, reducing several of the artifacts of the DWT that may affect the reconstruction of the signal [12]. In this work, the β frequency band was selected by using the DTCWT decomposition. Figure 1 shows the DTCWT decomposition of the EEG signal of subject MM05 when imagining the speech of the /uw/ phoneme. 2.3
Prototypical Networks
Learning from fewer data points is called few-shot learning or k-shot learning, where k represents the number of data points in each of the classes in the dataset [10]. Depending on the classes we want to identify, it is defined the n−way term, that is, n − way means the number of classes we have in our dataset. In few-shot learning, the model is trained with a support set and tested with a query set. In this type of learning, the training is performed in episodic fashion. During each episode, few data points are sampled from the dataset to prepare the support and query sets. Then, over a series of episodes, the model will modify its weights based on different and small datasets [5]. A Prototypical Network is a type of few-shot learning algorithm that creates a prototypical representation of each of the classes that the network must identify. A query point will be classified to the nearest prototype. In our implementation, a 1D Convolutional layer and Bidirectional Recurrent Neural Networks (Bi-RNN) are used as a feature extractor module that learns the embeddings of each input data. A class prototype is considered the mean embeddings of data points in the class. A query point will be analyzed by the feature extractor module, and
Imagined Speech Recognition
41
its embedding information is compared against each prototype. The distance between the class prototype and query point embedding will be calculated, and the nearest class will be assigned to the query point. Our proposal uses two feature embedding functions. One is based on a Bidirectional Long Short-Term Memory (Bi-LSTM) layer, and the other in a Bidirectional Gated Recurrent Unit (Bi-GRU) layer. The use of these types or RNNs was considered because of the accurate results reported in [13] with the analysis of EEG data for Epileptic Seizure Recognition. In both implementations of Proto-imEEG, a 1D-CNN is considered as the input layer whose configuration consists on a kernel size = 3, padding = 1, dilation rate = 2, 1D-batch normalization, ReLu activation function, and a 1DMaxPool layer. The Bi-LSTM and Bi-GRU configurations have 2 layers followed by a flatten connection. The learning procedure of the Prototypical Network is presented in Algorithm 1. Algorithm 1. Prototypical Networks Input: Training set Dtr , learning rate lr. Output: Model parameter θ. →Bi-RNN Embedding function fφ is 1D-CNN − Prepare support set S Prepare query set Q Randomly initiate θ; while not done do Generate episodes; for each episode do for each class do Class Prototype ← S1e (xi ,yi )∈Se fφ (xi ) ; end for for each query sample do Query Prototype ← fφ (xQ i ) ; Predict label yˆi based on Euclidean Distance; end for end for Evaluate loss on query samples: l l; Update model parameters θ ← θ − lr∇θ end while return θ
Figure 2 shows an approximation of Proto-imEEG by considering a 1-shot (EEG data of the 10 channels) and 3-way configuration. In all our experiments, all the classes of the KaraOne dataset were considered, that is, an 11-way configuration, but for illustrations purposes, Fig. 2 shows only a 3-way representation. The results reported in Table 1 present the specific k-shot/n-way configuration used in Proto-imEEG.
42
A. Hernandez-Galvan et al.
Fig. 2. Block diagram of Proto-imEEG considering a 1-shot and 3-way configuration. The specific k-shot/n-way configurations used are reported in Table 1.
3
Experimental Settings and Results
The PyTorch framework was used to implement our model (Proto-imEEG) using an NVIDIA RTX 3080Ti GPU with 12 GB of memory, an Intel Core i9-10850K and 64 GB of RAM. The Adam solver with a learning rate of 1e−5 was used in all our experiments. The training of Proto-imEEG uses a subject-independent approach with the Leave-One-Subject-Out and Leave-One-Trial-Out strategies for testing. Figure 3 shows the confusion matrices of the Leave-One-Subject-Out (LSO) configuration using a particular episode result. In this case, Subject MM05 was used for the evaluation. Figure 3(a) depicts the results of the Bi-LSTM and Fig. 3(b) the results of the Bi-GRU implementations, respectively. These results correspond to the use of a 7-shot/11-way configuration for episodic training. Subject MM05 has in total 15-shot, in the confusion matrices results, 8-shot were used for query. Figure 4 shows the confusion matrices of the One-Trial-Out (LTO) configuration. Figure 4(a) presents the results of the Bi-LSTM and Fig. 4(b) the results of the Bi-GRU implementation, respectively. These results correspond to using a 4-shot/11-way configuration for the episodic training. In this case, 4-shot was used in the query. The average inference time achieved by Proto-imEEG is 0.2 s. Table 1 presents a comparative performance by considering the average accuracy results reported by recent methods. As previously stated, there is no standardization in the evaluation methodology. For example, [2] only performed a binary classification of imagined words, [4] only classified imagined words as
Imagined Speech Recognition
43
Fig. 3. Confusion matrices of Proto-imEEG by using (a) Bi-LSTM and (b) Bi-GRU configurations with the Leave-One-Subject-Out strategy, 7-shot/11-way in episodic training.
Fig. 4. Confusion matrices of Proto-imEEG by using (a) Bi-LSTM and (b) BiGRU configurations with the Leave-One-Trial-Out strategy, 4-shot/11-way in episodic training.
verbs or nouns by using three CNNs. The most complete evaluation [7], achieved an accuracy of 77.37%, but the methodology considers the ensemble and combination of statistical features of four different algorithms. The average results of Proto-imEEG are 96% and 92.04% when using as feature extractors the 1DCNN → Bi-LSTM or 1D-CNN → Bi-GRU, respectively. These results surpass all other methods, and there is no need to extract different features of the raw EEG signal, only the β frequency band with the DTCWT algorithm.
44
A. Hernandez-Galvan et al. Table 1. Accuracy results of recent methods and our proposals. Methods
Task
Avg. Acc.
Method of [2]
4 words pair-way manner
90.25%
Method of [4]
verb or noun, LSO
82.2%
Method of [7]
4 words and 7 phonemic-syllabic 77.37%
Method of [11]
7 phonemic-syllabic
Proto-imEEG
4 words and 7 phonemic-syllabic 96%
24.19%
Bi-LSTM 7-shot/11-way LSO
4
Proto-imEEG
4 words and 7 phonemic-syllabic 92.04%
Bi-GRU 4-shot/11-way
LTO
Conclusions
This paper presents the implementation of a Prototypical Network adapted to automatically classify the four imagined words and seven imagined phonemic/syllabic prompts of the KaraOne dataset. Our proposal is named Proto-imEEG, and analyzes the β frequency band of the EEG data. Prototypical Networks are commonly used on classification tasks with few data available, a common issue in medical applications. The feature extractor module used to learn the embeddings of the EEG data is based on a 1D-CNN and Bi-RNN, particularly, Bi-LSTM and Bi-GRU. These configurations were not implemented in previous works, and the results achieved by Proto-imEEG surpass all other recent methods. The best average accuracy was obtained when using the 1D-CNN in combination with the Bi-LSTM. It is worth noting that classifying the eleven classes by only one method is not commonly reported in the literature because it is a difficult task to perform. However, we consider that this kind of evaluation should be the standard in BCI systems where it is desirable to perform different control actions by the same method. The average inference time achieved by Proto-imEEG is 0.2 s, demonstrating that a real-time application of our proposal can be easily achieved. In the future, we will further our research by considering different k-shot/nway configurations and combining EEG data of other datasets to test the robustness of Proto-imEEG.
References 1. Abdulkader, S.N., Atia, A., Mostafa, M.S.M.: Brain computer interfacing: applications and challenges. Egypt. Inform. J. 16(2), 213–230 (2015). https://doi.org/ 10.1016/j.eij.2015.06.002 2. Bakhshali, M.A., Khademi, M., Ebrahimi-Moghadam, A., Moghimi, S.: EEG signal classification of imagined speech based on Riemannian distance of correntropy spectral density. Biomed. Signal Proces. Control 59, 101899 (2020). https://doi. org/10.1016/j.bspc.2020.101899
Imagined Speech Recognition
45
3. DaSalla, C.S., Kambara, H., Sato, M., Koike, Y.: Single-trial classification of vowel speech imagery using common spatial patterns. Neural Netw. 22(9), 1334–1339 (2009). https://doi.org/10.1016/j.neunet.2009.05.008. Brain-Machine Interface 4. Datta, S., Boulgouris, N.V.: Recognition of grammatical class of imagined words from EEG signals using convolutional neural network. Neurocomputing 465, 301– 309 (2021). https://doi.org/10.1016/j.neucom.2021.08.035 5. Feng, Y., Chen, J., Xie, J., Zhang, T., Lv, H., Pan, T.: Meta-learning as a promising approach for few-shot cross-domain fault diagnosis: algorithms, applications, and prospects. Knowl. Based Syst. 235, 107646 (2022) 6. Huisman, M., van Rijn, J.N., Plaat, A.: A survey of deep meta-learning. Artif. Intell. Rev. 54, 4483–4541 (2021). https://link.springer.com/article/10. 1007/s10462-021-10004-4 7. Mini, P., Thomas, T., Gopikakumari, R.: EEG based direct speech BCI system using a fusion of smrt and MFCC/LPCC features with ANN classifier. Biomed. Signal Process. Control 68, 102625 (2021). https://doi.org/10.1016/j.bspc.2021. 102625 8. Nguyen, C.H., Karavas, G.K., Artemiadis, P.: Inferring imagined speech using EEG signals: a new approach using riemannian manifold features. J. Neural Eng. 15(1), 016002 (2017). https://doi.org/10.1088/1741-2552/aa8235 9. Nieto, N., Peterson, V., Rufiner, H.L., Kamienkowski, J.E., Spies, R.: Thinking out loud, an open-access EEG-based BCI dataset for inner speech recognition. Sci. Data 9(1), 1–17 (2022). https://doi.org/10.1038/s41597-022-01147-2 10. Ravichandiran, S.: Hands-On Meta Learning with Python: Meta learning using one-shot learning. Reptile, and Meta-SGD with TensorFlow. Packt Publishing, MAML (2018) 11. Rusnac, A.L., Grigore, O.: Generalized brain computer interface system for EEG imaginary speech recognition. In: 2020 24th International Conference on Circuits, Systems, Communications and Computers (CSCC), pp. 184–188 (2020). https:// doi.org/10.1109/CSCC49995.2020.00040 12. Selesnick, I., Baraniuk, R., Kingsbury, N.: The dual-tree complex wavelet transform. IEEE Signal Process. Mag. 22(6), 123–151 (2005). https://doi.org/10.1109/ MSP.2005.1550194 13. Xu, G., Ren, T., Chen, Y., Che, W.: A one-dimensional CNN-LSTM model for epileptic seizure recognition using EEG signal analysis. Front. Neurosci. 14 (2020). https://doi.org/10.3389/fnins.2020.578126 14. Zhao, S., Rudzicz, F.: Classifying phonological categories in imagined and articulated speech. In: 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 992–996 (2015). https://doi.org/10.1109/ICASSP. 2015.7178118
Design and Comparison of Artificial Intelligent Algorithms for Breast Cancer Classification Karen Valdez Hernández, Jhovana Cano Villalobos, Ana Castro Reyes, Andrea Gutiérrez Jurado, Sofia Moreno Terrones, Carlos Eduardo Cañedo Figueroa(B) , Abimael Guzmán Pando , and Gabriela Sámano Lira Facultad de Medicina y Ciencias Biomédicas, Universidad Autónoma de Chihuahua, 33000 Chihuahua, Chihuahua, México [email protected]
Abstract. Breast cancer early detection is a critical factor associated to patient survival and treatment cost reduction. Nevertheless, it is difficult to obtain a diagnose at earliest stages since it does not cause any symptoms. Recently, artificial intelligence field has demonstrated to be a suitable alternative to improve classification and early detection for this affection. Therefore, this paper proposes the design and comparison of three artificial intelligent algorithms for breast cancer classification. The algorithms used for classification were a naive Bayesian network (NBN), a support vector machine (SVM), and an artificial neural network (ANN). These algorithms were trained and validated in the Breast Cancer Prediction Database, located on the Kaggle platform. This database contains ten real-valued features computed from benign and malignant tumors. The evaluation results in F1 score shown 94%, 92% and 91% for the NBN, ANN and SVM respectively. These scores were compared with state-of-the-art algorithms to demonstrate the robustness of the proposed algorithms. The comparison was made considering feature-based and image-based models. Findings shown that our feature-based algorithms obtained competitive results requiring less computational resources than image-based models. Therefore, algorithms here proposed are a good option for the development of a high-fidelity system to classify the mentioned database into the cancer and non-cancer categories. Keywords: Breast cancer · Machine learning · Neural network
1 Introduction Breast cancer is one of the most common types of cancer, with more than 2.2 million cases in 2020 [1]. It begins mainly in the lining cells of the ducts and to a lesser extent in the lobes of the glandular tissue of the breasts. Progressively, the cancer may invade the surrounding breast tissue, and then spread to nearby lymph nodes or to other organs in the body. This last stage is known as metastasis and the probability of recovery is minimal [4]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 46–54, 2023. https://doi.org/10.1007/978-3-031-18256-3_5
Design and Comparison of Artificial Intelligent Algorithms
47
An accurate diagnosis is achieved through the integration of different clinical variables and mammographic features. An ideal diagnostic system must be able to discriminate between benign and malignant tumors since this ability is essential for risk estimation process [12]. The main difference between a benign and malignant is that benign tumor does not spread and does not cause any problem in other organs, while malignant or cancerous tumors easily expand to other organs and are life threatening [13]. Early-stage tumor detection can help to identify the best solution for possible treatment, if the malignant tumor is restricted to a specific area can be removed by surgery or radiation, but if the cancer has spread to an advance-stage it involved many body areas, so treatment includes chemotherapy, drug therapy, or immunotherapy. Therefore, cancer should be treated early in order to reduce mortality, but it is difficult to detect and diagnose the tumor at early-stage with common methods as manual detection because it is time-consuming and inefficient [13]. Therefore, the development of new automatic methods for early cancer detection are of vital importance. The goal is to reduce both the probability of death and the costs of treatment. Recently, the artificial intelligence algorithms have demonstrated the ability to automate the diagnostic process, since they allow to obtain reliable results in a short time compared to those obtained by a human [8]. Also, the development of artificial intelligent algorithms contributes to the reduction of errors in the detection of the degree of cancer affectation. Therefore, in this paper we propose the design and comparison of three different artificial intelligent techniques for breast cancer early detection. The algorithms are described on the next sections [10, 11].
2 Methodology 2.1 Database The database used in this work was the Breast Cancer Prediction Database from the research conducted by Dr. William H. Wolberd, at the University of Wisconsin Hospital in Madison. This dataset was generated from fluid samples taken from patients with solid breast masses and a graphical computer program called Xcyt, capable of performing analysis of cytological features from a digital scan [5]. The Breast Cancer Prediction Database contains 10 characteristics calculated for each tumor cell nucleus (radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, fractal dimension). We selected five of these characteristics for all algorithms design and training to obtain accurate results with a minimum error range. This selection was made considering only the characteristics without lack of information in the database. The selected characteristics were radius (mean of the distances from the center to the perimeter points), texture (standard deviation of the grayscale values), perimeter, area, and smoothness (local variation of the radius lengths). We used S = 424 samples for the training and validation process from the Breast Cancer Prediction Database, where 212 were taken from benign tumors (BT) and 212 malignant tumors (MT). For the training stage, we randomly selected 100 samples of benign (TSB) and 100 samples of malignant (TSM) tumors. For the validation stage, we used the rest samples 112 for both benign (VSB) and malignant tumors (VSM).
48
K. Valdez Hernández et al.
2.2 Artificial Neural Network The designed artificial neural network was based in [3] and it is shown in Fig. 1. It has an input, hidden and output layer. For the input layer, the five selected characteristics mentioned in Sect. 2.1 were fed it. The hidden layer consists of five neurons, with a sigmoidal tangent activation function. Finally, in the output layer there are two neurons, for the benign and malignant classification prediction. The number of neurons and hidden layers were determined experimentally until obtaining the best results.
Fig. 1. Neural network design with 5 inputs, 1 hidden layer of 5 neurons with sigmoidal tangential activation function and 1 output layer for two resulting classes.
The training schedule of the network was performed with a learning factor of 0.1, 3000 epochs and a minimum error of 1e–29. These values were obtained experimentally where 80%, 10% and 10% of the TSB and TSM data were used for training, validation and testing respectively in the complete training process for the ANN. We do not use the validation samples VSB and VSM for this stage. The output of the neural network was determined by Eq. (1), where ANN is the trained neural network, X is the input-sample vector of characteristics, and the outputs OB and OM indicate a cancer and non-cancer classification result. (OB , OM ) = ANN (X )
(1)
2.3 Naive Bayes Algorithm For the development of this model [2], we used TSB and TSM samples. At first step, we used Eq. (2) and Eq. (3) to obtain the probability that random data belong to a certain class: PB =
NTSB NTSB + NTSM
(2)
PM =
NTSM NTSM + NTSB
(3)
where NTSB and NTSM are the total number of TSB and TSM samples respectively, and PB and PM are the probability that a sample belongs to a cancer or non-cancer class, respectively. Subsequently, we obtained the statistical values of mean X jk and variance σjk 2 for each of the characteristics in both cancer and non-cancer classes using Eqs. (4) and (5) where Cijk refers to the value i R1x100 of the feature.
Design and Comparison of Artificial Intelligent Algorithms
49
j = {radius, texture, perimeter, area, smoothness} being evaluated in the class k = {benign, malignant} using TSM and TSB respectively, and Nj is the total amount of data for each feature, in this case, N j = 100. Cijk X jk = (4) Nj (Cijk − X jk )2 2 (5) σjk = Nj − 1 Then, Eq. (6) was applied to determine the probability P(Vij |k) of each sample Vij ∈ ((VSB) (VSM)) to belong to a class k.
P(Vij |k) =
−
1 2π σjk2
e
(Vij −X jk )2 2σ 2 jk
(6)
The priority Prk was obtained using Eq. (7) to calculate the product of all the probabilities for each class k. PB if k = Benign (7) Prk = Wk P(Vij |k);Wk = PM if k = Malignant Subsequently, the evidence Ev was calculated using Eq. (8), and finally the posteriori probability Ps of each class k was determined using Eq. (9). Ev = Prk (8) k
Psk =
Prk Ev
The final classification FC for each sample Vij was determined by: non − cancer if PsBenign > PsMalignant FC = cancer if Otherwise
(9)
(10)
2.4 Support Vector Machine The Support Vector Machine or SVM is a supervised learning algorithm that classifies samples in two classes (benign and malignant tumor) considering an optimal line. For this work, we used the features l = {area, perimeter}, since they achieved a better class discrimination. Therefore, we formed two feature sets, one for each class of the training set VBi,l ∈ TSB and VM i,l ∈ TSM . Once the training sets were created, we obtained the Euclidean distances Di by using Eq. (11). (11) Di = (VMi,area − VBi,area )2 + (VMi,perimeter − VBi,perimeter )2
50
K. Valdez Hernández et al.
Subsequently, the two lowest distance values were used as [6, 7] to define and generate the coefficients of the optimal equation line, Eq. 12: Y = a · Vi,area + b · Vi,perimeter + c
(12)
where a = 0.0531, b = 0.0053, c = -8.7773 and Vi,l ∈ ((VSB)U(VSM )) is the sample to be analyzed. If the result of Eq. (12) shows that Y is positive, the sample will be classified as malignant tumor and as benign tumor otherwise.
3 Results and Discussions The algorithms were evaluated whit VSB and VSM datasets. These evaluations are presented next using confusion matrices for each algorithm developed. The Table 1 show the results obtained whit the artificial neural network. With this data we calculated a precision of 96%, an accuracy of 92% a recall of 88% and a 92% of F1 score (View Fig. 2).
Real
Table 1. Confusion matrix of the neural network Benign Malignant
107 14 Benign Predicted
5 98 Malignant
Fig. 2. Results of artificial neural network model.
The results obtained from the Bayesian algorithm can be seen in the Table 2. This was the best algorithm and its metrics results were precision of 91%, accuracy of 94%, recall of 96% and F1 of 94% (View Fig. 3).
Design and Comparison of Artificial Intelligent Algorithms
51
Real
Table 2. Confusion matrix of the Bayesian algorithm Benign Malignant
102 4 Benign Predicted
10 108 Malignant
Fig. 3. Results of Bayesian algorithm.
For the support vector machine, the results can be seen in the Table 3. From these results it was obtained a precision of 88%, an accuracy of 91%, a recall of 94% and an F1 of 91%, only 1% less from the neural network. This information is showed on Fig. 4.
Real
Table 3. Confusion matrix of the support vector machine Benign Malignant
98 6 Benign Predicted
14 106 Malignant
Fig. 4. Results of SVM model.
52
K. Valdez Hernández et al.
Table 4 shown the comparison of the proposed algorithms results respect some state-of-the-art algorithms to evaluate their performance. It indicates the algorithms performance on accuracy, precision, recall and F1 score metrics and it is included the type of input data employed for each algorithm. It can be noted that in comparison with ensemble learning models, the algorithms proposed preserve metric consistency having high similar values, above 91%, in the F1 score. The highest F1 score was obtained by the SVM algorithm [14]. However, our proposed SVM and Bayesian algorithm achieved comparable or even better results than [14] on Recall metric. The next highest F1 score is 95.29% obtained by an ensemble deep Learning algorithm [20] where digital image is processed. However, due to complete image processing this type of algorithm requires very specialized computer equipment such as GPUs and huge memory to be trained or executed. On the other hand, our proposed algorithms are less computational expensive preserving good metric scores. Therefore, it can be inferred that it is not necessary to base breast cancer classification on image-based models to obtain results above the average. Table 4. Comparison of the state-of-the-art algorithms results with the proposed algorithms. Autor
Input representation
Accuracy
Precision
Recall
F1 Score
Ours SVM
Features
91%
88%
94%
91%
Ours Bayesian
Features
94%
91%
96%
94%
Ours ANN
Features
92%
96%
88%
92%
SVM Z.Suhail, et. al. [14] Features
95%
98%
94%
95.9%
SVM K. Wadkar, et al. [15]
Features
95%
-
-
-
CNN M. Nawaz, et al. [16]
Image
95.4%
-
-
-
CNN Y. J. Tan, et al. [17]
Image
82.7%
70.5%
82.7%
76.13%
KNN J.C. Peña [18]
Features
95%
90%
-
89%
KNN Lévano and Cérdan [19]
Features
85%
85%
85%
85%
Ensemble deep learning [20]
Image
95.29%
93.48
97.73%
95.29%
Ensemble learning [21]
Features
90%
97%
81%
89%
4 Conclusion The efficiency results among the proposed algorithms show that the Bayesian algorithm obtained the best results with an F1 score of 94%. Also, ANN and SVM models preserve metric consistency having high F1 scores of 91% and 92% respectively. However, the possibility of implementing ensemble learning algorithms or performing data augmentation to increase the classification efficiency is foreseen. Regarding state-of-the-art
Design and Comparison of Artificial Intelligent Algorithms
53
comparison, we compared feature-based and image-based models. Findings showed that feature-based models proposed Bayesian, ANN and SVM were a good option to obtain satisfactory results on breast cancer classification with reduced memory and computational resources in comparison to image-based models. Considering that this type of studies can generate a very positive impact in the health area, further research will be continued to achieve a more efficient algorithm in this case study. As future work we would like to generate or explore other databases to consider information on culture and ethnicities from different regions. Adding these datasets in the training process would help to develop a more generic algorithm for early detection of breast cancer. Also,
References 1. Arceo-Martínez, M.T., López-Meza, J.E., Ochoa-Zarzosa, A., Palomera-Sanchez, Z.: Estado actual del cáncer de mama en México: principales tipos y factores de riesgo. Gaceta mexicana de oncología 20(3), 101–110 (2021). http://www.scielo.org.mx/scielo.php?pid=S2565-005 X2021000300101&script=sci_arttext 2. Castrillón, O.D., Castaño, E., Castillo, L.F.: Sistema predictivo Bayesiano para detección del cáncer de mama. Información tecnológica 29(3), 257–270 (2018). https://scielo.conicyt.cl/ scielo.php?pid=S0718-07642018000300257&script=sci_arttext&tlng=en 3. Flores Gutiérrez, H.: Redes Neuronales Aplicadas a la detección de Cáncer de Mama (Tésis para obtener el grado de Maestría en Tecnología de Cómputo). Instituto Politécnico Nacional (2015) 4. Jiménez, M., et al.: Recomendaciones para mejorar la práctica de la medicina en la atención del paciente con cáncer. CONAMED, pp. 5–12 (2002) 5. Kalshtein, Y.: Wisconsin Breast Cancer (Diagnostic) DataSet Analysis. Recuperado de (2017). https://rstudio-pubs-static.s3.amazonaws.com/344010_1f4d6691092d4544bfbddb09 2e7223d2.html 6. León, E.: Introducción a las máquinas de vector soporte (SVM) en aprendizaje supervisado. Trabajo de Fin de Grado en Matemáticas, Universidad de Zaragoza (2017). https://zaguan. unizar.es/record/59156/files/TAZ-TFG-2016-2057.pdf?version=1 7. Ramirez-Quintana, J.A., Madrid-Herrera, L., Chacon-Murguia, M.I., Corral-Martinez, L.F.: Brain-computer interface system based on P300 processing with convolutional neural network, novel speller, and low number of electrodes. Cogn. Comput. 13(1), 108–124 (2020). https://doi.org/10.1007/s12559-020-09744-2 8. Pazmiño Martínez, L.A., Ruiz Baquerizo, K.R.: Contribución de la implementación de Inteligencia Artificial para el diagnóstico precoz del cáncer de mama. Universidad Central del Ecuador (2021). http://www.dspace.uce.edu.ec/handle/25000/24990 9. Singh, M.: Breast Cancer Prediction Dataset. Kaggle (2018). https://www.kaggle.com/dat asets/merishnasuwal/breast-cancer-prediction-dataset 10. World Health Organization: Cáncer de mama (2021). https://www.who.int/es/news-room/ fact-sheets/detail/breast-cancer 11. Harbeck, N., Penault-Llorca, F., Cortes, J., et al.: Breast cancer. Nat. Rev. Dis. Primers 5, 66 (2019). https://doi.org/10.1038/s41572-019-0111-2 12. Mojtaba, S., Maryam, T., Abbas, R., Salar, R.: Assessing breast cancer risk with an artificial neural network. Asian Pac. J. Cancer Prev. 19, 1017–1019 (2018) 13. Pinagel, R., Lokhande, S., Pasalkar, S., Solanke, A.: "Breast cancer detection using ANN classifier. Int. J. Innov. Eng. Res. Technol. 4, 58–61 (2017)
54
K. Valdez Hernández et al.
14. Suhail, Z., Denton, E.R.E., Zwiggelaar, R.: Classification of micro-calcification in mammograms using scalable linear Fisher discriminant analysis. Med. Biol. Eng. Compu. 56(8), 1475–1485 (2018). https://doi.org/10.1007/s11517-017-1774-z 15. Wadkar, K., Pathak, P., Wagh, N.: Breast cancer detection using ANN network and performance analysis with SVM. Int. J. Comput. Eng. Technol. 10(3), 75–86 (2019) 16. Nawaz, M., Sewissy, A.A., Soliman, T.H.A.: Multi-class breast cancer classification using deep learning convolutional neural network (IJACSA). Int. J. Adv. Comput. Sci. Appl. 9(6), 316–322 (2018) 17. Tan, Y.J., Sim, K.S., Ting, F.F.: Breast cancer detection using convolutional neural networks for mammogram imaging system. In: Conference: 2017 International Conference on Robotics, Automation and Sciences (ICORAS), Malaysia (2017) 18. Peña, J.C.: Despliegue de un modelo de clasificación de tumores de cáncer de mama. Universidad de Antioquia, Medellín, Trabajo de grado (2022) 19. Lévano-Rodriguez, D., Cerdán-León, F.E.: Discriminación de masas mamográficas mediante K-Nearest Neighbor y atributos BIRADS. Rev. Cient. Sist. Inform. 2(1), e225 (2022) 20. Hameed, Z., Zahia, S., Garcia-Zapirain, B., Javier Aguirre, J., María Vanegas, A.: Breast cancer histopathology image classification using an ensemble of deep learning models. Sensors 20(16), 4373 (2020) 21. Mohebian, M.R., Marateb, H.R., Mansourian, M., Mañanas, M.A., Mokarian, F.: A hybrid computer-aided-diagnosis system for prediction of breast cancer recurrence (HPBCR) using optimized ensemble learning. Comput. Struct. Biotechnol. J. 15, 75-85 (2017)
Electrophysiological Signals Simulation with Machine Learning Mario Axel L´ opez Agui˜ naga1(B) , Arturo Valdivia Gonz´ alez2 , 1 and Laura Paulina Osuna Carrasco 1
Translational Bioengineering Department, University Center of Exact Sciences and Engineering, University of Guadalajara, 44430 Guadalajara, Jalisco, Mexico [email protected] 2 Computational Sciences Department, University Center of Exact Sciences and Engineering, University of Guadalajara, 44430 Guadalajara, Jalisco, Mexico
Abstract. The simulation of electrophysiological signals has been a challenging task for the computer sciences for a long time due to their complex morphology and the high environmental noise present within the human body. In an attempt to replicate these signals a variety of conventional models have been developed but all of those presented complexity restrictions that prevented them to reach an acceptable similarity with the real signals. Viewing this situation the goal of this work was set to demonstrate that it is possible to simulate electrophysiological registers using machine learning algorithms, to achieve this goal electromyographic registers (EMG) from the Biceps Brachii muscle of three subjects were recorded using surface electrodes and each register was split in time fragments with a length of 1 s to be used as a data set for the model. Subsequently, the proposed machine learning model based in a recurrent neural network (RNN) and an autoencoder was trained to simulate EMG signals based on the previously obtained data set. Finally, the resultant simulations of this model were analyzed using digital signal processing techniques and the results showed that the simulations behaved in a very similar way to the real electromyographic registers. Keywords: Machine learning · Computer science · Electrophysiology Time series forecasting · Signal simulation · Simulation · Electromyography · EMG · Surface EMG · Matlab · Python · Tensorflow · Keras · Recurrent neural network · RNN · Long short term memory · LSTM · Autoencoder
1
·
Introduction
In recent years the field of the machine learning has been experimenting an exponential growth around its practical applications due to the fact that this kind of algorithms allow the computers learn for themselves how to execute complex tasks that by any other means such as conventional programming would be very c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 55–72, 2023. https://doi.org/10.1007/978-3-031-18256-3_6
56
M. A. L´ opez Agui˜ naga et al.
difficult or impossible to solve [1]. Among these tasks we find the automatic classification and pattern recognition in variate media files, single values forecasting and time series forecasting [2,3]. This last task is important for the biomedical engineering because it could allow the generation of synthetic signals that can be used in a variety of applications such as references for a variety of signal processing techniques and to evaluate the performance of medical acquisition devices without the need of a living subject. A variety of models have been developed in an attempt to achieve the electrophysiological signals forecasting such as the mathematical models, the electronical models and the differential equation models, like the Izhikevich model [4] that is capable of use differential equations to forecast up to 8 different types of neuron firing patterns; however, the complexity restrictions that imply the definition of differential equations make this model only capable to synthesize simple signals. Following this restriction David Belo et al. [5] developed a machine learning model based on recurrent neural networks to forecast three types of signals: electromyographic, electrocardiographic and respiratory signals. The electrocardiographic and respiratory datasets were obtained from the Physiobank Database website while the electromyographic dataset was created by the authors themselves. Their model was trained in 410 epochs and the subsequent evaluation showed that the electrocardiographic and respiratory forecasted signals presented an error rate smaller than 15%, in the other hand, the electromyographic forecasted signals presented an error rate higher than 50%. Another machine learning model developed to forecast electrophysiological signals is the one created by Rafael Zanini et al. [6] to generate synthetic Parkinson’s deceased affected electromyographic signals. To accomplish the task of forecast pathological signals 4 different neural architectures were tested, being the most prominent in performance terms the autoencoded LSTM because its computational costs were significantly smaller than the rest of architectures and having a loss function score of 0.14 at the moment of forecasting the interest signals. Viewing the results of the previously mentioned works motivated us to developed a similar machine learning model based on the recurrent neural network architecture with the purpose of forecasting electromyographic activity from the Biceps Brachii muscle taking as reference a dataset of real registers of this same muscle and test these generated signals utility.
2 2.1
Materials and Methods Data Acquisition and Processing
The first step to acquire electromyographic registers for the machine learning model was the selection of the study subjects. We selected three subjects within an age range from 18 to 30 years old with no background neither of neuropathies nor myopathies and no recent muscular injuries. Prior to the recording sessions all the study subjects were informed about the non-invasive nature of the procedure and its scientific intentions, following this explanation the three study subjects gave their explicit consent to perform the data recording procedure. In
Electrophysiological Signals Simulation
57
order to capture the interest signals three surface electrodes were placed in the following positions on the subjects’ right arm: The first electrode was placed on the ventral portion of the Biceps Brachii muscle. The second electrode was placed on the distal portion of the Biceps Brachii muscle. Finally the third electrode was placed on the olecranon bone (elbow). Figure 1 shows the surface electrode configuration used to acquire the electromyographic signals.
Fig. 1. Surface electrodes placement on a study subject’s right arm. In the left image the positive (upper) and negative (lower) electrodes are displayed. In the right image the reference electrode is displayed.
To process the electrical readings a data acquisition circuit board was designed and built. The circuit board included all of the following features: 1) Three mechanical connections for the wires connected to the surface electrodes. 2) Two instrumentation amplifiers INA128P. 3) One analog filtering module consisting in a 4th order Butterworth pass-band filter with a frequency range from 0.5 500 Hz. 4) One selector to switch between the filtering module output and the Notch filter output. 5) One Notch filter 60 Hz. 6) One controllable gain circuit for the second INA128P. 7) One controllable offset circuit from 0 to 5 V. 8) One voltage output for the analog to digital converter. The proposed circuit was energized using symmetrical voltage from two 9V batteries connected together, due to this fact the selector switch was toggled to get the electromyographic signal directly from the filtering module output and pass it through the rest of the circuit. The selected analog to digital converter was an Arduino Uno board due to its acceptable resolution of 10 bits and ease to use. Figure 2 shows a flowchart that describes all the stages that compose the proposed data acquisition circuit. Figure 3 shows both the printed circuit boards employed to fabricate the acquisition and filtering boards.
Fig. 2. Simplified feature flowchart of the data acquisition circuit.
58 M. A. L´ opez Agui˜ naga et al.
Electrophysiological Signals Simulation
59
Fig. 3. Circuit boards used to acquire the electromyographical data necessary for the machine learning model. The left image displays the main acquisition board and its measures: 100 mm × 10 mm. The right image displays the filtering board and its measures: 50 mm × 50 mm.
The processed data was captured in real time using a Matlab script where the recording time was set at 10 s for each study subject and the sampling rate was set at 1 kHz. To obtain the interest signals each selected subject was told to contract the Biceps Brachii muscle for periods of one second without any kind of extra weight, with these instructions given 10 contractions of the Biceps Brachii muscle were recorded per session and a total of 100 s were obtained per subject resulting in 300 s of activity. The reason why only 3 study subjects were studied in this work was that the complex nature of electromyographic signals made that each muscular contraction signal was different from the previous ones. To convert the obtained signals into a useful dataset for the machine learning model each signal was split into time fragments with a duration of 1 s using another Matlab script and the DC offset was removed from them using the following formula: (1) XDataset = XF ragment − X F ragment where X F ragment is the mean of each obtained time fragment. Once the offset removal was completed all of the time fragments were concatenated into a rectangular matrix which will work as the dataset for the machine learning model and said matrix was consequently saved as a .csv file. 2.2
Data Reshape for Machine Learning Model
The .csv file previously saved was loaded into a Python script to reshape it for the proposed model understanding. Before starting with the data reshape all the values contained in the original dataset matrix were scaled within a range of values from 0 to 1 and split in two smaller matrices with the same size, one of them working as the training set for the machine learning model and the other one working as the testing set. The required dataset shape for a recurrent neural network model consists in a 3-dimensional matrix where the
60
M. A. L´ opez Agui˜ naga et al.
first dimension is the number of samples recorded in each sequence, the second dimension is the number of data units covered in each time step of the sequence and the third dimension is the number of different features observed in each sample. The required data shape can be represented as the follows [7]: [NSamples , TStep , NF eatures ]. Both created datasets were reshaped into this new shape to be used by the machine learning model in their respective stages. 2.3
Machine Learning Model Design, Training and Testing
The proposed machine learning model was developed using Tensorflow and Keras employing the following architecture: Autoencoder Input → First LSTM layer → Repeat Vector function → Second LSTM layer → Autoencoder Output → Time distribuited neuron layer. The reason why the LSTM neuron layers were chosen for this proposed machine learning model is that these neurons are specialized in processing time sequences and therefore have the capability to remember information, update it or forget it according with the structure of input data. With the machine learning model designed the training phase was performed using the training dataset through 100 epochs employing as optimization function the ADAM function, the mean squared error was set as the loss function and the accuracy parameter was set as the evaluation metric. The testing phase of the model was performed using the testing dataset to predict 1 s of muscular activity. 2.4
Model Utility
To test the model utility for the biomedical engineering three applications were developed and tested. The first one is a synthetic signal generator, which would be able to generate complete electromyographic registers with a duration set by the user, to fabricate this signal generator the model was integrated within a loop cycle where the number of iterations was the number of seconds introduced by the user, after each iteration is completed the forecasted signals are reshaped into a one dimensional vector and at the end of them a numeric value equal to 0 is appended to have 1,000 samples per second. Consequently, the 1 forecasted registers are concatenated sequentially to create a synthetic electromyographic signal, due to the fact that the machine learning model can only output values in a range from 0 to 1 the synthetic registers are scaled to a maximum value of 1024 to replicate the amplitude of the real signals acquired with a 10-bit analog to digital converter. Once the complete signal is created a digital 4th order high-pass Butterworth filter with a cutoff frequency of 0.01 Hz eliminates the low frequency noise coming from the model’s signal forecasting procedure. The Listing 1.1 shows the pseudocode for the electromyographic signal simulator using the proposed machine learning model.
Electrophysiological Signals Simulation
61
n=u s e r t i m e i n p u t f o r i in range n : S i n g l e m o d e l p r e d i c t i o n=m o d e l p r e d i c t S i g=S i n g l e m o d e l p r e d i c t i o n S i g 1 s e c=append ( Sig , 0 ) S i m S i g=c o n c a t e n a t e ( S i g 1 s e c ) f i l t e r=b u t t e r w o r t h ( 4 t h o r d e r , 0 . 0 1 Hz , h i g h p a s s ) S i m f i l=a p p l y f i l t e r ( f i l t e r , S i m S i g ) Listing 1.1. Pseudocode for signal simulator
The second application is centered in the generated signals utility for the signal processing techniques and the properties that they exhibit when they are processed. The following signal processing techniques were employed in the generated signals: the discrete fourier transform [8], the normalized correlation coefficient [9], the cross correlation [10,11], the coherence [12], the computed wavelet transform using the Morlet waveform [13] and the spectrogram [14]. In the case of the normalized correlation coefficient the original equation was modified to work with 1-dimensional signals instead of 2-dimensional images. Passing from this equation [9]: ¯ ¯ (i,j)∈R (I(r + i, s + j) ∗ R(i, j)) − N ∗ I(r, s) ∗ R (2) CL (r, s) = 2 2 ¯ (i,j)∈R (I(r + i, s + j)) − N ∗ (I(r, s)) ∗ σR To this one:
(j)∈R (I(s
CLM od (s) =
¯ ∗R ¯ + j) ∗ R(j)) − N ∗ I(s)
2 2 ¯ (j)∈R (I(s + j)) − N ∗ (I(s)) ∗ σR
(3)
where I is the interest signal and R is the signal fragment used as reference, 1 1 ¯ ¯ ¯ 2 I = N (j)∈R I(s + j), R = N (j)∈R R(j) and σR = (j)∈R (R(j) − R) The third application is a device tester that uses the signals generated by the model to evaluate the performance of acquisition devices by broadcasting a signal to the interest device and compare the output signal with the transmittted one. A circuit based on the digital to analog converter MCP4725 was designed, the circuit included a bipolar voltage converter provided by the manufacturer in its datasheet [15] and a modified Wheatstone bridge where the resistances are fixed and the voltage in one of the sides is variable, with this circuit we pretend to replicate the nature of the muscular signals and its amplitude range when measured with surface electrodes which is from 1 to 10 mV [16]. Figure 4 shows the printed circuit board employed to design the signal transmitter.
62
M. A. L´ opez Agui˜ naga et al.
Fig. 4. Circuit board for the signal transmitter circuit. Measures: 50 mm × 50 mm.
The designed circuit is connected to the computer via an Ardino Uno board that serves as a data transmitter using serial and I2C communication to broadcast the selected signal to the interest instrument, once the transmitted signal is sampled by the acquisition device and captured by another Arduino Uno serving as a data receiver the original signal and the obtained one are scaled in a range within 0 and 1 and the euclidean distance between them is measured using the following formula to calculate the existing error rate between them: (4) dE (i) = (ST ransmitted (i) − SReceived (i))2
3
Results
The designed circuit and the interface to record the muscular activity from the study subjects had an acceptable real time performance in spite of the large amount of recorded data, with this program we managed to capture good-quality EMG signals to create the dataset for the model. Figure 5 shows a real time captured signal from one of the selected study subjects. The creation of the machine learning model dataset was successful, all signals obtained from the study subjects were split in 300 time fragments with a duration of 1 s (1000 points) each. The calculated mean for all the fragments were approximately 0 which means that the DC component of the signals were successfully removed. Figure 6 shows a waterfall plot where all the obtained time fragments are displayed.
Electrophysiological Signals Simulation
63
Fig. 5. Electromyographic register obtained in real time from a study subject.
Fig. 6. Created dataset for the machine learning model.
From the original dataset two new datasets were created, one for training and one for testing, with the following dimensions: [999, 1, 150]
(5)
where 999 are the number of samples per sequence, 1 is the number of data units covered in each time step and 150 is the number of features observed at each sample. Table 1 shows the used hyperparameters for the machine learning model. Figure 7 shows the diagram with the proposed machine learning model architecture and the number of neurons present in each layer.
64
M. A. L´ opez Agui˜ naga et al. Table 1. Model hyperparameter table. Hyperparameters
Values
Training set sample size
150 (50%)
Testing set sample size
150 (50%)
Batch size
50
Learning epochs
100
LSTM layers neurons
150
Autoencoder layers neurons
15
Autoencoder compression rate
1/10
Time distribuited layer neurons 1
Fig. 7. Model architecture diagram.
In model training phase it was detected that the accuracy metric results were too low for the conventional recurrent neural networks applications, this effect was caused due to the non-periodical nature of the electromyographic signals and therefore we can say that, the lower the resulting accuracy the better is the model’s performance also, the loss function remained in low values during almost all the learning epochs meaning that the most of the information given
Electrophysiological Signals Simulation
65
to the model was retained. Listing 1.2 contains the obtained metrics while the machine learning model was trained using its respective dataset. Epoch 1/100 20/20 − 4 s 4ms/ s t e p − l o s s : 0 . 1 1 4 1 − a c c u r a c y : 0 . 0 0 5 2 Epoch 10/100 20/20 − 0 s 5ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 20/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 30/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 40/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 50/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 60/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 70/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 80/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 90/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Epoch 100/100 20/20 − 0 s 4ms/ s t e p − l o s s : 0 . 0 1 5 7 − a c c u r a c y : 0 . 0 0 5 2 Listing 1.2. Recorded metrics during model training phase.
To see the model training results this last one was ordered to forecast only one second (1000 data samples) of muscular activity. Figure 8 shows the forecasted signal with the original model amplitude range (from 0 to 1) and no DC offset.
66
M. A. L´ opez Agui˜ naga et al.
Fig. 8. Forecasted EMG signal. The model generated signal presents the same relaxation and contraction regions that the real EMG signals.
To show the signal simulator performance a time of 5 s was introduced in the simulator interface, with this parameter defined the software was executed and the forecasted signal was created. Figure 9 shows the 5 s electromyographic signal created by the model.
Fig. 9. Electromyographic signal generated using the signal simulator interface and a time of 5 s introduced by the user. The unions between the forecasted time fragments are almost imperceptible, giving the illusion that the displayed signal is a real EMG register.
The performance of the synthetic signals under the application of the previously described processing techniques were tested using one forecasted signal with a duration of 1 s and two forecasted registers with a duration of 10 s, in this way we can use the one second signal as a template to compare all the values that compose the interest signal in a local level and also these long registers can be used to visualize the signal properties easier. Figure 10 displays the results of a discrete Fourier transform applied on a 10 s forecasted signal. Figure 11 displays the results of a normalized correlation coefficient analysis over a 10 s signal using a 1 s reference signal. Figure 12 shows the results of a cross-correlation analysis
Electrophysiological Signals Simulation
67
over a 10 s signal using a 1 s signal as reference. Figure 13 shows the results of a cross-correlation analysis over a 10 s signal using another 10 s signal as reference. Figure 14 shows he results of a coherence analysis between the two synthetic signals with a duration of 10 s. Figure 15 shows a computed wavelet transform applied to a 10 s signal using the Morlet wavelet. Figure 16 shows the results of an spectrogram analysis on a 10 s synthetic signal.
Fig. 10. Discrete Fourier transform applied on a 10 s EMG synthetic signal. In spite of that the spectral power of the signal is dispersed across the whole frequency range the highest peaks are located among the 150 Hz and the 300 Hz.
Fig. 11. Normalized correlation coefficient analysis applied on a 10 s synthetic EMG signal using a 1 s EMG synthetic signal reference. The obtained correlation values are low due to the applied formula which normalizes each data comparison made by the analysis creating a local level comparison.
68
M. A. L´ opez Agui˜ naga et al.
Fig. 12. Cross-correlation analysis applied on a 10 s synthetic EMG signal using a 1 s synthetic EMG signal. The higher correlation values are result of the cross-correlation formula which normalizes all the data at once instead of normalizing each data comparison giving as result a global level comparison.
Fig. 13. Cross-correlation analysis applied on a 10 s synthetic EMG signal using a 10 s synthetic EMG signal. The obtained correlation values are the highest in comparison with the previous correlation analyses.
Electrophysiological Signals Simulation
69
Fig. 14. Coherence analysis applied on two synthetic EMG signals with the same length of 10 s. The highest coherence peak value can be found in between 200 Hz and 300 Hz.
Fig. 15. Computed Wavelet transform applied on a synthetic EMG signal with a length of 10 s using the Morlet waveform. The highest activity spots detected by the Morlet wavelet can be found from 100 Hz to 300 Hz.
Fig. 16. Spectrogram applied on a synthetic EMG signal with a length of 10 s. The highest activity spots of the signal can be found between the 200 Hz and the 500 Hz.
70
M. A. L´ opez Agui˜ naga et al.
To show the acquisition instrument evaluator performance a synthetic EMG signal with a duration of one second was loaded by the user and transmitted using the proposed digital to analog converter (DAC) circuit. Figure 17 shows the results of transmitting the loaded signal with the DAC circuit to an acquisition device and he consequent error rate calculation.
Fig. 17. Acquisition instruments evaluator test with a 1 s EMG synthetic signal. The received signal is only electrical noise and there is no similarity with the transmitted signal.
4
Discussions
Comparing the proposed machine learning model with the ones created by the previously mentioned authors David Belo and Rafael Zanini serves to realize that this is a smaller model in terms of the utilized hyperparameters because we only worked with one type of electrophysiological signals, and with only one signal channel being that the Belo’s model worked with three different types of physiological signals, and the Zanini’s model worked with electromyographic signals affected by the Parkinson’s decease; however none of these works attempted to explain the generated signals utilities in the biomedical engineering field beyond the capability of generate synthetic electrophysiological signals with determined characteristics. In the other hand, this work attempted to explain the utilities for the biomedical engineering field of the proposed model and the properties exhibited by its generated signals using digital to analog conversion circuits and signal processing techniques.
Electrophysiological Signals Simulation
5
71
Conclusions
The proposed machine learning model successfully forecasted complete electromyographic registers based on the real signals given during its training and testing. The simulated signals exhibited a similar but not totally identical behavior and properties to the real electromyographic registers based on the results of the employed signal processing techniques. The part of instrument evaluation using the simulated signals must be reframed to find new methods to simulate a bipolar signal and attenuate it to the interest signals’ amplitude range, despite of these inconveniences the use of machine learning models to simulate complex signals generates a new field of opportunities for the computer sciences and the biomedical engineering as well because it would allow the replication of the electrical activity from individual organs to complete physiological systems without the need of an alive study subject and apply them as a template to detect potential pathologies in real registers and to evaluate the performance of medical devices specialized in the recording of this kind of signals as well.
References ´ 1. Valdemar, E.-C., Avalos, O., D´ıaz, P.-E., Valdivia, A.-G., Cisneros, M.-A.: Introducci´ on al Machine Learning con MATLAB, 1st edn. Marcombo S.L., Barcelona (2021). ISBN: 978-84-267-3282-8 2. Kubat, M.: An Introduction to Machine Learning, 2nd edn. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-63913-0. ISBN: 978-3-319-63913-0 3. Aggarwal, C.C.: Recurrent Neural Networks. In: Aggarwal, C.C. (ed.) Neural Networks and Deep Learning. Springer, Cham (2018). https://doi.org/10.1007/978-3319-94463-0 7. ISBN: 978-3-319-94463-0 4. Izhikevich, E.-M.: Simple model of spiking neurons. IEEE Trans. Neural Netw. 14(6), 1569–1572 (2003). https://doi.org/10.1109/TNN.2003.820440 5. Belo, D., Rodrigues, J., Vaz, J.-R., Pezarat-Correia, P., Gamboa, H.: Biosignals learning and synthesis using deep neural networks. Biomed. Eng. Online 16(1), 115 (2017). https://doi.org/10.1186/s12938-017-0405-0 6. Zanini, R.-A., Colombini, E.-L., Ferrari de Castro, M.-C.: Parkinson’s decease EMG signal prediction using neural networks. In: 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC), pp. 2446–2453 (2019). https:// doi.org/10.1109/smc.2019.8914553 7. Brownlee, J.: How to reshape input data for long short-term memory networks in Keras. Machine Learning Mastery (2019). www.machinelearningmastery.com/ reshape-input-data-long-short-term-memory-networks-keras/ 8. Numpy.org: Discrete Fourier Transform (numpy.fft). NumPy v1.22 Manual (2022). www.numpy.org/doc/stable/reference/routines.fft.html ´ 9. Valdemar, E.-C., Avalos, O., D´ıaz, P.-E., Valdivia, A.-G., Cisneros, M.-A.: Coeficiente de correlaci´ on. In: Introducci´ on al Machine Learning con MATLAB, 1st edn., pp. 186–189. Marcombo S.L., Barcelona (2021). ISBN: 978-84-267-3282-8 ´ 10. Valdemar, E.-C., Avalos, O., D´ıaz, P.-E., Valdivia, A.-G., Cisneros, M.-A.: Correlaci´ on cruzada normalizada. In: Introducci´ on al Machine Learning con MATLAB, 1st edn., pp. 184–186. Marcombo S.L., Barcelona (2021). ISBN: 978-84-267-3282-8
72
M. A. L´ opez Agui˜ naga et al.
11. Scipy.org: scipy.signal.correlate. SciPy v1.8.1 Manual (2022). www.docs.scipy.org/ doc/scipy/reference/generated/scipy.signal.correlate.html 12. Scipy.org: scipy.signal.coherence. SciPy v1.8.1 Manual (2022). www.docs.scipy.org/ doc/scipy/reference/generated/scipy.signal.coherence.html 13. PyWavelets: Continuous Wavelet Transform (CWT). PyWavelets Documentation (2022). www.pywavelets.readthedocs.io/en/latest/ref/cwt.html 14. Scipy.org: scipy.signal.stft. SciPy v1.8.1 Manual (2022). www.docs.scipy.org/doc/ scipy/reference/generated/scipy.signal.stft.html 15. Microchip: 12-Bit Digital-to-Analog Converter with EEPROM Memory in SOT23-6. MCP4725, June 2009 [Rev. D]. www.digikey.cz/htmldatasheets/production/ 48676/0/0/1/mcp4725.html 16. Raez, M., Hussain, M., Mohd-Yasin, F.: Techniques of EMG signal analysis: detection, processing, classification and applications. Biol. Proced. Online 8, 11–35 (2006). https://doi.org/10.1251/bpo115
Quantification of a Lip and Palate Clefts Classification Beatriz Gutiérrez-Sánchez1 , José Maya-Behar2 , and Martha Ortiz-Posadas1,2(B) 1 Universidad Autónoma Metropolitana Iztapalapa, 09340 CDMX, Mexico
[email protected] 2 SUMA Cleft Leadership Center, Smile Train, 03103 CDMX, Mexico
Abstract. The team of plastic surgeons that collaborates in the Smile Train Cleft Leadership Center (CLC) SUMA in Mexico, is interested in developing a quantitative classification for representing the surgical complexity of clefts and provides more objective criteria for the treatment of patients. The objective of this work was to propose a cleft classification and a relevance factor for each cleft. The classification is a problem of multiple attribute decision making, and that is why we use the Analytic Hierarchy Method for developing the relevance factor for each cleft. The relevance factor was validated in a sample of fifty patients treated at the SUMA-CLC. A total of twenty-nine classes were defined: nine for primary palate, two classes for secondary palate, and eighteen classes for labial-palatal clefts. The relevance factors obtained for the clefts grouped in these classes are in the interval (3, 298), where three means the lowest relevance and 298 means the highest. In general, the values increase as the cleft characteristics become more complex. Surgeons consider that the class relevance factor effectively represents the complexity of the cleft, and the surgical treatment requires by the patient. Keywords: Cleft lip and palate · Classification · Relevance factor · Surgical complexity
1 Introduction Congenital malformations of lip cleft and palate cleft, clinically termed primary and secondary palate cleft respectively, have an incidence across the world accepted to be 1 in 1,000 births [1]. In Mexico, the incidence is 1 per 700 live births [2]. Clefts can have distinctive characteristics. For example, clefts can affect the upper lip, nose, and teeth, in which case it is called cleft of primary palate. It can only affect the palate; in which case it is denominated as cleft of secondary palate. It can affect both palates simultaneously; in which case it is denominated labial-palatal cleft [3]. The Smile Train Cleft Leadership Center (CLC) SUMA in Mexico [4] provides integral treatment to its patients through thirteen health specialties. The general objective of the plastic surgery specialty is to correct any congenital, acquired, tumoral or involutive process that affects shape and/or body function. In the case of lip and palate malformations, plastic surgeons perform various surgical procedures: cheiloplasty (lip reconstruction), palatoplasty (palate reconstruction), rhinoplasty (aesthetic and functional surgery of the nose), alveolar bone graft (gum reconstruction), among others. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 73–82, 2023. https://doi.org/10.1007/978-3-031-18256-3_7
74
B. Gutiérrez-Sánchez et al.
The group of plastic surgeons that collaborates in SUMA-CLC is interested in developing a quantitative classification for representing the surgical complexity of clefts and provides more objective criteria for the patient’s treatment. The objective of this work was to propose a classification and a relevance factor for cleft lip and palate considering its characteristics, to unify criteria for the surgical treatment of the patients. The relevance factor (ρ) represents the surgical complexity of the clefts and therefore the patient treatment. A relevance factor was associated with each cleft, and it was validated in a sample of fifty patients treated at the SUMA-CLC.
2 Methodology The classification of primary and secondary palates clefts is a problem of multiple attribute decision for making preference decisions (such as evaluation, prioritization, and selection) over the available alternatives that are characterized by multiple, usually conflicting, attributes [5]. In this regard, there are some studies that have been carried out to classify clef lip and palate. For example, to assess the clinical condition of the child according to the severity of cleft lip and palate by the International Classification of Functioning [6] or a classification which categorizes the fissure incorporating all the anatomical variants and specific complexity that can be presented in an abbreviated scheme [7]. In this work we use the Analytic Hierarchy Method (AHP) for developing a relevance factor (ρ) for each cleft. As in a hierarchy, the first level indicates the goal for the specific decision problem. In the second level, the goal is decomposed of several aspects known as dimensions. The criteria in the third level and the lower levels can follow up this principal to the division into other sub-criteria [5]. The hierarchy for cleft of primary and secondary palates is shown in Fig. 1. Note that there are three hierarchical levels. Level 1 contains unilateral clefts of the primary and secondary palates that occur independently in patients. In level 2, bilateral primary palate clefts were incorporated since they are the combination of two unilateral clefts. Note that this type of cleft has an associated bilaterality factor (FB = 1.5), since surgeons
Fig. 1. Hierarchy for cleft of primary and secondary palates
Quantification of a Lip and Palate Clefts Classification
75
argue that bilaterality is more than the simple sum of the two unilateral primary palate clefts. Level 3 contains the clefts that occur in both palates (primary and secondary) simultaneously, and therefore, they are the ones with the greatest surgical complexity. In this case, a labial-palatal factor (FLP = 2.0) was also associated. To know the relevance factor (ρ) for each class, after we built the hierarchy of evaluation criteria, we determined the weights to obtain the ranking for each class.
3 Results 3.1 Level 1. Unilateral Primary and Secondary Clefts The proposal in this work was worked together with SUMA-CLC surgeons. Initially, a previous classification that assigns a relevance factor to clefts [8], shown in column 3 of Table 1, was taken as a reference. A new grouping of the classes was proposed, and a more concentrated classification was obtained. It should be mentioned that each class grouped clefts with distinct characteristics (microform, incomplete or complete), and a modified relevance factor (ρ) was obtained for each class. For unilateral clefts of the primary palate, three classes were proposed (C1, C2, C3). Class C1 contains three clefts; C2 contains only one cleft, y C3 contains clefts without segment contact in the interval (1, 20 mm), generating 20 different clefts (one for each mm measurement apart). These clefts were divided into three groups according to the millimeters segments separation: (1, 4), (5, 12), (13, 20). In the case of the secondary palate, two classes were proposed: C4, which groups submucous and incomplete (1/3, 2/3) clefts, and C5, which groups complete clefts in GI, GII, GIII. The relevance factor of each of these five classes is shown in Table 1. 3.2 Level 2. Bilateral Primary Clefts Classification of bilateral clefts in the primary palate was defined with six classes (Table 2) combining the three main classes C1 , C2 and C3 (Table 1) of the primary palate. As mentioned, a bilateral cleft is more than just the sum of unilateral primary palate clefts, therefore, a bilateral factor (FB = 1.5) was defined that must be multiplied by the sum of the relevant factors that make up the bilateral cleft. The relevance factor for primary bilateral clefts is calculated with the Eq. (1). ρ = ρright + ρleft FB (1)
76
B. Gutiérrez-Sánchez et al.
Table 1. Classification of unilateral primary palate clefts and secondary palate clefts with their relevance factor (Level 1) Palate
Level 1 Description
ρ [8]
ρ (surgeons)
Class
Unilateral primary palate
Microform
1
3
C1
Incomplete 1/3
3
Secondary palate
Incomplete 2/3
6
Segment contact (sc)
12
12
C2
No segment contact (nsc) (1, 20 mm)
(13, 36)
15, 22, 32
C3
Submucous (no bifid uvula)
1
7
C4
Submucous (bifid uvula)
4
Incomplete 1/3
8
Incomplete 2/3
13
Complete GI (unilateral)
25
27
C5
Complete GI (bilateral)
28
Complete GII (unilateral)
34
Complete GII (bilateral)
37
Complete GIII (unilateral)
50
Complete GIII (bilateral)
55
36 53
To illustrate the calculation of the relevance factor of bilateral cleft of primary palate an example is shown in Table 2. The relevance factor (ρ) calculated for each class of Level 2 is shown in Table 3. Table 2. Relevance factor (ρ) calculation for three different clefts. Cleft description
Relevance Factor
Primary palate
Secondary palate
Left side
Right side
ρ
C1. Incomplete 1/3 ρleft = 3
C2. sc ρright = 12
ρ = (3 + 12) = 15 ρ = (15)(1.5) = 23
C2. Sc ρC2 = 12 C2. Sc ρC2 = 12
C2. sc ρC2 = 12
C4. Incomplete 1/3 ρC4 = 7
ρ = (12 + 7)2 = 38
C4. Incomplete 1/3 ρC4 = 7
ρ = (12 + 12)1.5 = 36 ρ = (36 + 7)2 = 86
Quantification of a Lip and Palate Clefts Classification
77
Table 3. Classification for bilateral cleft of primary palate and their relevance factor (Level 2) Level 2 Description
ρ (surgeons)
(C1 + C1)FB
9
C6
(C1 + C2)FB
23
C7
(C2 + C2)FB
36
C8
(C1 + C3)FB
27, 38, 53
C9
(C2 + C3)FB
41, 51, 66
C10
(C3 + C3)FB
45,56, 66, 71, 81, 96
C11
Class (alternative)
3.3 Level 3. Labial-Palatal Cleft (Primary and Secondary Palates) This classification refers to the clefts that occur in both palates (primary and secondary) simultaneously. The classification of labial-palatal clefts was defined in eighteen classes: six with unilateral primary palate cleft (C12, C17), and twelve with bilateral primary palate cleft (C18, C29), which are describe in Table 4. For this case, a labial-palatal factor was defined (FLP = 2.0). The relevance factor for labial-palatal cleft is calculated with the Eq. (2). (2) ρ = ρprimary + ρsecondary FLP It is important to mention that relevance factor calculation change depending on the primary palate cleft. For unilateral primary palate clefts, we sum the relevance factors of primary and secondary palates clefts and multiply by FLP . On the other hand, if the primary palate cleft is bilateral, we use Eq. 1, and then we add the relevance factor of the secondary palate to finally multiply by FLP . Some examples of the calculation of ρ for labial-palatal clefts are illustrated in Table 2. The relevance factor calculated for each class of Level 3 is shown in Table 4. 3.4 Validation of the Relevance Factor (ρ) To validate the relevance factor (ρ), fifty patients with different clefts treated at SUMACLC were chosen. In each case ρ was calculated and the results obtained are presented below, according to the corresponding level of each cleft. Level 1. A sample of fifteen clefts was obtained, distributed as follows: nine in the primary unilateral palate and six in the secondary palate (Table 5). Note that patients Pt7, Pt8, and Pt9, located in class C3, have a unilateral primary palate cleft with a segment separation of 5, 8, and 10 mm, respectively, and all obtained ρ = 22. This means that clefts have a similar surgical complexity and therefore, the three patients have the same surgical treatment. On the other hand, patients 10, 11, y 12, located in class C4, present different incomplete clefts in the secondary palate and obtained the same relevance factor (ρ = 7). As in the previous case, this means that the surgical treatment of patients is the same.
78
B. Gutiérrez-Sánchez et al.
Table 4. Classification for labial-palatal clefts (primary and secondary palates) and their relevance factor (ρ) (Level 3) Level 3 Description Labial-palatal Primary unilateral (C1 + C4)FLP palate (C2 + C4)FLP
Primary bilateral palate
ρ (surgeons)
Class (alternative)
20
C12
38
C13
(C3 + C4)FLP
44, 58, 78
C14
(C1 + C5)FLP
60, 78, 112
C15
(C2 + C5)FLP
78, 96, 130
C16
(C3 + C5)FLP
84, 98, 102, 116, 118, 136, 150, 170
C17
(C6 + C4)FLP
32
C18
(C7 + C4)FLP
60
C19
(C8 + C4)FLP
86
C20
(C9 + C4)FLP
68, 89, 119
C21
(C6 + C5)FLP
72, 90, 124
C22
(C10 + C4)FLP 95, 116, 146
C23
(C7 + C5)FLP
100, 118, 152
C24
(C8 + C5)FLP
126, 144, 178
C25
(C11 + C4)FLP 104, 125, 146, 155, C26 176, 206 (C9 + C5)FLP
108, 126, 129, 147, C27 159, 160, 177, 181, 211
(C10 + C5)FLP 135, 153, 156, 174, C28 186, 187, 204, 208, 238 (C11 + C5)FLP 144, 162, 165, 183, C29 186, 195, 196, 204, 213, 216, 217, 234, 238, 246, 247, 264, 268, 298
Quantification of a Lip and Palate Clefts Classification
79
Table 5. Description of nine patients with primary palate cleft, and six patients with secondary palate clefts (Level 1). Level 1 Unilateral primary palate
Primary palate
ρ
Class C1
Patient
Left
1
Inc 2/3
3
2
Microform
3
3 4
Wright
Inc 1/3
3
sc
12
5
Sc
12
6
4 mm
15
7
5 mm
22
8
8 mm
22
9 Secondary palate
Secondary palate
10 mm
C2
C3
22
10
Inc 1/3
7
11
Submucous
7
12
Inc 2/3
7
13
G-I
27
14
G-II
36
15
G-III
53
C4
C5
Level 2. A sample of six patients with bilateral primary palate clefts was studied. See its description in Table 6. Note that each patient is in each of the classes defined at this level.
Table 6. Description of six patients with primary bilateral clefts ρ
Level 2
Primary palate
Patient
Left
Wright
16
Microform
Microform
17
Inc 1/3
18
sc
19
Class
9
C6
Sc
23
C7
Sc
36
C8
Inc 1/3
7 mm
38
C9
20
3 mm
1 mm
45
C10
21
sc
8 mm
51
C11
80
B. Gutiérrez-Sánchez et al.
Level 3. A sample of twenty-nine patients with labial-palatal clefts (primary palate and secondary palate clefts simultaneously) was obtained, distributed as follows: sixteen with unilateral primary palate cleft and thirteen with bilateral primary palate cleft (Table 7). Table 7. Description of twenty-nine patients with labial-palatal clefts (sixteen with clefts of primary unilateral palate and thirteen with primary bilateral clefts). Level 3 Patient
Primary palate Left
Labial-palatal
Unilateral primary
ρ
Class
22
Inc 1/3
G-II
78
23
Inc 2/3
G-II
78
24
sc
G-I
78
G-II
96
25
sc
26
sc
G-III
130
27
3 mm
G-I
84
28
4 mm
G-I
84
29
9 mm
G-I
98
30
4 mm
G-II
102
G-II
116
31
5 mm
32
G-II
116
33
8 mm
12 mm
G-II
116
34
3 mm
G-III
136
35
5 mm
G-III
150
36
12 mm
G-III
150
37 Bilateral primary
Secondary palate
Wright C15
C16
C17
8 mm
G-III
150
38
Inc 2/3
3 mm
G-II
126
39
Inc 1/3
8 mm
G-I
129
40
Inc 2/3
Inc 2/3
G-II
90
41
Inc 2/3
Inc 2/3
G-III
124
42
Inc 1/3
sc
G-II
117
C24
43
sc
sc
G-II
144
C25
44
1 mm
1 mm
G-I
144
45
3 mm
3 mm
G-II
162
46
8 mm
2 mm
G-II
183
47
5 mm
8 mm
G-II
204
48
5 mm
3 mm
G-III
217
49
10 mm
4 mm
G-III
217
50
7 mm
10 mm
G-III
238
C21 C22
C29
Quantification of a Lip and Palate Clefts Classification
81
Note that some clefts located in different classes have the same ρ. For example, patients 22 and 24 located in classes C15 and C16, respectively, obtained ρ = 78. Likewise, different clefts located in the same class also obtained a different ρ. Such is the case of patients 27 and 28 located in class C17, both obtained a ρ = 84. Three more patients (31–33), also located in C17, obtained ρ = 116. Similarly, patients 35–37, located in C17, obtained a higher relevance factor (ρ = 150). The premise is that the clefts that obtain the same ρ imply that they have the same surgical complexity and therefore, the surgical treatment of the patient is the same. Once the relevance factors (ρ) were calculated for all clefts in the sample studied, a graph was made for visual analysis (Fig. 2). Note that in general the values increase as the cleft characteristics become more complex. The five main classes (C1, C5) have the lowest ρ values because they are the clefts with the lowest complexity. On the other hand, the clefts in the last classes (labial-palatal) have the highest ρ values. It is evident that there is a directly proportional relation between the cleft anatomical characteristics and the relevance factor (ρ), which is visualized through the exponential curve plotted. On the other hand, note that in class C17 there are six clefts with different relevance factor, because it can take eight values {84, 98, 102, 116, 118, 136, 150, 170} in this class, depending on the cleft description. The same situation occurs for C29. Multiple values of ρ occur for the last four classes of the classification (Table 3). In this regard, it is necessary to study the domain of ρ more precisely and model it together with plastic
Relevance Factor ρ
CLP Patients 240 230 220 210 200 190 180 170 160 150 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C15 C16 C17 C21 C22 C24 C25 C29 Classes
Fig. 2. Relevance factor of the sample of fifty patients with different cleft palates seen at SUMACLC.
82
B. Gutiérrez-Sánchez et al.
surgeons to obtain a domain with fewer admissible values, which represents the relevance of the corresponding cleft.
4 Conclusions A total of twenty-nine classes were defined: nine for primary palate (three unilateral and six bilateral), two classes for secondary palate, and eighteen classes for labial-palatal clefts. The relevance factors obtained for the clefts grouped in these classes are in the interval (3, 298), where three means the lowest relevance and 298 means the highest. Clefts that obtain the same value in the relevance factor (ρ) mean that they have an equivalent surgical complexity, and that the surgical treatment of the patient is the same. This is a result of the integration of the experience and diverse clinical criteria of the plastic surgeons who collaborate in The Center of Comprehensive Care of Cleft Lip Palate SUMA [3]. On the other hand, surgeons consider that the classification of cleft lip and palate proposed in this work is still very scattered. Therefore, work will continue towards a model that reduces the twenty-nine classes, without compromising that the class relevance factor (ρ) effectively represents the complexity of the cleft and the surgical treatment requires by the patient. Likewise, it will be necessary to work on a one-to-one correspondence where only one relevance factor (ρ) corresponds to each cleft in each class.
References 1. Oner, D.A., Tastan, H.: Cleft lip and palate: epidemiology and etiology. Otorhinolaryngology Head Neck Surg 5 (2020). https://doi.org/10.15761/OHNS.1000246 2. Palmero Picazo, J., Rodríguez Gallegos, M.F.: Cleft lip and palate. Current concepts. Acta Méd. Grupo Ángeles 17(4), 372–379 (2019). Available at: http://www.scielo.org.mx/scielo.php?scr ipt=sci_arttext&pid=S187072032019000400372&lng=es. Last accessed 08 September 2021. (In Spanish) 3. Children´s Wisconsin Hospital: Cleft lip and/or palate. https://childrenswi.org/medicalcare/fetal-concerns-center/conditions/infant-complications/cleft-lip-or-palate. Last accessed 11 June 2022 4. Smile Train Homepage: https://www.smiletrain.org/cleft-leadership-centers/suma 5. Tzeng, G.H.: Huang JJ Multiple attribute decision making methods and applications. CRC Press, Florida (2011) 6. Zamurayeva, A., Aldabergenova, T., Orynbayeva, B.A., Detochkina, V.: The main criteria for determining disability in children with congenital cleft upper lip and palate according to the international classification of functioning (ICF). Systematic Reviews in Pharmacy 11(4), 413–418 (2020) 7. Behar, J.M., Gallardo, D.D.L., De La Cerda, R.M., Zamudio, S.T.: A new classification approach: center of integral care of cleft lip palate “SUMA” center (México). Journal of Cleft Lip Palate and Craniofacial Anomalies 4(3), 20 (2017) 8. Ortiz-Posadas, M.R., Vega-Alvarado, L., Maya-Behar, J.: A new approach to classify cleft lip and palate. Cleft Palate and Craniofacial J 38(6), 545–550 (2001). https://doi.org/10.1597/ 1545-1569_2001_038_0545_anatcc_2.0.co_2
Artificial Intelligence Applied to Breast Cancer Classification Samara Acosta-Jiménez1(B) , Javier Camarillo-Cisneros2 , Abimael Guzmán-Pando2 , Susana Aideé González-Chávez1 , Jorge Issac Galván-Tejada3 , Graciela Ramírez-Alonso4 , César Francisco Pacheco-Tena1 , and Rosa Elena Ochoa-Albiztegui5 1 PABIOM Laboratory, Faculty of Medicine and Biomedical Sciences, Autonomous University
of Chihuahua, 31125 Chihuahua, Mexico [email protected], [email protected] 2 Laboratory of Computational Physical Chemistry, Faculty of Medicine and Biomedical Sciences, Autonomous University of Chihuahua, 31125 Chihuahua, Mexico 3 Academic Unit of Electrical Engineering, Autonomous University of Zacatecas, 98000 Zacatecas, Mexico 4 Faculty of Engineering, Autonomous University of Chihuahua, 31125 Chihuahua, Mexico 5 Radiology Department, Memorial Sloan Kettering Cancer Center, New York 10065, USA
Abstract. One in eight women is likely to develop breast cancer at some stage in her life, with a 12.5% average risk rate of developing breast cancer. Early detection and treatment are of vital importance to ensure the patient’s survival. Currently, mammography is the main diagnostic study to identify breast cancer. However, since mammography requires a human, medical radiologist, to make a diagnosis, it is prone to errors. Recently, deep learning techniques have proven to be a suitable tool for breast cancer classification and detection. Therefore, this research proposes an algorithm based on convolutional neural networks (CNN) for screening classification of cancer in mammography images. The evaluation results of the proposed algorithm respect state-of-the-art algorithms demonstrate competitive accuracy results of up to 99% and the fastest training time. Therefore, our algorithm is well suitable for automatic breast cancer detection using the public All-MIAS database. Keywords: AI · Breast-cancer · Mammography
1 Introduction Statistics show that one in eight women (approximately 12%) will develop breast cancer at some stage in her life. In 2019, 268,600 invasive and 62,930 non-invasive cancer cases were reported within the United States [1]. Furthermore, according to the National Institute of Statistics and Geography (INEGI) during 2019 in Mexico, 15,286 new cases were reported in women over 20 years of age, which translates into an incidence rate of 18.55 new cases per 100,000 women [2]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 83–93, 2023. https://doi.org/10.1007/978-3-031-18256-3_8
84
S. Acosta-Jiménez et al.
According to CONACEM [3], on March 1, 2019, 430 radiologists certified as subspecialists in breast imaging analysis were registered. On the other hand, INEGI reported that in 2015 there was a population of 51.4 million women [4], of which 17.1% corresponded to women ≥ 40 years old, approximately 10.5 million. Therefore, to diagnose all of them, radiologists specializing in breast cancer identification would have to interpret 67 studies daily for a year to compensate for the national demand, which is a strenuous procedure for any specialist and prone to errors. Due to the above, it is necessary to develop robust and automatic methods for the diagnosis and analysis of breast images. Therefore, this paper presents the development of a deep learning algorithm for screening cancer detection in mammography images. 1.1 Mammography as a Method of Breast Cancer Screening Mammography is one of the main diagnostic methods for the identification of breast cancer. This technique consists of obtaining images of several projections of the breast structure by X-ray exposure [5]. The resulting images are subsequently interpreted by a radiologist specializing in the breast who gives a classification using the BI-RADS score [6]. However, being a method that requires human intervention, the diagnosis may have a certain degree of subjectivity. In addition, there are other factors that increase the number of erroneous diagnoses [7, 8], such as the amount of work for the radiologist, the quality of the image, the young age of the patient, breast density, family history, current estrogen treatment, among others. Therefore, the importance of early detection of breast cancer has driven computational tools have been created and applied to speed up the analysis and improve the diagnosis [9–11]. One example is computer-aided detection (CAD), which employs image processing techniques and pattern recognition theory. However, CAD models lack accuracy in mammography diagnosis [12]. Another solution is artificial intelligence (AI) methods, specifically convolutional neural networks (CNNs) [12]. CNNs can learn to extract features and establish relationships across data to solve different task. Thus, many CNNs algorithms have been developed for detecting cancer in mammography images [13–16]. The main contributions of this work are described next. Regarding to our findings in the comparison of state-of-the-art CNNs algorithms carried out in this paper, our proposed model obtained the best performance on accuracy metric, the fastest training time (42 min), and it has less resource requirements. Therefore, our model demonstrates its robustness to classify cancer in mammography images with a fast development time and lower costs involved due to the reduced resource requirements.
Artificial Intelligence Applied to Breast Cancer Classification
85
1.2 Convolutional Neural Networks (CNN) Convolutional neural networks or CNNs are a breakthrough in the design of models for image classification and recognition. A CNN is a specialized type of artificial neural network that uses the mathematical operation of convolution to process data [17]. In addition, a CNN is an extension of an artificial neural network (ANN) model, which contains a set of interconnected artificial neurons [18]. The advantages of CNNs over ANNs are that they eliminate the need for handcrafted features and “learn” to extract features from the input data on their own. There are four basic operations in CNNs, which are described below: Convolutional layer: is the main element for a convolutional neural network. It is composed for a set of filters or kernels that are convolved with a feature map to obtain a new encoded feature map [17]. That is, the input feature G and the j-th convolution filters Ke are mixed to form a new feature map fm, by means of: fmk+1,j = Gk ∗ Kek,j
(1)
The j-th filters Ke used in the convolution will define the number of feature maps fm that form the depth of the k + 1 layer. Nonlinearity (ReLU): replaces negative pixel values with zero. Its purpose is to introduce nonlinearity to help approximate complex functions since most data encountered in the real world cannot be described by linear relationships [18]. Figure 2 shows the mapping performed by the ReLU function on pixel i. Other nonlinear functions such as the Sigmoid or the Tanh function can be used. 0 if i < 0 f ReLU(i) = (2) i if i ≥ 0 Clustering or subsampling: reduces the size of the feature map through a sampling process that preserves the most distinctive or prominent features of the map [17]. The two types of pooling layers are max-pooling and average-pooling. In this paper, Max-Pooling is used in the CNN network which is defined by: fmp = max(fm(x + w, y + h)); w = {1, ..., W }, h = {1, ..., H }
(3)
where (x,y) are spatial positions of the feature map fm. W and H are the sizes in width and height of the window or max operator. (w,h) are the positions that are shifted by traversing the window size to generate the new feature map fmp. Classification: its objective is to identify classes based on the features extracted by the convolutional layers. Generally, the classification stage is composed of fully connected layers and a softmax activation function [19]. This softmax function yields a probability value between 0 and 1 for each of the classification labels that the model is trying to predict. The winning class will be the one with the highest probability. The output equation of the classification layer is defined by:
86
S. Acosta-Jiménez et al.
Os = fsoftmax (
N
(Xn · Wn + bn ))s
(4)
n=1
where fsoftmax is the softmax function, bn is the bias, Xn is the input vector that is multiplied by the weights Wn for each neuron n and N is total number of neurons used to generate the OS output, where S = {cancer, non-cancer}.
2 Materials and Methods The CNN receives a feature map in tensor form, where the initial size in width, height and depth corresponds to the dimensions of the image. In mammograms, grayscale images have only one dimension in depth and the image has values from 0 to 255. The value 0 indicates a totally black pixel and 255 indicates a totally white pixel [20]. 2.1 Digital Image Processing The database used in this work was ALL-MIAS [21], which were conditioned using the GDCM library. The size of each image was adjusted as a first standardization step to 1024 × 1024 pixels by means of Phyton 3 [22]. PGM format was used for the analysis, which in turn was derived from the DICOM clinical image. The DCMTK library was used in this step. The ALL-MIAS database were subjected to a data augmentation process for increasing the number of mammographic images in the CNN training. Each image was rotated in a factor of 1 and 2 degrees until reaching its initial position. Resulting in 180 images for the factor 2, due to the generation of one image each to two degrees, and 360 images for the factor 1, by the generation of one image every one degree. 2.2 Development Environment For the CNN implementation, a virtual environment was created through Anaconda [23] and the following libraries were installed: Spyder [24], pandas [25], pillow [26], matplotlib [27], tqdm [28], opencv-contrib-Python [29], sklearn [30] and keras [31]. The experiments in this study were conducted on a Linux workstation equipped with an NVIDIA RTX 3080 Ti GPU card with 12 GB of RAM.
Artificial Intelligence Applied to Breast Cancer Classification
87
2.3 Proposed Algorithm Figure 1 shows the structure of the CNN network we worked with. The model was trained with the public database AllMIAS [21], which consists of 330 mammography images with approximately 66% normal and 33% with abnormalities. For the training of the proposed CNN, 80% of the mammograms were used for training and the rest for validation. The model was trained to predict whether the breast image belonged to cancer or non-cancer class. The convolutional layers were designed with 3 × 3 filters, while the clustering layers with 2 × 2 filters. The activation function used for the convolution layers was ReLU, while a Sigmoid function was used for the classification layers [32]. In addition, dropout and flatten layers were used. The dropout layers select random values of coefficients within the filters and set them to zero, which helps to reduce overfitting. The flattened layer is used to generate a one-dimensional space that facilitates the transition from convolutional to fully connected layer characteristics. 2.4 Training Models The model was trained according to the number of total images after the increase of data and the number of epochs, obtaining a total of 4 configurations, in order to see the behavior of the Accuracy and Loss graphs. Table 1 shows the training configurations that were applied to the CNN. Table 1. Configuration of training models Configuration
Databaste
Loss function
Optimizer
Learning rate
Epochs
A1
MIAS
Binary Crossentroy
Adam
0.001
50
A2
MIAS
Binary Crossentroy
Adam
0.001
50
A3
MIAS
Binary Crossentroy
Adam
0.001
30
A4
MIAS
Binary Crossentroy
Adam
0.001
30
A1, A3: Data enlargement, 180 new images per each image of data base ALL-MIAS. A2, A4: Data enlargement, 360 new images per each image of data base ALL-MIAS.
88
S. Acosta-Jiménez et al.
3 Results Table 2 shows the values of the validation accuracy of each configuration at the end of the training stage. We experimentally reduced the number of epochs to improve the model training time and to verify the performance behavior. When using 50 epochs the accuracy in each of the results was 95.5% for A1 and 99.57% for A2, when decreasing to 30 epochs, it was 90.77% for A3 and 99.00% for A4. The training results are shown in the accuracy graphs, which are shown in Figs. 2 and 3. In the accuracy graphs (Figs. 2 and 3), the orange line indicating the validation accuracy and the blue line indicates the training accuracy. These figures are presented to visualize models’ generalization. It can be noted that A1 and A3 configurations show overfitting in their accuracy graphs; the validation accuracy line (orange) is under the training accuracy line (blue). On the other hand, A2 and A4 configurations, show adequate behavior and therefore better generalization. These results showed that the A2 configuration with 50 epochs turned out to be optimal because it has the best performance, and no overfitting is observed. Table 3 is presented to visualize and compare our model implementation details respect other CNN models applied to cancer detection. It can be noted that our results are competitive and even better in accuracy than the models shown in Table 3. Additionally, our model has the fastest training time, and it does not require huge amount of GPU resources. Therefore, these findings imply a less development time and a less cost involved in a real application. Table 2. Accuracy of the different configurations of the CNN-Optimizer Adam Configuration
Training base
Loss function Optimizer
Learning rate
Epochs
Accuracy (%)
A1
MIAS
Binary Crossentroy
Adam
0.001
50
95.5
A2
MIAS
Binary Crossentroy
Adam
0.001
50
99.57
A3
MIAS
Binary Crossentroy
Adam
0.001
30
90.77
A4
MIAS
Binary Crossentroy
Adam
0.001
30
99.00
2021
2015
2016
2017
2021
Our
Xi et al. [13]
Jiao et al. [16]
Hang et al. [14]
Sun et al. [15]
Year
12
7
1500 × 2000
8
152
10
Number of layers
3000 × 4000
227 × 227
2244 × 224
1024 × 1024
Image Resolution
3
2
2
151
5
Number of Conv-Pool
1
2
2
1
1
Number of fully connected layers
1024
4M
20.4K
1000
57K
Full connected layer size
3,5
3
1,3,7
1,3,7
3,2
Filter sizes
✘
✘
✓
✓
✓
Data Augmentation
✘
✘
✘
✓
✓
Dropout
Table 3. Comparison of implementation details with state-of-the-art algorithms.
1-GTX 1080 Ti
1-GTX TITAN
No reported
8-GPUs
1-RTX
GPU
-
10 h
22 h
2–3 weeks
40 min
Training Time
85.82
66
96.70
90.22
99.57
Accuracy (%)
Artificial Intelligence Applied to Breast Cancer Classification 89
90
S. Acosta-Jiménez et al.
Fig. 1. CNN structure.
Artificial Intelligence Applied to Breast Cancer Classification
91
Fig. 2. Evaluation of the accuracy metric in training and validation; A) the A1 configuration, and B) A2 configuration.
Fig. 3. Evaluation of the accuracy metric in training and validation; C) the A3 configuration, and D) A4 configuration.
4 Conclusions In this study, we proposed a CNN algorithm that can be used to aid radiologists in detecting normal and abnormal mammograms. The designed CNN algorithm showed high accuracy results for the classification of benign and malignant mammography using the database ALL-MIAS. We proposed four training configurations that performed above 90% on the accuracy metric, being the best option the A2 configuration with a performance of 99.57%. We demonstrate that our model results are competitive and even better to those stateof-the-art algorithms. Additionally, our proposed algorithm obtained the fastest training time without the necessity of a huge number of computational resources. Therefore, these findings indicate that our A2 configuration might be useful as a screening tool for mammogram classification reducing the workload of the breast radiologist in the analysis of breast images. In addition, it represents an automatic and fast method to obtain a diagnosis. In future work, the algorithm is expected to be applied to larger mammogram databases, such as screening mammography and the image retrieval in medical applications. Additionally, the possibility of implementing our model in mammography images of the Mexican population is foreseen.
92
S. Acosta-Jiménez et al.
References 1. Breastcancer.org - Breast Cancer Information and Support. https://www.breastcancer.org/ Accessed 03 Aug. 2022 2. EAP_LUCHACANCER2021.pdf. [Online]. Available: https://www.inegi.org.mx/conten idos/saladeprensa/aproposito/2021/EAP_LUCHACANCER2021.pdf. Accessed: 03 Aug. 2022 3. CMRI | Inicio. https://www.cmri.org.mx/ Accessed 03 Aug. 2022 4. I.N. de E. y Geografía (INEGI): Población. Censos y conteos. Población y Vivienda (Jan. 01 1910). https://www.inegi.org.mx/temas/estructura/default.html#Informacion_general,% 20last%20accessed%202020/02/25 Accessed 03 Aug. 2022 5. Hmida, M., Hamrouni, K., Solaiman, B., Boussetta, S.: Mammographic mass segmentation using fuzzy contours. Comput. Methods Programs Biomed. 164, 131–142 (2018). https://doi. org/10.1016/j.cmpb.2018.07.005. Oct. 6. Spak, D.A., Plaxco, J.S., Santiago, L., Dryden, M.J., Dogan, B.E.: BI-RADS® fifth edition: a summary of changes. Diagn. Interv. Imaging 98(3), 179–190 (2017). https://doi.org/10.1016/ j.diii.2017.01.001. Mar. 7. Ventura-Alfaro, C.E.: Measurements errors in screening mammogram interpretation by radiologists. Rev. Salud Publica Bogota Colomb. 20(4), 518–522 (2018). https://doi.org/10.15446/ rsap.V20n4.52035. Aug. 8. Aguilar-Torres, C.: Current overview of screening for the detection of breast cancer in the state of Chihuahua, Mexico. Ginecol. Obstet. México 89, 91–99 (2021). https://doi.org/10. 24245/gom.v89i2.4466. Feb. 9. Xia, Q., et al.: Differential diagnosis of breast cancer assisted by S-detect artificial intelligence system. Math. Biosci. Eng. MBE 18(4), 3680–3689 (2021). https://doi.org/10.3934/mbe.202 1184. Apr. 10. Akselrod-Ballin, A., et al.: Predicting breast cancer by applying deep learning to linked health records and mammograms. Radiology 292(2), 331–342 (2019). https://doi.org/10.1148/rad iol.2019182622. Aug. 11. Tran, W.T., et al.: Computational radiology in breast cancer screening and diagnosis using artificial intelligence. Can. Assoc. Radiol. J. J. Assoc. Can. Radiol. 72(1), 98–108 (2021). https://doi.org/10.1177/0846537120949974. Feb. 12. Lehman, C.D., Wellman, R.D., Buist, D.S.M., Kerlikowske, K., Tosteson, A.N.A., Miglioretti, D.L.: Diagnostic accuracy of digital screening mammography with and without computeraided detection. JAMA Intern. Med. 175(11), 1828–1837 (2015). https://doi.org/10.1001/jam ainternmed.2015.5231. Nov. 13. Xi, P., Shu, C., Goubran, R.: Abnormality detection in mammography using deep convolutional neural networks. In: 2018 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 1–6 (Jun. 2018). https://doi.org/10.1109/MeMeA.2018.843 8639 14. Hang, W., Liu, Z.: GlimpseNet : Attentional Methods for Full-Image Mammogram Diagnosis (2017). https://www.semanticscholar.org/paper/GlimpseNet-%3A-Attentional-Methodsfor-Full-Image-Hang-Liu/267ca984017bcadd47cff605bcc5e5a4d136a540 Accessed 03 Aug. 2022 15. Sun, L., Sun, H., Wang, J., Wu, S., Zhao, Y., Xu, Y.: Breast mass detection in mammography based on image template matching and CNN. Sensors 21(8), 2855 (2021). https://doi.org/10. 3390/s21082855. Apr. 16. Jiao, Z., Gao, X., Wang, Y., Li, J.: A deep feature based framework for breast masses classification. Neurocomputing 197, 221–231 (2016). https://doi.org/10.1016/j.neucom.2016.02.060. Jul.
Artificial Intelligence Applied to Breast Cancer Classification
93
17. Aggarwal, C.C.: Neural Networks and Deep Learning: A Textbook. Springer (2018) 18. Ciaburro, G., Venkateswaran, B.: Neural Networks with R: Smart models using CNN, RNN, deep learning, and artificial intelligence principles. Packt Publishing Ltd. (2017) 19. CS231n Convolutional Neural Networks for Visual Recognition. https://cs231n.github.io/con volutional-networks/ accessed 31 Jul 2022 20. Gonzalez, R.C., Woods, R.E.: Digital Image Processing. Prentice Hall (2008) 21. The mini-MIAS database of mammograms. http://peipa.essex.ac.uk/info/mias.html accessed 12 Nov. 2021 22. Welcome to Python.org, Python.org. https://www.python.org/ Accessed 22 Sep. 2021 23. Anaconda | The World’s Most Popular Data Science Platform. https://www.anaconda.com/ Accessed 22 Sep. 2021 24. Home — Spyder IDE. https://www.spyder-ide.org/ Accessed 22 Sep. 2021 25. pandas - Python Data Analysis Library. https://pandas.pydata.org/ Accessed 07 Dec. 2021 26. Pillow. https://pillow.readthedocs.io/en/stable/index.html Accessed 07 Dec. 2021 27. Matplotlib — Visualization with Python. https://matplotlib.org/ Accessed 07 Dec. 2021 28. tqdm: Fast, Extensible Progress Meter. [MacOS, MacOS :: MacOS X, Microsoft, Microsoft :: MS-DOS, Microsoft :: Windows, POSIX, POSIX :: BSD, POSIX :: BSD :: FreeBSD, POSIX :: Linux, POSIX :: SunOS/Solaris, Unix]. Available: https://tqdm.github.io Accessed 07 Dec. 2021 29. opencv-contrib-python: Wrapper package for OpenCV python bindings. [MacOS, Microsoft :: Windows, POSIX, Unix]. Available: https://github.com/skvark/opencv-python Accessed: 07 Dec. 2021 30. scikit-learn: machine learning in Python — scikit-learn 1.0.1 documentation. https://scikitlearn.org/stable/ Accessed 07 Dec. 2021 31. Keras: the Python deep learning API. https://keras.io/ Accessed 07 Dec. 2021 32. Waoo, A.A., Soni, B.K.: Performance Analysis of Sigmoid and Relu Activation Functions in Deep Neural Network. In: Sheth, A., Sinhal, A., Shrivastava, A., Pandey, A.K. (eds.) Intelligent Systems. AIS, pp. 39–52. Springer, Singapore (2021). https://doi.org/10.1007/978-981-162248-9_5
Computational Chemistry as an Educational Tool in Health Sciences Alexica Celine Márquez-Barreto1 , Celia María Quiñones-Flores1 Graciela Ramírez-Alonso2 , Gabriela Sámano-Lira1 , and Javier Camarillo-Cisneros1(B)
,
1 Laboratorio de Química Física Computacional, Facultad de Medicina y Ciencias Biomédicas,
Universidad Autónoma de Chihuahua, 31125 Chihuahua, Chihuahua, Mexico [email protected] 2 Facultad de Ingeniería, Universidad Autónoma de Chihuahua, 31125 Chihuahua, Chihuahua, Mexico
Abstract. Experimentation as a teaching technique allows the understanding and relationship of concepts as well as the acquisition of problem-solving skills. Computational chemistry is a tool for studying chemical phenomena through computational experiments. The use of simulation in chemistry and biochemistry education is evolving the teaching techniques and developing computational skills. Teaching chemistry and biology through simulations and structural analysis is mainly limited to graduate students. However, we are moving toward a future where computational skills, including programming and simulation, will no longer be optional. In the present research, we use a pharmaceutical example for computational modeling and molecular docking to study and design drugs. Physicochemical characterization of the drug Remdesivir was carried out to demonstrate that the acquisition and learning of theoretical concepts are more practical when performing computational experiments. Keywords: Chemistry · Computational chemistry · Education · Remdesivir · SARS CoV-2
1 Introduction Basic, middle, and higher education have implemented traditionalist pedagogical models and educational strategies in which rote and repetitive methods acquire learning. These methods result in the adaptation of young minds, such as those of the students with difficulties relating knowledge or generating skills that allow problem-solving. In the case of chemistry applied to health sciences, learning strategies that include experimentation and technological recourses within a laboratory are very important for understanding concepts [1].
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 94–103, 2023. https://doi.org/10.1007/978-3-031-18256-3_9
Computational Chemistry as an Educational Tool in Health Sciences
95
Computational chemistry can be conceptualized as the application of computation in chemistry, biology, and physics for studying atoms, molecules, and macromolecules through three-dimensional modeling representation. Thanks to computational chemistry, it is possible to study and analyze physicochemical properties [2, 3] usually impossible, unsuitable, or impractical in experimental laboratories. This fact is due to the necessary conditions for the experiments, such as temperature, radioactivity, vacuum conditions, etc. Another limitation of experimental studies is the high investment or resources in infrastructure and reagents necessary for such experiments [4, 5]. Computational chemistry has been a significant contribution to the pharmaceutical industry, and an example of this is the current pandemic caused by the SARS-CoV-2 virus. Very little structural information was available for developing drugs and vaccines at the beginning of the pandemic. Using computational tools made it possible to know the proteins involved in viral infection, the viral structure, and its possible interactions with different drugs [6, 7]. For example, among the drugs used for the treatment of COVID19, a disease caused by the SARS CoV-2 virus, is the antiviral drug called Remdesivir, the first drug approved by the U.S. Food and Drug Administration (FDA) [8, 9]. The computational studies allowed performing simulations with accurate measurements on the interactions of Remdesivir with the SARS CoV-2 virus. The results generated by these techniques provided information for understanding the virus or planning future experiments [6]. This research presents the results of the structural analysis of the drug Remdesivir with the primary objective of showing the applicability of computational tools in teaching basic chemistry concepts. As a secondary objective, we intend to present an applicable workflow for students and professionals in chemistry teaching. 1.1 Importance of the Study There is an interest in seeking continuous improvement and increasing the quality of teaching, part of the search involves the implementation of didactic strategies. The application of computational methods provides significant learning by incorporating models approximating the molecular nature. Another advantage of integrating computational methods is the learning through visual representations that help understand the structure of matter and its behavior when interacting with other complex elements. In 2005, the National University of San Luis conducted a study to analyze how computational chemistry improved group performance in the subject of chemistry. The first results were deficient compared to those obtained after implementing computational chemistry. Nevertheless, the study concluded that the students surpassed their level by using learned concepts and cognitive processes and complementing what they observed with what they had previously learned [10]. Another study, conducted in 2021 in Chilean Patagonia, analyzed the pedagogical benefits obtained by students and teachers, through the development of a research project in computational chemistry. They demonstrated that computational chemistry applied in research projects develops scientific-technological skills in undergraduate and graduate students [11].
96
A. C. Márquez-Barreto et al.
Both studies show the positive impact of computational chemistry in the educational field. In the present research, we demonstrate the ease of understanding theoretical concepts supported by experimental results by obtaining and explaining the physicalchemical properties of Remdesivir.
2 Methodology The open-access database PubChem was used to obtain the different structures of remdesivir previously published by several researchers. The metabolic pathway used to perform the computational experiment is shown in Fig. 1 with five different conformations of remdesivir. The FHI-aims software was used to minimize the structures within the Density Functional Theory (DFT); thus, obtaining the most stable conformation of the molecule. For the latter case, open-access software like ABINIT or SIESTA could be used as an alternative. On this occasion, it was decided to plot the molecular orbitals as a character with crucial chemical significance that only theoretical or computational techniques can obtain. To prepare the selected molecules and perform the molecular docking of each system with the SARS CoV-2 virus, we used the open-access software Chimera and Auto-Dock Vina Suite. The SARS CoV-2 virus was downloaded from the freely accessible Protein Data Base (code 7bv1). Chimera software (open-access software) was used to clean up the remdesivir and SARS CoV-2 virus structures by removing residues that were not interesting for molecular docking. Normalization of the remdesivir structures within AutoDock 4 consisted of adding non-polar hydrogens, binding the hydrogens, and adding Kollman charges. Then, the SARS CoV-2 virus was loaded as a ligand and prepared by adding hydrogens, binding hydrogens, and adding Gasteiger charges. With the assembly complete, it was performed the docking of the drug-virus complex in AutoDock 4 and AutoDock Vina software (open-access software).
Fig. 1. Representation of the research process and the software used at each step.
Computational Chemistry as an Educational Tool in Health Sciences
97
All calculations were performed on a desktop computer with an i7-processor with 32 GB of RAM and a dedicated video card Nvidia GTX-1080 Ti with 3584 CUDA cores. However, the minimum computer equipment can be a school computer with an i3 processor and 8 GB of RAM for molecules similar to or smaller than the one used in the present work.
3 Results 3.1 Structure Analysis Results Figure 2 shows the metabolic pathway represented by the most stable structures obtained by the FHI-aims software. PubChem was used to search and download the molecules.
Fig. 2. The metabolic pathway of Remdesivir represented by the most stable structures obtained by the FHI-aims software.
Figure 3 compares the unloaded structure (Fig. 3a) and the structure obtained by energy minimization (Fig. 3b). The visualization of the more stable structure lets us know which structure requires less energy for its conformation.
Fig. 3. Molecular structure of Remdesivir obtained from the open access database PubChem (a) and the minimized molecular structure.
98
A. C. Márquez-Barreto et al.
In Fig. 4, the energy of the molecular orbitals is shown. In the upper part of the diagram, the empty boxes represent places available for an electron to arrive when it receives additional energy to that possessed by the molecule. The empty boxes are known as LUMO and refer to the molecule’s ability to accept electrons. In the case of the lower part of the diagram, there are boxes filled with a maximum of two electrons, which are necessarily represented by an up arrow and a down arrow. These full boxes are known as HOMO and indicate the molecule’s capacity to donate electrons. Boxes 1 to 154 represent all the energy levels the electrons have accommodated. Level 1 is the most stable, so it is difficult for an electron to migrate from this location. On the other hand, the box numbered 154 is less stable, and its electrons can migrate easily by adding energy to the molecule by either adding heat, electric field, or magnetic field of some mass next to it.
Fig. 4. Molecular orbitals of Remdesivir. The energy values of each orbital are shown, as well as the HOMO and LUMO orbitals.
Figure 5 shows an electronic structure plot, i.e., where the electrons are most likely to be found in the antiviral molecule. The red and blue colors are equivalent to the up and down arrows in the representation we commonly see in traditional chemistry. Figures 5a, 5b, 5c represent the LUMO-2, LUMO-1, and LUMO levels, respectively. While Figs. 5d, 5e, 5f represent the HOMO, HOMO-1, and HOMO-2 levels, respectively. As can be seen in the LUMO group of levels, there are no two equal areas where the electrons of the molecular orbitals can be; each electron has its place where it is more likely to be located and thus create bonds. This same observation can be extended to the HOMO orbitals. The volume occupied by the HOMO orbital is very similar to the LUMO, so when receiving energy, the electrons continue to occupy volumes positioned in practically the same region of the molecule, giving a sense of bond creation. Therefore, we should expect the antiviral remdesivir molecule to bind at the left end, shown by the three HOMO levels, to some protein or RNA of the SARS-COV-2 virus. This binding requires 5.234 eV (electronic GAP shown in Fig. 5) to take place, so it would be expected to occur only at temperatures far from the freezing point, such as, for example, that of the human body.
Computational Chemistry as an Educational Tool in Health Sciences
99
Fig. 5. Molecular orbitals of Remdesivir and GAP energy. The red clouds show the surfaces that define the volume where the electrons and thus the chemical bonds are located.
Finally, Fig. 6 shows an illustrative table of the molecular dockings performed with the different structures of the remdesivir drug and the SARS CoV-2 viral complex. AutoDock allows the visualization of the different possible conformations between molecules and the chemical properties of these conformations, such as binding energy, electrostatic energy, and ligand efficiency, among other properties that are possible to know through molecular docking. Figure 7 shows the binding energy, total internal energy, and torsional energy and Fig. 8 shows the ligand efficiency and electrostatic energy. The binding energy is the energy required to break or form a bond. For example, in Table 1, we can see that the binding energy for the Remdesivir structure in its triphosphate form is −5.37 eV. In other words, it is required −5.37 eV for this structure to bind with the SARS CoV-2. Another concept we can visualize through molecular docking is torsional stress, defined as the resistance to torsion of the bonds, which occurs when the bonds are in a conformation different from their most stable conformation. For example, observing Fig. 7, we can determine that the conformation with the lowest torsional energy was that of Remdesivir in its first conformation (RDV).
100
A. C. Márquez-Barreto et al.
Fig. 6. Molecular structures of Remdesivir (RDV), nucleotide monophosphate (RMP), nucleotide triphosphate (RTP), and RN nucleoside (RN) obtained with FHI-aims software and their respective molecular docking with the SARS CoV-2 viral complex obtained with AutoDock 4 and AutoDock Vina software.
5.37
Energy (eV)
6 5
Binding energy
4 3
Total internal energy
2 1
Torsional energy
0 0
1
2
3
4
Molecules Fig. 7. Binding energy, total internal energy, and torsional energy obtained through AutoDock 4 and AutoDock Vina for RDV (1), RMP (2), RTP (3), and RN (4) molecules.
Energy (eV)
Computational Chemistry as an Educational Tool in Health Sciences
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
101
Ligand efficiency Electrostatic energy 0
1
2
3
4
Molecules Fig. 8. Ligand efficiency and electrostatic energy were obtained through AutoDock 4 and AutoDock Vina for RDV (1), RMP (2), RTP (3), and RN (4) molecules.
3.2 Pedagogical Results Molecules can be found in databases indexed by familiar names. Using trade names of drugs for teaching purposes may be more user-friendly as this is the information they come in contact with daily. Using IUPAC names is helpful for structure identification. However, it can be challenging to memorize, and it is just the learning that we propose changing for meaningful learning. Energy minimization is finding an arrangement in space for an atomic structure where the force between atoms is acceptably close to zero. Students must know and use the most energetically stable structures since they require less energy for their conformation and correspond to their natural geometry or the closest to it, allowing them to obtain results closer to the real ones. Teachers using blackboard drawings usually teach concepts such as HOMO-LUMO and GAP energy. However, with the support of computational chemistry, students can come to the construction of basic concepts through experimentation. For example, students can visualize the number of orbitals and their energy, as shown in Fig. 4. In the same way, they can visualize how much energy is required for electronic migration and, therefore, for the creation of new bonds. With this methodology, teachers and students can perform structural analyses of different molecules and learn about concepts that can be complicated to learn only theoretically. In addition, students can compare Remdesivir against some antiviral options against SARS-CoV-2 since evaluating its chemical activity would be a matter of comparing the electronic GAP. Small GAP energy indicates higher chemical activity and thus lower molecular stability, while high GAP energy indicates higher excitation energy of the excited state and thus good chemical stability.
4 Conclusions 4.1 Discussion In the structural analysis of the Remdesivir, it was possible to visualize the most energetically stable structure, analyze the molecular orbitals with their respective energies, and analyze the molecular docking between Remdesivir and SARS CoV-2 virus.
102
A. C. Márquez-Barreto et al.
Although traditional chemistry allows us to obtain calculations and results based on experimentation, computational chemistry allows the analysis with the greatest possible detail. Computational methods are a valuable tool in studying molecules and their properties, as they have provided results for current problems such as emerging diseases and the design or study of drugs. We have learned chemistry concepts through text and two-dimensional structures to represent molecules. However, the molecular structure is three-dimensional, and it is crucial to describe the molecular bonds, their distances, angles, and arrangements in space. It is important to emphasize that computational chemistry should not be considered a substitute for traditional chemistry in the laboratory or classroom. However, computational chemistry allows adding concepts and increasing the understanding of traditional chemistry. Therefore, it would be ideal for creating an adequate balance between chemical knowledge and the contribution of computational development. Another point to consider is that using computational chemistry requires an investment of resources such as computer equipment, appropriate software, and knowledge of intermediate concepts of physics and chemistry. However, if these resources are unavailable, the published results that computational chemistry has generated through scientific journals can be used. 4.2 Author’s Comment Among the objectives of current pedagogical models is the training of professionals prepared with tools that allow the solution of societal problems. Enriching traditional teaching with current technologies can prepare students from an early age not only to improve the understanding or relationship of concepts but also to imagine and visualize solutions to current problems such as the current SARS CoV-2 pandemic, antibiotic resistance, or the lack of effective drugs for emerging diseases. Another point of view is that the global pandemic and the switch to remote education forced teachers to use all the available technologies to teach concepts previously approached with other traditional methods. As a result, there has been a notable progression in teachers and students using available technologies, moving toward a future where computational skills, including programming and simulation, will no longer be considered optional but a required skill.
References 1. Espinoza, C.: Estrategias empleadas desde la quimica computacional para la enseñanza de la quimica organiza en estudiantes del plan comun de secundaria (9 Marzo 2020). [Online]. https://es.scribd.com/document/450799395/Estrategias-empleadas-desde-la-quimica-com putacional-para-la-ensenanza-de-la-quimica-en-estudiantes-de-secundaria-1. Accessed Septiembre 2021 2. Lewars, E.: Computational Chemistry: Introduction to the Theory and Applications of Molecular and Quantum Mechanics, 2nd ed. Kluwer Academic Publishers (2003)
Computational Chemistry as an Educational Tool in Health Sciences
103
3. Garofalo, M., Grazioso, G., Cavalli, A., Sgrignani, J.: How computational chemistry and drug delivery techniques can support the development of new anticancer drugs. Molecules 25(7), 1756 (2020) 4. Saldivar, F., Prieto, F., Medina, J.: Descubrimiento y desarrollo de farmacos: un enfoque computacional. Educacion Quimica 28(1), 51–58 (1 Enero 2017) 5. Rossi, J.-P., Mondelo, N., Cascone, O.: Desarrollo de nuevos farmacos mediante inteligencia artificial. In: Inteligencia Artificial: una mirada multidisciplinaria, Buenos Aires, Academia Nacional de Farmacia y Bioquimica, pp. 239–257 (2019) 6. Narkede, R., Cheke, R., Ambhore, J., Shinde, S.: El estudio del acoplamiento molecular de posibles farmacos candidatos que muestran actividad anti-COVID-19 mediante la exploracion de dianas terapeuticas del SARS CoV-2. Eurasian Journal of Medicine and Oncology 4(3), 185–195 (2020) 7. Yin, W., et al.: Structural basis for inhibition of the RNA-dependent RNA-polymerase from SARS CoV-2 by remdesivir. Science 368(6498), 1499–1504 (2020). June 8. Sanchez Contreras, G.: La FDA aprueba el primer tratamiento para el COVID-19 (22 Octubre 2020). [Online]. https://www.fda.gov/news-events/press-announcements/la-fda-aprueba-elprimer-tratamiento-para-el-covid-19. Accessed Septiembre 2021 9. U.S. Food and Drug Administration: U.S. Food and Drug Administration (10 octubre 2020). [Online]. https://www.fda.gov/news-events/press-announcements/la-fda-aprueba-el-primertratamiento-para-el-covid-19. Accessed 7 Marzo 2022 10. Villagra, S., Guidugli, S., Masman, M.: Propuesta de enseñanza para la Fisicoquimica Computacional y Modelizacion Molecular en tercer año de Polimodal (Septiembre 2005). [Online]. http://sedici.unlp.edu.ar/handle/10915/22833. Accessed Septiembre 2021 11. Oyarzun, A.M., Moya, I.I., Navarro, D.J.: Formacion en quimica computacional y sus aplicaciones a traves de un proyecto de investigacion desarrollado en la Patagonia chilena. Formacion Universitaria 15(2), 103–116 (2022) 12. McDonald, A., Roberts, R., Koeppe, J., Hall, B.: Undergraduate structural biology education: a shift from users to developers of computation and simulation tools. Curr. Opin. Struct. Biol. 72, 39–45 (2022)
A Gene-Community Overview of Transcriptional Dynamics During Neurodevelopment Gustavo Guzm´ an1 , Elsa Maga˜ na-Cuevas1 , Juan Serna-Grill´ o1 , 1,2 1 , Hugo V´elez-P´erez , Rebeca Romo-V´azquez1 , Omar Paredes and Jose Alejandro Morales1(B) 1
Bioengineering Translational Department, CUCEI, Guadalajara University, 44430 Guadalajara, Mexico [email protected] 2 ITESM Campus Guadalajara, 45201 Zapopan, Mexico
Abstract. The brain is a complex network of anatomic functional modules called brain circuits. A brain circuit comprises structures that share transcriptional dynamics, enabling them to interact together to meet cognitive functions. The brain circuit emergence and its complexification begins at the early stages of neurodevelopment. Studies that approach transcriptional dynamics only focus on a few genes in limited structures. Thus, complex transcriptional behaviors underlying the brain circuits’ emergence remain hidden. In this work, we examine the whole-brain spatiotemporal transcriptomes to capture transcriptional dynamics during neurodevelopment. Through hyperbolic representations like Poincar´e maps, we illustrate a transcriptional machinery difference in embryonic development and adulthood. Examples like dendrogenesis and axonogenesis indicate that this difference results from constituting and later striving transcriptional dynamics. This qualitative analysis evidences expression dynamics that lead to brain circuit emergence.
Keywords: Neurodevelopment
1
· Brain circuits · Poincar´e maps
Introduction
Brain architecture is a complex network made up of anatomical connections among neuronal elements. This network develops a hierarchical organization where anatomical structures and functional circuitry emerge [1]. Connectomics studies such brain organization to unveil the principles which enable the brain to reach its diverse cognitive functions [2]. There are multiple approaches to model brain organization that spans from electrical and functional activity (macroscale) to temporal and spatial transcriptomics (microscale). Some works aim to integrate both scales to settle the brain c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 104–110, 2023. https://doi.org/10.1007/978-3-031-18256-3_10
Transcriptional Dynamics
105
organizing principles [3,4]. The main building elements within brain organization are functional anatomical networks called brain circuits. Such brain compartmentalization is congruent to anatomical structures featuring well-defined functions i.e. Default Mode Network, Dorsal Attention Network, etc [5]. Tooley et al. evidenced that brain circuits had already developed in childhood and that these circuits continue during neurodevelopment [5]. Paredes et al. proposed a mesoscale approach that revealed that brain circuits or modules have gene coexpression that leads to singular transcriptional dynamics [3]. Both works suggest that circuit transcriptional signatures are already shaped and will continue during neurodevelopment. Others recently works highlights the efficiency of the Poincare maps for observing changes across the time, specially for maintain all the important informative characteristics in single cell data, that permits a better understanding and visualization of the development and finally its potential use for any type of data with a hidden hierarchical structure [6]. In this work, we explore brain structure transcriptomes to illustrate that at any stage of neurodevelopment the brain holds its co-expression dynamics,employing the Poincare maps which allow us represent highly complex biological data in a easy way for reliable results visualization.
2 2.1
Methods Database
To scan expression dynamics throughout neurodevelopment, we retrieved gene expression data from the BrainSpan database [7]. The BrainSpan is a transcriptional atlas of human brain development across 26 structures from an embryonic stage (8 post-conception weeks) throughout adulthood (40 years). Next, we fetch the gene communities proposed by Paredes et al. [3] bottom-up brain model. Such gene communities are gene pools that co-express throughout the brain. By these transcriptional dynamics the brain builds all the brain circuits. Since the transcriptomes used in BrainSpan and gene community identification are different, we screened both transcript sets to keep those community genes and their gene expressions listed in the BrainSpan transcriptomes, the coincidence are shown in Table 1. Then we built the gene expression profiles for each gene community across all structures and stages. 2.2
Community Neurodevelopment
Spatial transcriptomes are multidimensional datasets showing gene expressions and their changes at different loci [8]. When also considering their temporal trend, their dimensions increase in number. To study and characterize the gene community transcriptional dynamics throughout neurodevelopment, we employed dimensionality reduction approaches. The most widespread reduction method is principal component analysis (PCA). PCA simplifies the data information into a characteristic space built
106
G. Guzm´ an et al.
by singular vectors. Further reduction approaches are mappings to topological invariant euclidean spaces such as T-distributed Stochastic Neighbor Embedding and Locally Linear Embedding. Yet, all these three methods have constraints to represent highly complex data such as spatiotemporal transcriptomes. The main restrictions are the omission of local topological relations and the simplification of complex hierarchies into euclidean relations [9]. Poincar´e maps are non-Euclidean manifolds that optimize spaces to map local and hierarchical topological relationships. Poincar´e maps proved to be reliable when representing highly complex biological data like single-cell. Such reduction method represented reliable biological trends as cell lineages hierarchies and gene expression time trajectories [6]. Hence, we implemented Poincar´e maps to model the individual expressions of gene communities. The goal of this method is to illustrate that gene co-expression temporal dynamics remain highly conserved during neurodevelopment.
3
Results and Discussion
3.1
Dataset
To trace gene expression dynamics in gene communities, we downloaded gene community information proposed in [3]. Then, we retrieved the gene communities from the BrainSpan database [7]. The gene numbers in the proposed communities and those retrieved from BrainSpan are in Table 1. Table 1. Number of genes retrieved from the BrainSpan database [7] that are present int the communities proposed by Paredes et al. [3] Donor 1
Donor 2
Community Genes in [3] Genes in BS
Community Genes in [3] Genes in BS
C1
1138
1106
C1
888
862
C2
1356
1317
C2
1127
1101
C3
1830
1787
C3
862
838
C4
1841
1786
C4
872
848
C5
753
727
C5
1873
1830
C6
941
908
C6
887
862
C7
1141
1107
To verify that the gene difference does not influence the study, we enriched the retrieved genes from BrainSpan and the gene difference set in the STRING database [10]. Figure 1 shows an example, where we observed that the difference set does not enrich any relationships in the STRING database, even some transcripts are missing. We suggest that the difference genes are peripheral genes within the communities, and therefore are not relevant to the study. Next, we built the individual expression matrices for each gene community at each structure and at all stages of neurodevelopment. We have 13 communities including six from donor 1 and seven from donor 2 derived from the Allen’s database.
Transcriptional Dynamics
107
Fig. 1. Co-expressed interaction STRING-enriched genes from D1 C1 community from [3]. In the upper left corner are shown the genes that were not retrieved from the BrainSpan dataset for this community. The network was built with any kind of direct interaction with a 0.01 p-value
3.2
Gene Community Expression Dynamics
To generate the Poincar´e maps for each gene community, we used the default parameters of the algorithm proposed by Klimovskaia et al, which are batch size = 52, gamma = 2, sigma = 1, epoch = 1000. We varied the neighbor number since we noticed that any representation with a higher gene number neighbourhood yielded high density maps which overlook gene expression trends. Thus, we fixed the parameter to 10 neighbors. Based on these parameters, we obtained a mean
108
G. Guzm´ an et al.
loss value below 0.28 for all the communities, indicating that the Poincar´e maps retain the hierarchical and local topological data relationships. A global trend in all communities is a noticeable expression difference between adulthood and embryonic development (See Figs. 2 and 3). This indicates that the communities’ transcriptional dynamics diverge between the two stages. Such behavior is congruent with neurodevelopment since at the embryonic stage the aim is to develop the brain circuits while in adulthood we only strive for their proper functioning [5].
Fig. 2. Poincar´e maps from Donor1 gene communities expression across the age development. The black dot is the central reference. Note that this projection clusters genes according to age.
Particular cases to highlight are those related to the D2 C6 community and the D1 C4 and D2 C5 communities. Paredes et al. suggested that the D2 C6 community is a gene ensemble that promotes dendrogenesis and its regulation, while the D1 C4 and D2 C5 communities play a role within axonogenesis. The dendritic spike regulation is high regulated for spikes generation and elongation during the childhood, nevertheless, become negatively regulated during the adulthood [11,12]. Concerning axonogenesis, there is evidence that during early stages of neurodevelopment, axonogenesis aims to develop brain neuronal
Transcriptional Dynamics
109
Fig. 3. Poincar´e maps from Donor2 gene communities expression across the age development. Most communities show projections very similar to Donor1.
layers that enable normal brain function until the brain reaches functional niches with mechanisms to control and promote brain plasticity [13,14]. These two cases reveal differentiated purposes that the transcriptional machinery aims at, implying differentiated transcriptional behaviors. This work exemplifies these qualitative trends among gene communities during neurodevelopment. Such observations call for further insight to determine the pathways and functions implicated during these transcriptional dynamics shifts.
110
4
G. Guzm´ an et al.
Conclusions
In the literature has been demonstrated that dendritic spike regulation, axogenesis and its regulation are important in neurodevelopment, furthermore, this development can conserve across the time some gene expression communities. Our results show that the basic neurodevelopmental circuits can be determined mesoconectomically, allowing a better understanding of their organization and regulation. Finally, future work should aim to implement further analysis where the complexity of exploration is increased by taking into account other gene elements.
References 1. Changeux, J.-P., Goulas, A., Hilgetag, C.C.: A connectomic hypothesis for the hominization of the brain. Cereb. Cortex 31(5), 2425–2449 (2020) 2. Bassett, D.S., Sporns, O.: Network neuroscience. Nat. Neurosci. 20(3), 353–364 (2017) 3. Paredes, O., Jhonatan, B.L., Covantes-Osuna, C., Ocegueda-Hern´ andez, V., RomoV´ azquez, R., Alejandro Morales, J.: A transcriptome community-and-module approach of the human mesoconnectome. Entropy 23(8), 1031 (2021) 4. Fornito, A., Arnatkeviˇci¯ ut˙e, A., Fulcher, B.D.: Bridging the gap between connectome and transcriptome. Trends Cogn. Sci. 23(1), 34–50 (2019) 5. Tooley, U.A., Bassett, D.S., Mackey, A.P.: Functional brain network community structure in childhood: unfinished territories and fuzzy boundaries. NeuroImage 247, 118843 (2022) 6. Klimovskaia, A., Lopez-Paz, D., Bottou, L., Nickel, M.: Poincar´e maps for analyzing complex hierarchies in single-cell data. Nat. Commun. 11(1), 1–9 (2020) 7. Home :: Brainspan: Atlas of the developing human brain (2022) 8. Moses, L., Pachter, L.: Museum of spatial transcriptomics. Nat. Methods 19(5), 534–546 (2022) 9. Kobak, D., Berens, P.: The art of using t-SNE for single-cell transcriptomics. Nat. Commun. 10(1), 1–4 (2019) 10. Szklarczyk, D., et al.: STRING v11: protein-protein association networks with increased coverage, supporting functional discovery in genome-wide experimental datasets. Nucl. Acids Res. 47(D1), D607–D613 (2018) 11. Kirch, C., Gollo, L.L.: Single-neuron dynamical effects of dendritic pruning implicated in aging and neurodegeneration: towards a measure of neuronal reserve. Sci. Rep. 11(1), 1–15 (2021) 12. Duan, H.: Age-related dendritic and spine changes in corticocortically projecting neurons in macaque monkeys. Cerebral Cortex 13(9), 950–961 (2003) 13. Finlay, B.L., Uchiyama, R.: The timing of brain maturation, early experience, and the human social niche. Evol. Nerv. Syst. 123–148 (2017) 14. Miyata, T., et al.: Migration, early axonogenesis, and reelin-dependent layerforming behavior of early/posterior-born purkinje cells in the developing mouse lateral cerebellum. Neural Dev. 5(1), 23 (2010)
CNNs for ISCI Stage Recognition on Video Sequences Gabriela Aguirre-Espericueta
and Gerardo Mendizabal-Ruiz(B)
Departamento de Biología Traslacional, Universidad de Guadalajara, 44430 Guadalajara, Jalisco, México [email protected]
Abstract. Intracytoplasmic sperm injection (ICSI) is one of the most commonly applied techniques for in vitro fertilization. This technique consists of the single selection of a spermatozoon followed by the injection of this sample into the oocyte’s plasma. The embryologists perform the ICSI by using their judgment to select the spermatozoon to inject. Additionally, they decide the best technique to penetrate the oocyte with the needle. Therefore, the success of an ICSI procedure can be affected by subjective decisions such as the characteristics of the sperm selected, the angle at which the needle pierces the oocyte, and the speed at which it is performed. The main objective of this project is to develop a computational tool that can automatically identify the different stages of the ICSI procedure. A tool like this will automate the activation of other artificial intelligence tools that can assist the embryologist while performing the ICSI (e.g., assistants to select the best sperm to inject and guides for the injection technique). Our results indicate the feasibility of employing a deep neural network architecture to determine the stage of the ICSI procedure from a video stream from a camera attached to a microscope. Keywords: ICSI · Neural network · Deep learning
1 Introduction 1.1 Development of ICSI Intracytoplasmic sperm injection (ICSI) is one of the most commonly applied techniques for in vitro fertilization (IVF). Palermo et al. made one of the first mentions in July 1992 [1]. While their team was performing a sub-zonal injection of an oocyte procedure, the oolemma was accidentally breached, causing the deliberation of spermatozoon directly into the ooplasm. This incident inspired the development of the ICSI procedure as we know it these days. This technique consists of the single selection of a spermatozoon followed by the injection of this sample into the oocyte’s plasma [2]. During ICSI, an embryologist selects what they determine to be the best sperm to inject into the oocyte directly. Sperm are selected subjectively by evaluating the morphology (shape) and progression (movement) of the sperms from a sample in a droplet. The selected sperm is then aspirated from the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 111–118, 2023. https://doi.org/10.1007/978-3-031-18256-3_11
112
G. Aguirre-Espericueta and G. Mendizabal-Ruiz
sperm sample into a microtool called an ICSI needle. Once the sperm is in the ICSI needle, the embryologist moves it to a media drop containing the oocyte to be fertilized. The egg to be injected is held in place by a holding pipette, which exerts a light suction on the oocyte, allowing the embryologist to place the oocyte in the perfect injection position. The embryologist then brings the ICSI needle with the sperm down to the other media drop and lines it up with the oocyte. The ICSI needle is pressed into the side of the egg below the polar body. The zona pellucida and oolemma are punctured. Then, a small part of the ooplasm is aspirated into the needle to mix it before the sperm is placed inside the oocyte. Embryologists typically perform several ICSI procedures in one session depending on the number of oocytes retrieved during that cycle. Initially, ICSI was presented as a promising solution or alternative for cases where the pregnancy was unsuccessful even after implementing other techniques. Some techniques mentioned in the original publication were partial zona dissection (PZD) and sub zonal insemination (SUZI). At first, the ICSI technique was recommended for assisted reproduction in cases where the principal factors for infertility were masculine (i.e., low spermatozoon quantity, morphologic abnormalities, or mobility). However, during the constant improvement and development of the ICSI techniques, it was found that it can also be used for cases where the oocytes had been cryopreserved, where there was low oocyte yield when there were undersaturated or dysmorphic oocytes, or as an alternative where another standard insemination technique was unsuccessful [1, 2]. ICSI can prevent certain illnesses caused by genetic defects or ploidy status because of the testing performed on the oocytes, such as preimplantation genetic diagnosis (PGD) and preimplantation genetic screening (PGS) [3]. The embryologists perform the ICSI by using their judgment to select the spermatozoon to inject and decide the best technique to penetrate the oocyte with the needle. The success of an ICSI can be affected by subjective decisions or human error, such as the angle at which the needle pierces the oocyte and the speed at which it is performed. For this reason, there is an increasing attempt to develop artificial intelligence (AI) assistants that can provide information to the embryologist during the ICSI procedure to improve the chances of success. The use of AI in the IVF lab is made possible since most microscopes already have digital cameras. Therefore, it is possible to develop software that can analyze the video feed from the cameras attached to the microscopes in real-time. A significant challenge in using these assistants is being able to activate them at the right time during the ICSI procedure. For example, an algorithm can automatically analyze and select the best sperm to inject [4]. Still, it only works when the microscope is on a droplet where sperms are in a solution that makes them slow so they can be manually immobilized. This method would need to be inactivated when the microscope stage is moved to the drop where an oocyte is waiting to be injected since it may interfere with the procedure. At the same time, it would be necessary to activate an assistant that can indicate the best technique to penetrate the oocyte to inject a sperm to fertilize it. Therefore, the main objective of this project is to develop a computational tool that can identify the different stages of the ICSI procedure.
CNNs for ISCI Stage Recognition on Video Sequences
113
1.2 Convolutional Neural Networks Convolutional neural networks (CNN) are a class of artificial neural network architecture designed to perform pattern analysis on two or more dimensional images [5]. The difference between traditional fully connected neural networks and CNN is that the former contains layers that perform convolutions. This is an operation that linearly combines the values of the pixels of the images based on a sliding matrix (i.e., kernel) that defines the contributions of each pixel to the convolution operation. Compared with other traditional computer-based approaches that employ machine learning, the main advantage of CNN is that the neural networks can estimate optimal kernel values to solve the problem being tackled automatically [6]. CNN architectures have been recently used in many applications related to image recognition and analysis in different fields, including IVF [7, 8]. One of the principal characteristics of neural networks is how they try to emulate the functioning of the human brain cortex. Therefore, they are highly used for the pattern and feature identification [9]. This is because using local receptive fields, tied weights, and special subsampling enables the ability to extract elementary features like edges or corners when combined with the convolutional layers, allowing the detection of higherorder features. CNNs are structured by layers roughly classified into three essential categories: one input layer, many hidden layers of processing (including convolutional blocks, pooling elements, and flattening), and an output layer, consecutively (Fig. 1). The convolutional layers oversee convoluting the whole image by implementing kernels, intermediate feature maps, and generating many feature maps. The pooling layers are in charge of reducing the dimension, such as the width and height of the input image for the next convolutional layer, not affecting the depth dimension. This is also known as subsampling or downsampling. In this subsampling process, mentioning the beneficial loss of information is essential. Moreover, finally, fully connected layers, where the conversion of a 2D feature map into a 1D feature vector offers the option of feeding it into a certain number of categories for classification or the consideration of the vector for further processing [10, 11].
Fig. 1. Example of the structure and functioning of a convolutional neural network.
114
G. Aguirre-Espericueta and G. Mendizabal-Ruiz
2 Methods The database employed consisted of 2,550 frames from 50 video sequences where ICSI was recorded at 15 frames per second at a resolution of 640 × 480 pixels. The data corresponded to different ICSI procedures collected from ten clinics from January 2021 to January 2022. The dataset is not publicly available. Each frame was organized into two classes according to the two main stages of the ICSI process: selecting a spermatozoon and injecting a spermatozoon (Fig. 2). Training and testing datasets were generated by separating the frames on four digital folders using an 80–20% ratio, in which each category contained two folders, one for each stage class. We assured not to include frames from the same videos in the different folders, meaning that the training folder did not contain frames from the videos in the testing folder. The manual classification of the dataset for building the ground truth for training and validation sets followed the following criteria: (i) for the code to predict the injection stage, there had to be at least two of the following on the frame: an oocyte, a needle, or a pipette, and (ii) for the code to predict the selection stage, there had to be at least four spermatozoa cells, and presence of the needle could be optional. We implemented a custom convolutional neural network architecture consisting of four convolution layers, one max-pooling layer, and a final layer of flattening. Table 1 list the network architecture along with the number of parameters for each layer. Table 1. Architecture and parameters of the proposed CNN. Layer (type)
Output shape
conv2d (Conv2D)
(None, 62, 62, 32)
max_pooling2d
(MaxPooling2D (None, 31, 31, 32))
conv2d_1 (Conv2D)
(None, 29, 29, 32)
Param # 896 0 9248
max_pooling2d_1
(MaxPooling (None, 14, 14, 32)2D)
conv2d_2 (Conv2D)
(None, 12, 12, 32)
max_pooling2d_2
(MaxPooling (None, 6, 6, 32)2D)
conv2d_3 (Conv2D)
(None, 4, 4, 32)
max_pooling2d_3
(MaxPooling (None, 2, 2, 32)2D)
0
flatten (Flatten)
(None, 128)
0
dense (Dense)
(None, 128)
16512
dense_1 (Dense)
(None, 1)
Total params: 45,281 Trainable params: 45,281 Non-trainable params: 0
0 9248 0 9248
129
CNNs for ISCI Stage Recognition on Video Sequences
115
We trained this network using the Adam optimizer employing the binary cross entropy as the loss function. The network was trained for 30 epochs using a batch size of 16 and eight validation steps.
Fig. 2. Example of frames from the dataset. Left: example of selection stage. Right: example of injection stage.
3 Results The proposed CNN was implemented using Keras and Tensorflow using an Intel Core i7 with 16 Gb of RAM and a GPU Nvidia 3070. Figure 3 depicts the training and validation set loss curves during training. The lowest training and validation loss are 0.027 and 0.021, respectively. In Fig. 4, we can observe that the most significant values for the accuracy in the training and validation datasets are 0.992 and 1, respectively. It is important to note that the cost function converged by the final epochs of the training and validation stage. This indicated that a successful training was reached.
Fig. 3. Loss graph during training.
116
G. Aguirre-Espericueta and G. Mendizabal-Ruiz
Fig. 4. Accuracy graph during training.
On the implementation of the network for the classification of the video sequences, a sign in yellow of the ICSI stage was shown on the superior left corner of the frame according to the prediction of the network. Figure 5 depicts examples of frames where the neural network succeeded in classifying the stage of the ICSI procedure as sperm selection despite the different situations regarding the presence or absence of the needle, sperms, or other tissues.
Fig. 5. Examples of successful results of selection.
Figure 6 depicts examples of frames where the neural network succeeded in classifying the stage of the ICSI procedure as the injection. Note that there are different magnifications and sizes of oocytes and that there may be instruments on the image, such as the needle and the holder, with different characteristics. It was noted that the performance of CNN was incorrect in some frames. For example, where the semicircular shape of a significant drop was on the frame at the same time
CNNs for ISCI Stage Recognition on Video Sequences
117
Fig. 6. Examples of successful results of injection.
as the needle (Fig. 7, left), the network predicted it as the stage of injection even when spermatozoa were present. Alternatively, in some cases on frames where none of the elements (needle, pipette, oocyte, or spermatozoa) were present, and none of the two stages established applied, the code predicted the selection stage (Fig. 7, right).
Fig. 7. Examples of wrong classification.
4 Discussion CNN is an excellent alternative for analyzing and understanding medical data, such as image recognition. This has been demonstrated in many recent works that employ them to analyze medical data. In our case, we have found that it is feasible to use a CNN architecture to classify video sequences taken under a microscope that can automatically indicate the stage of the ICSI procedure.
118
G. Aguirre-Espericueta and G. Mendizabal-Ruiz
However, one of the possible limitations of the application of CNN-based methods is the requirement of high computing power and memory capabilities to perform the inference. This is particularly true when real-time (at least 15fps) is desired. However, most of the equipment that can be found in an IVF lab (microscopes, micromanipulators, etc.) are high-priced. Therefore, it may be possible to convince the clinics to acquire a PC that at least counts with a GPU compatible with CNN models. In the next version of our network, we will divide the ICSI procedure into more stages, including the movement from one droplet to the other, the immobilization of sperm, and the three parts of the injection: oocyte holding, penetration, and insemination.
5 Conclusions We have presented an automatic method that can detect an ICSI Procedure’s stage using a microscope video feed. The results indicate the feasibility of using this methodology for real-time applications where other AI-based assistants need to be activated depending on the step of the ICSI.
References 1. Palermo, G.D., et al.: Intracytoplasmic sperm injection: state of the art in humans. Reproduction 154(6), F93–F110 (2017) 2. O’neill, C.L., Chow, S., Rosenwaks, Z., Palermo, G.D.: Development of ICSI. Reproduction 156(1), F51–F58 (2018) 3. Palermo, G., Joris, H., Devroey, P., Van Steirteghem, A.C.: Pregnancies after intracytoplasmic injection of single spermatozoon into an oocyte. The Lancet 340(8810), 17–18 (1992) 4. Mendizabal-Ruiz, G., et al.: Computer software (SiD) assisted real-time single sperm selection correlates with fertilization and blastocyst formation. Reproductive BioMedicine Online (2022) 5. Valueva, M.V., Nagornov, N.N., Lyakhov, P.A., Valuev, G.V., Chervyakov, N.I.: Application of the residue number system to reduce hardware costs of the convolutional neural network implementation. Math. Comput. Simul. 177, 232–243 (2020) 6. Géron, A.: Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow: Concepts, tools, and techniques to build intelligent systems. O’Reilly Media Inc. (2019) 7. Fernandez, E.I., et al.: Artificial intelligence in the IVF laboratory: overview through the application of different types of algorithms for the classification of reproductive data. J. Assist. Reprod. Genet. 37(10), 2359–2376 (2020). https://doi.org/10.1007/s10815-020-018 81-9 8. Louis, C.M., Erwin, A., Handayani, N., Polim, A.A., Boediono, A., Sini, I.: Review of computer vision application in in vitro fertilization: the application of deep learning-based computer vision technology in the world of IVF. J. Assist. Reprod. Genet. 38(7), 1627–1639 (2021). https://doi.org/10.1007/s10815-021-02123-2 9. Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural network. In: 2017 international conference on engineering and technology (ICET), pp. 1–6. IEEE (August 2017) 10. Voulodimos, A., Doulamis, N., Doulamis, A., Protopapadakis, E.: Deep learning for computer vision: A brief review. Computational intelligence and neuroscience 2018 (2018) 11. Chauhan, R., Ghanshala, K.K., Joshi, R.C.: Convolutional neural network (CNN) for image detection and recognition. In: 2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC), pp. 278–282. IEEE (December 2018)
Stacked Spatial and Temporal Deep Learning Methods for Identification of Parkinson’s Disease Using Gait Signals Brenda Guadalupe Mu˜ noz-Mata1(B) , Guadalupe Dorantes-M´endez1 , and Omar Pi˜ na-Ram´ırez2 1
2
Facultad de Ciencias, Universidad Aut´ onoma de San Luis Potos´ı, 78295 San Luis Potos´ı, San Luis Potos´ı, M´exico [email protected] Departamento de Bioinform´ atica y An´ alisis Estad´ısticos, Instituto Nacional de Perinatolog´ıa “Isidro Espinosa de los Reyes”, Miguel Hidalgo, 11000 Ciudad de M´exico, M´exico
Abstract. Parkinson’s disease (PD) is a progressive condition that affects dopaminergic neurons, causing motor alterations. Motor disturbances, such as gait impairment, can be used to assess the disease. Unfortunately, gait disturbances, such as decreased walking speed and step variability, can also occur due to aging, affecting the identification of abnormal PD gait. Therefore, developing an adequate tool to evaluate PD patients’ gait is essential. This paper proposes a deep learning algorithm to differentiate between PD gaits and normal walking using vertical ground reaction force (VGRF) signals. CLDNN is a single framework composed of a convolutional neural network, a long-short term memory network, and a deep neural network. To train and validate a CLDNN classifier gait cycles were obtained from VGRF signals. The VGRF signals were from a public database with recordings from 93 PD patients and 73 healthy adult controls. The CLDNN performance was evaluated by five-fold cross-validation. The combined spatial and temporal methods in CLDNN enabled the effective identification of PD gait with less complex architecture. The best weighted accuracy was 98.28 ± 0.38. Thus, our model is compact and efficient for future embedded or portable implementations.
Keywords: Deep learning ground reaction force
1
· Gait cycle · Movement disorders · Vertical
Introduction
Parkinson’s disease (PD) is a neurodegenerative disorder characterized by the presence of motor and non-motor alterations [4]. The motor disturbances result from the progressive loss of dopaminergic neurons in the substantia nigra [8]. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 119–126, 2023. https://doi.org/10.1007/978-3-031-18256-3_12
120
B. G. Mu˜ noz-Mata et al.
The decrease of dopaminergic neurons is reflected in motor disturbances such as tremor, stiffness, bradykinesia, and postural instability [4,8]. PD alterations are progressive and affect the quality of life of people who suffer from it [4]. PD is one of the most frequently occurring neurological diseases in society. The incidence of PD varies in different reports. However, it is generally accepted that PD affects 1–2 per 1000 of the population [9], more than 10 million people worldwide [10]. PD prevalence increase with age, being aging the most significant risk factor for developing idiopathic PD [6], affecting 1% of the population above 60 years [9]. Consequently, the increase in life expectancy, thus the increase in the elderly population, indicates that more older people could be affected by PD in the coming years. Although many learning algorithms have been developed in the last decade, only a few studies can potentially be helpful for clinical use. Thus, PD remains a clinical diagnosis. The current evaluation of PD uses the descriptions of medical experts, patients, or caregivers. To diagnose the disease, clinicians evaluate motor alterations, including gait disturbances [4]. Nonetheless, gait abnormalities of PD, such as decreased walking speed, shorter step length, increased step variability, and a shuffling gait [11], could also be developed during normal aging. The similarity between walking in the elderly and PD affects the early diagnosis of Parkinson’s since its clinical analysis is subjective. Therefore, an objective classification system for PD assessment using gait information is of general interest to clinicians and PD patients, as it can help increase the efficiency of diagnosis. Considering a model implementation in clinical diagnosis, the simplest classifier that provides adequate accuracy is preferable over complex algorithms that may provide slightly higher accuracy. Some classifiers consider the differences in walking patterns to classify the disease because gait patterns can assess PD without requiring specialized tests. In addition, gait can be evaluated using vertical ground reaction force (VGRF) signals easily acquired by placing inexpensive, non-invasive sensors on the sole [12]. In the literature, some studies have focused on PD identification by deep learning methods using VGRF sequences without prior manual feature extraction since the sequences provide spatial and temporal information about the gait. In the work of Hoang et al. [3] a maximum percentage accuracy of 88.7% was reported using two-dimensional and one-dimensional stacked convolutional neural networks (CNNs). In another study, a percentage accuracy of 98.61% was presented considering segments of the VGRF signals as inputs in a two-channel network composed of a CNN and long short-term memory (LSTM) network [13]. Also, one-dimensional CNNs have been used to model the VGRF segments, reporting an accuracy of 98.7% [1]. On the other hand, using gait cycles obtained from VGRF signals, a dual model based on deep learning, where each foot was modeled separately by a CNN followed by an attention-enhanced bidirectionalLSTM network, got an accuracy of 99.07% [11]. The classification systems mentioned above consider the spatio-temporal information provided by the VGRF sequences and use algorithms to analyze spatial (CNN) and temporal (LSTM)
Stacked Spatial and Temporal Deep Learning Methods
121
features. However, they implement classifiers per foot [11], per the number of sensors [1], or network [13], increasing the complexity of the models. The main objective of this work is to develop a non-complex accurate classification algorithm that, in the future, can be used as a supportive tool for medical diagnosis to identify PD gait from normal gait. Therefore, VGRF signals from both feet were used in a network combining CNN, LSTM, and a fully connected network (FCN). The purpose of the network, known as CLDNN, was to extract spatial and temporal information from gait cycles to identify walking patterns. To the best of our knowledge, the proposed architecture based on CLDNN has never been proposed to identify the gait of PD patients. For medical diagnosis, it is preferable to have compact and efficient classification models for their application. In this sense, the CLDNN design has fewer parameters to compute for future implementation. Besides, CLDNN provides acceptable percentages that compete with state-of-the-art models.
2 2.1
Materials and Methods Databases Description
A public database [2] was used. The database includes the VGRF signals of 93 PD patients (mean age 66.3 years) and the VGRF signals of 73 healthy control (HC) participants (mean age 66.3 years). The VGRF signals were acquired by placing eight sensors on the soles of each participant. Subsequently, each participant was asked to walk on level ground for two minutes at their own pace. In the database, 18 time series with a sampling rate of 100 Hz per participant are available: 16 VGRF signals from the eight sensors of each foot and two records that are the sum of the sensors of each foot (total force). 2.2
Signal Processing: Gait Cycles
The gait cycle is the sequence of movements that begins when the reference foot hits the ground and ends when the foot loses contact [5]. In the VGRF signal, particularly in the total force signal, a gait cycle begins when the force jumps to non-zero values corresponding to heel strike and ends when the force values return to zero representing toe-off [11]. Figure 1 shows examples of gait cycles from a PD patient and a healthy control participant. 2.3
CLDNN
A CLDNN architecture involves feeding the input data to a CNN. Subsequently, the CNN output becomes the input of an LSTM network, and then the output of the recurrent network is used to classify the data with a deep neural network (DNN), which can be an FCN. The individual modeling capabilities of the CNN, LSTM, and DNN networks are limited. Thus, CLDNN aims to enhance the classification performance by combining deep networks in a unified framework [7]. In
122
B. G. Mu˜ noz-Mata et al.
Fig. 1. Gait cycles of VGRF signals of a PD patient and a healthy control adult from the left and right foot. S1 to S8 are the signals of each of the eight sensors under the left foot. On the other side, S9 to S16 are the signals of each of the eight sensors under the right foot. TFL and TFR correspond to the total force of the left and right foot, respectively.
CLDNN, the primary purpose of CNN is to reduce frequency changes (spectral changes) in the input data to facilitate the LSTM learning procedure. Meanwhile, the LSTM network is good at time modeling reducing time variations [7]. Therefore, the combination of CNN and LSTM helps to interpret the spatial and temporal information of the data. Thereafter, that information, which corresponds to the data features, is classified using a DNN. 2.4
Classification
VGRF signals were normalized and partitioned into gait cycles. Due to fluctuating gait speed and stride length, the length of gait cycles is not always the same. Since the network requires vectors of the same size for training and validation, zero-padding was used to have sequences of the same length. A length of 150 samples (1.5 s) was considered, based on the maximum time of the gait cycle observed in previous works [11]. Subsequently, the input data vectors were formed by concatenating the 1.5-s gait cycle segments from the sensors of each foot without the total forces. As a result, the dimensions of the input vectors were 1 × 150 × 16. Each input vector was tagged with its corresponding class label. The proposed CLDNN was trained to classify vectors of 1 × 150 × 16. Figure 2 depicts the proposed architecture. In the architecture CNN is formed by two onedimensional convolutional layers of size 7. The first and second convolutional layers have 32 and 64 filters respectively, and a max-pooling layer of size 2 follows every convolutional layer. The LSTM network is formed by one layer of 100 units. In the FCN, the number of nodes in the input layer is 100, while
Stacked Spatial and Temporal Deep Learning Methods
123
the output layer has one node with a sigmoid activation function for binary classification. The models were trained for 600 epochs using a Nadam optimizer, a batch size of 32, and an initial learning rate of 0.0001. The loss function used was cross-entropy. In addition, the regularization technique of Dropout (50%) was applied after each pooling layer. To evaluate the algorithm, it was used a simple validation (80% train, 20% test) and a five-fold cross-validation procedure. The metrics were calculated from the number of true positives (T P ), false positives (F P ), true negatives (T N ), and false negatives (F N ) of the confusion matrix. The evaluation metrics reported are accuracy = [(T P + T N ) / (T P + T N + F P + F N )] × 100, sensitivity or recall = [(T P/ (T P + F N )] × 100, specif icity = [T N/ (T N + F P )]×100, precision = [T P/ (T P + F P )]×100, and F 1−score = (2 × P recision × Recall)/ (P recision + Recall) [1]. AUC values were also considered. The control group is the negative (N ) group, and the PD group is the positive (P ). Therefore, sensitivity corresponds to correctly classified PD patients, and specificity corresponds to the proportion of correctly classified HC participants.
Fig. 2. Proposed CLDNN architecture for binary classification.
3
Results and Discussion
This section presents the evaluation performance of the CLDNN models for PD gait identification. The learning curves and the confusion matrix in Fig. 3 illustrate the model performance with simple validation. Visually there was no significant overfitting, i.e., the accuracy value of the validation data increased in the same way that the accuracy value of the training data. At the same time, the loss curves decreased similarly for both data sets. Hence, this indicates adequate learning during the training stage. Both classes in the confusion matrix had a percentage accuracy greater than 98% on the test data. Although more healthy control participants were misclassified, the predictive capability of the model was high, with an overall accuracy of 98.55%. The classification report of the predictions in each class is represented
124
B. G. Mu˜ noz-Mata et al.
in Table 1. In this table, we observe that the F1-score was similar between classes (0.981 and 0.988). Therefore, the model seems consistent in its predictions, even if the data set is unbalanced. Finally, five-fold cross-validation was performed to evaluate the proposed CLDNN architecture and guarantee that the results are independent of the training and test sets. The percentage of weighted accuracy obtained was 98.28 ± 0.38%. Almost all evaluation metrics exceed 98% in both types of validation (see Table 2). The simple validation and cross-validation results were consistent with resembling values. The lowest value of the evaluation metrics is related to the identification of normal gait since the percentage of specificity obtained was 97.79 ± 0.33%. However, there was no big difference between the percentages of specificity and sensitivity. Besides, the AUC value was 0.98, indicating a balanced identification of PD gait and normal walking.
Fig. 3. Learning curves and confusion matrix of the CLDNN model trained and evaluated with the gait cycle segments of 1.5 s. Table 1. Gait cycle prediction results on test data with simple validation. Class
%Precision %Recall F1-score Samples
HC
98.02
98.24
0.981
1766
PD
98.88
98.74
0.988
2779
Weighted averaged 98.55
98.55
0.985
Table 3 presents a comparative summary of the results of different studies using the same database. The results demonstrated that using VGRF signals, accurate classification is possible by combining temporal and spatial features. In the literature, the model with the lowest performance only uses spatial features [3], while in the works of Xia et al. [11] and Maachi et al. [1] the combination of CNN and LSTM (spatial and temporal features) results in better models, agreeing with our results. Our work obtained a weighted accuracy of 98.28 ± 0.38% with five-fold crossvalidation. Our model provided accurate results comparable with other studies in the literature. Although, the performance of the proposed CLDNN was slightly lower than other models presented in the current state of the art. The 1% difference is offset by a less complex architecture and fewer training epochs.
Stacked Spatial and Temporal Deep Learning Methods
125
Table 2. Simple validation and cross-validation (CV) evaluation metrics that were obtained from the models. Validation %Accuracy
%Precision
%Recall (Sensitivity)
F1-score
%Specificity AUC
Simple
98.55
98.55
98.74
0.985
98.24
CV
98.28 ± 0.38 98.28 ± 0.38 98.58 ± 0.50
0.983
0.982 ± 0.38 97.79 ± 0.33 0.981 ± 0.00
Table 3. Results of models for binary classification trained and evaluated using the public database of PD gait signals. Method
Input sequence Size %Accuracy
Dual 2D-CNN+Bi-LSTM+Attention layer [11] GC
1.5 s 99.07
Stacked CNNs: 2D-CNN+1D-CNN [3]
VGRF
2.5 s 88.7
Two channels: LSTM and CNN [13]
VGRF
1s
98.61
VGRF
1s
98.7
GC
1.5 s 98.28
18 parallel 1D-CNN [1] ∗ ∗
CLDNN: 1D-CNN+LSTM+FCN Proposed architecture, GC = gait cycle
Deep learning can be enhanced in different ways to make it better for realworld applications. The proposed CLDNN is of considerable importance due to its straightforward design. One model for each VGRF sensor [1], two-channel models [13], and a model per foot [11] imply higher model complexity than our proposed approach. One benefit of having a less complex network involves low computational cost. Another advantage is that the proposed CLDNN could be implemented in an embedded system (portable) in diagnostic-evaluation equipment.
4
Conclusions
Abnormal gait dynamics can be an indicator of pathological alterations. By analyzing VGRF signals, discriminatory gait features can be derived to evaluate gait-related diseases such as PD. This paper proposed a neural network for PD gait identification using gait cycles. Our deep learning classifier focused on spatial and temporal feature extraction and classification. With less complex architecture, CLDNN provided state-of-the-art performance. Therefore, it can be implemented eventually in a portable system. Additionally, further tuning can be applied to get better models. Also, in future work, a classification of PD severity can be carried out using the CLDNN architecture proposed with gait cycles.
126
B. G. Mu˜ noz-Mata et al.
References 1. El Maachi, I., Bilodeau, G.A., Bouachir, W.: Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait. Expert Syst. Appl. 143, 113,075 (2020). https://doi.org/10.1016/j.eswa.2019.113075 2. Goldberger, A.L., et al.: Physiobank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 101(23), e215–e220 (2000). https://doi.org/10.1161/01.CIR.101.23.e215 3. Hoang, N.S., Cai, Y., Lee, C.W., Yang, Y.O., Chui, C.K., Chua, M.C.H.: Gait classification for Parkinson’s disease using stacked 2D and 1D convolutional neural network. In: 2019 International Conference on Advanced Technologies for Communications (ATC), pp. 44–49. IEEE (2019). 10.1109/ATC.2019.8924567 4. Jankovic, J.: Parkinson’s disease: clinical features and diagnosis. J. Neurol. Neurosurg. Psychiatry 79(4), 368–376 (2008). https://doi.org/10.1136/jnnp.2007.131045 5. Prabhu, P., Pradhan, N.: Recurrence quantification analysis of human gait in neurological movement disorders. IJERT 5(03), 447–452 (2016) 6. Reeve, A., Simcox, E., Turnbull, D.: Ageing and Parkinson’s disease: why is advancing age the biggest risk factor? Ageing Res. Rev. 14, 19–30 (2014). https://doi. org/10.1016/j.arr.2014.01.004 7. Sainath, T.N., Vinyals, O., Senior, A., Sak, H.: Convolutional, long short-term memory, fully connected deep neural networks. In: 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4580–4584. IEEE (2015). https://doi.org/10.1109/ICASSP.2015.7178838 8. Standaert, D.G., Saint-Hilaire, M.H., Thomas, C.A.: Parkinson’s Disease Handbook. American Parkinson’s Disease Association, Staten Island (2019) 9. Tysnes, O.-B., Storstein, A.: Epidemiology of Parkinson’s disease. J. Neural Transm. 124(8), 901–905 (2017). https://doi.org/10.1007/s00702-017-1686-y 10. Veeraragavan, S., Gopalai, A.A., Gouwanda, D., Ahmad, S.A.: Parkinson’s disease diagnosis and severity assessment using ground reaction forces and neural networks. Front. Physiol. 11, 587,057 (2020). https://doi.org/10.3389/fphys.2020.587057 11. Xia, Y., Yao, Z., Ye, Q., Cheng, N.: A dual-modal attention-enhanced deep learning network for quantification of Parkinson’s disease characteristics. IEEE Trans. Neural Syst. Rehabil. Eng. 28(1), 42–51 (2019). https://doi.org/10.1109/TNSRE. 2019.2946194 12. Zeng, W., Yuan, C., Wang, Q., Liu, F., Wang, Y.: Classification of gait patterns between patients with Parkinson’s disease and healthy controls using phase space reconstruction (PSR), empirical mode decomposition (EMD) and neural networks. Neural Netw. 111, 64–76 (2019). https://doi.org/10.1016/j.neunet.2018.12.012 13. Zhao, A., Qi, L., Li, J., Dong, J., Yu, H.: A hybrid spatio-temporal model for detection and severity rating of Parkinson’s disease from gait data. Neurocomputing 315, 1–8 (2018). https://doi.org/10.1016/j.neucom.2018.03.032
Diversity of Genotyping Chlamydia Trachomatis Serovars in Urogenital Samples from Mexican Patients: A Molecular and Bioinformatic Characterization Fabiola Hernández-Rosas1 , Socorro Mariana García-González2 , Shumeyker Susmith Franco-González2 , Ana Paola Salgado-Álvarez1 , and Mercedes Piedad de León-Bautista2,3(B) 1 Biomedical Engineering Faculty, Anáhuac University, El Marqués, 76246 Querétaro, Mexico 2 Escuela de Medicina, Universidad Vasco de Quiroga, 58090 Morelia, Michoacán, Mexico
[email protected] 3 Translational Medicine, Vanguard and Technology Transfer Sector, Human Health
Department, Central ADN Laboratories, 58341 Morelia, Michoacán, Mexico
Abstract. Chlamydia trachomatis (CT) is the most frequent bacterial sexually transmitted infection (STI) in the world. Therefore, the identification of serovars is essential for epidemiological surveillance and the development of prevention methods in our population. We aimed to demonstrate the diversity and frequency of serovars of CT in the urogenital tract in Mexican patients. We carried out an observational, prospective and cross-sectional study through 40 samples positive for CT. For the serovars identification, PCR multiplex, PCR of the ompA gene, automated sequencing, multiple sequence alignment, and phylogenetic analysis were used. Moreover, the variables like serovars, sex, aged groups, anatomical sites and, concomitant pathogens, were statistically analyzed. We reported that the prevalence of the most common serovars in our Mexican population were F (38.1%), E (27.3%) and, D (18.2%). Serovar F was the most prevalent in men, meanwhile, serovar D was in women and, E was equally prevalent in both genders. The most common concomitant pathogens were Ureaplasma spp. (22.7%), Mycoplasma spp. (13.6%) and, Haemophilus spp. (9.1%). In summary, epidemiological surveillance remarks the necessity of detecting serovars CT to elucidate the molecular implication and reduce sexual and reproductive complications. Keywords: Chlamydia trachomatis · Serovar · PCR · ompA gene · Automated sequencing
1 Introduction Chlamydia is the most prevalent obligate intracellular pathogen worldwide, with different implications and clinical manifestations [1]. In particular, Chlamydia trachomatis (CT) is a gram-negative and obligate intracellular bacterium and lacks peptidoglycan in its cell wall. It has an ovoid morphology of 0.3–1 µm in diameter, depending on the stage © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 127–135, 2023. https://doi.org/10.1007/978-3-031-18256-3_13
128
F. Hernández-Rosas et al.
in the replication cycle. Its envelope contains an outer membrane lipopolysaccharide and proteins [2]. Currently, 19 serovars of CT were described: A, B, Ba, C, D, Da, E, F, G, Ga, H, I, Ia, J, K, L1, L2, L2a and L3, classified depending on its antigenic variation and the adhesion of its major outer membrane protein (MOMP). The MOMP protein is coded by the omp1 gene [3]. CT is classified into different serovars depending on its antigenic variation and the adhesion of its MOMP protein, which affects different human tissues [4]. Serovars A, B, Ba, and C produce trachoma, a severe eye disease that can cause blindness, described as endemic in Africa and Asia. Serovars D and K produce genital tract infections and neonatal infections. Nevertheless, 75% of the cases are asymptomatic. Finally, serovars L1 and L3 are associated with genital ulcer. The clinical manifestations are swollen nodes in the inguinal area, edema, and redness [5–7]. It has been reported that the identification of serovars has clinical implications, furthermore, each population has different serovars. In this context, a new CT variant was identified in Mexico [8]. Contrasting and comparing this information, the identification of serovars is essential for epidemiological surveillance and the development of prevention methods in our population. Thus, the objective of our work was to demonstrate the diversity and frequency of serovars of CT in the urogenital tract in Mexican patients.
2 Methodology 2.1 Ethics Statement Our study was approved by the Ethics Committee of the Human Health from Central ADN Laboratories (approval no. 017/21). The samples were collected with informed consent from all patients. This study was carried out according to the ethical principles of the Declaration of Helsinki and Mexico’s General Health Law. 2.2 Research Population A cross-sectional, prospective, and analytical study was performed including patients of both sexes and ages between 18 and 60 years, who voluntarily donated their DNA samples from the cervix, urethra, and glans to the Human Health Department, Central ADN Laboratories, in September 2020 to April 2021. Also, the clinical data were collected from all patients in our validated questionnaire. 2.3 Samples and DNA Collection Clinical specimens from the urogenital tract were obtained from all participants. All samples were stored in a UTM medium (Copan Inc., USA) and then analyzed at the Central ADN Laboratories of Michoacán, Mexico. DNA extraction was performed using Instagene Matrix (Bio-Rad, USA) as described by Hernández-Rosas et al. [9].
Diversity of Genotyping Chlamydia Trachomatis Serovars in Urogenital Samples
129
2.4 Sampling The sampling was intentional. It was carried out from samples of patients with a positive result for CT and also other concomitant pathogens associated with Sexually Transmitted Infections. 2,100 STI samples were analyzed, of which 40 were CT-positive. The quality and purity of DNA were evaluated by spectrophotometry in a nanodrop spectrophotometer (Thermo Fisher Scientific Inc., MA USA) and through electrophoresis on a 2.0% agarose gel dyed with Red Safe (iNtRON Biotechnology, Inc., Korea). The bands were visualized in a system for DNA analysis and gel documentation (ENDURO™ GDS, LabNet international, Edison, NJ, USA). Only 26 samples met the quality and purity criteria for sequencing analysis. 2.5 Multiplex PCR Assay for the Diagnosis of STI 2,100 urogenital samples from patients of both sexes were analyzed by the multiplex PCR assay to detect Mycoplasma spp., Candida spp., Neisseria gonorrhoeae, Trichomonas vaginalis, Chlamydia trachomatis, and herpes virus 1 and 2, previously validated by our research group [9]. Only 40 samples were positive for CT. These ultimate samples were evaluated for the quality and integrity of DNA by a nanodrop spectrophotometer (Thermo Fisher Scientific Inc., MA USA). 2.6 OmpA-Targeted PCR PCR process consisted of the separation of the sample, which was placed in a 200 µl microtube, The multiplex PCR reactions were carried out as it was described previously Hernández-Rosas et al. [9]. To confirm that the amplicon corresponded to the expected size, a 100 base pair (bp) ladder was used on the gel. Oligonucleotide synthesis was performed by Integrated DNA Technologies, Inc., (San Diego, CA, USA); the catalog numbers are PI, 226910440; OMP2, 266910443; RVS, 266910443. Primer sequences for the amplification of the ompA gene are shown in Table 1. Table 1. Oligonucleotides and sequences for ompA gene amplification by PCR. Sequence 5 -3
Gene-Region
Primer name
Ref
OMPA
P1
ATGAAACTCTTGAAATCG
[10]
OMPA
OMP2
CTCAACTGTGACTGCGTATTT
[10]
OMP
MOMP 87
TGAACCAAGCCTTATGATCGACGACGGA
[11]
OMPA
RVS
TCTTCGAYTTTAGGTTTAGATTGA
[11]
130
F. Hernández-Rosas et al.
2.7 Identification of CT Serovars To identify the CT serovar, the amplification of the ompA sequence was developed with oligonucleotides, MOMP 87 + RVS, and P1 + OMP2. The amplification with the primers P1 + OMP2 required a higher level of purity, unlike the pair of oligonucleotides MOMP 87 + RVS. Therefore, we used MOMP 87 + RVS for sequencing by duplicates. Subsequently, the amplicons were automated sequencing at the Macrogen Laboratory (Macrogen Inc., Seoul, South Korea). The nBLAST program from NCBI was used to analyze the CT sequences serovars and confront them in GenBank [12]. 2.8 Statistical and Bioinformatic Analysis The data analysis was performed using Statistica 7 program (StatSoft, Inc., USA). The qualitative variables were described by frequency and percentages. The quantitative continuous with normal distribution variables were expressed as means ± SD. Chi2 test was used to assess statistical relationships for qualitative variables. P value 0 kJ/mol). The binding is driven by electrostatic interactions and, in minor, weaker non-electrostatic interactions (for example, hydrophobic contacts) mainly for megalin-kanamycin (Table 1). By contrast, flutamide established weak contacts (non-electrostatic interactions) in the binding with the megalin receptor. The binding is determined between charges groups of megalin and ototoxic, which implies that through the Coulombic component, an electrostatic force is established.
150
G. D. Hernández Cornejo et al.
(A)
(B)
Fig. 4. Complex that presented the most favorable binding energy of flutamide with megalin receptor: (A) Interaction map in 3D determined by PyMOL (B) Interaction diagram in 2D determined by LigPlot+. Table 1. Energy components to the Gb for megalin-kanamycin and megalin-flutamide systems at three different times during the simulation trajectory System
Gsolv (kJ/mol)
GCoul (kJ/mol)
Gnon-elec (kJ/mol)
Gb * (kJ/mol)
Megalin-kanamycin Complex1
42
−1302
−12
−1271
Complex2
41
−1268
−10
−1237
Complex3
44
−1141
−10
−1107
Mean
42(2)
−1237(85)
−11(1)
−1205(87)
Megalin-flutamide Complex1
66
271
−7.6
330
Complex2
42
194
−7.6
228
Complex3
49
191
−7.5
233
Mean
52(13)
219(45)
−7.6(0.1)
264(57)
The parentheses numbers indicate the complex’s standard deviation was more favorable at three different times during the simulation trajectory. *Eq. (1).
Chun et al. [9] recently reported that the otoprotective effects of flutamide on kanamycin-megalin induced hearing loss in rats; they carried out in vivo experiments inducing hearing loss in the rat, and by supplying flutamide, they observed through ABR
A Comparative Study on the Interaction of an Ototoxic and an Otoprotective
151
recordings that rat preserve hearing. However, until now, no computational studies have described the mechanism of interaction of kanamycin or flutamide with the megalin receptor. In this work, our models indicate that the interactions leading to the binding mainly of ototoxic with megalin are electrostatic. The present study also demonstrates that docking assays, molecular dynamics, and binding energy determinations provide a qualitative understanding of the driving forces responsible for binding ototoxic or otoprotective to megalin. Our computational model suggests flutamide does not promote endocytosis like aminoglycoside antibiotics; therefore, it does not contribute to hearing loss. In the future, we will complement our molecular recognition model with ABR simulations to monitor hearing loss due to ototoxic.
4 Conclusions Although aminoglycoside antibiotics such as kanamycin have been very effective in treating lung diseases, they are also ototoxic. They damage the hair cells in the cochlea, and megalin is known to be the primary transporter of kanamycin within the cochlea. Flutamide prevents kanamycin-induced hearing loss in animal models; however, the molecular mechanisms that produce this otoprotective are unknown. In this work, our interest was focused on the binding of megalin-kanamycin and megalin-flutamide involvement in ototoxicity and otoprotective, respectively. We demonstrate that docking assays, molecular dynamics, and binding energy determinations provide a qualitative interpretation of the driving forces responsible for binding ototoxic or otoprotective to megalin. Kanamycin favors endocytosis through the megalin receptor by binding with high affinity, inducing hearing loss, and flutamide does not favor this process. In summary, we characterize the molecular recognition between megalin and ototoxic (in our case, kanamycin). We observe a high affinity which could imply the activation of signaling cascades (inflammation factors) toward hearing loss.
References 1. Francis, S.P., Cunningham, L.L.: Non-autonomous cellular responses to ototoxic druginduced stress and death. Front. Cell. Neurosci. 11(August), 1–12 (2017). https://doi.org/ 10.3389/fncel.2017.00252 2. Ganesan, P., Schmiedge, J., Manchaiah, V., Swapna, S., Dhandayutham, S., Kothandaraman, P.P.: Ototoxicity: a challenge in diagnosis and treatment. J. Audiol. Otol. 22(2), 59–68 (2018). https://doi.org/10.7874/jao.2017.00360 3. Ciorba, A., et al.: Don’t forget ototoxicity during the SARS-CoV-2 (Covid-19) pandemic! Int. J. Immunopathol. Pharmacol. 34, 2 (2020). https://doi.org/10.1177/2058738420941754 4. Coffin, A.B., Boney, R., Hill, J., Tian, C., Steyger, P.S.: Detecting novel ototoxins and potentiation of ototoxicity by disease settings. Front. Neurol. 12(August), 1–14 (2021). https://doi. org/10.3389/fneur.2021.725566 5. Liu, K., et al.: Cochlear inner hair cell ribbon synapse is the primary target of ototoxic aminoglycoside stimuli. Mol. Neurobiol. 48(3), 647–654 (2013). https://doi.org/10.1007/s12 035-013-8454-2 6. Liu, K., et al.: Spontaneous and partial repair of ribbon synapse in cochlear inner hair cells after ototoxic withdrawal. Mol. Neurobiol. 52(3), 1680–1689 (2014). https://doi.org/10.1007/ s12035-014-8951-y
152
G. D. Hernández Cornejo et al.
7. Bas, E., et al.: Efficacy of three drugs for protecting against gentamicin-induced hair cell and hearing losses. Br. J. Pharmacol. 166(6), 1888–1904 (2012). https://doi.org/10.1111/j.14765381.2012.01890.x 8. König, O., et al.: Estrogen and the inner ear: megalin knockout mice suffer progressive hearing loss. FASEB J. 22(2), 410–417 (2008). https://doi.org/10.1096/fj.07-9171com 9. Chun, K.J., Lee, C.H., Kim, K.W., Lee, S.M., Kim, S.Y.: Effects of androgen receptor inhibition on kanamycin-induced hearing loss in rats. Int. J. Mol. Sci. 22(10), 5307 (2021). https:// doi.org/10.3390/ijms22105307 10. Saito, A., Pietromonaco, S., Loo, A.K.C., Farquhar, M.G.: Complete cloning and sequencing of rat gp330/megalin, a distinctive member of the low density lipoprotein receptor gene family. Proc. Natl. Acad. Sci. U.S.A. 91(21), 9725–9729 (1994). https://doi.org/10.1073/pnas.91.21. 9725 11. Dagil, R., O’Shea, C., Nykjaer, A., Bonvin, A.M.J.J., Kragelund, B.B.: Gentamicin binds to the megalin receptor as a competitive inhibitor using the common ligand binding motif of complement type repeats insight from the NMR structure of the WTH complement type repeat domain alone and in complex with gentamicin. J. Biol. Chem. 288(6), 4424–4435 (2013). https://doi.org/10.1074/jbc.M112.434159 12. Parasuraman, S.: Protein data bank. J. Pharmacol. Pharmacother. 3(4), 351–352 (2012). https://doi.org/10.4103/0976-500X.103704 13. Fuentes-Martínez, Y., Godoy-Alcántar, C., Medrano, F., Dikiy, A., Yatsimirsky, A.K.: Protonation of kanamycin A: detailing of thermodynamics and protonation sites assignment. Bioorg. Chem. 38(4), 173–180 (2010). https://doi.org/10.1016/j.bioorg.2010.04.003 14. John, T., Thomas, T., Abel, B., Wood, B.R., Chalmers, D.K., Martin, L.L.: How kanamycin A interacts with bacterial and mammalian mimetic membranes. Biochim. Biophys. Acta Biomembr. 1859(11), 2242–2252 (2017). https://doi.org/10.1016/j.bbamem.2017.08.016 15. Allouche, A.: Software news and updates gabedit – a graphical user interface for computational chemistry softwares. J. Comput. Chem. 32, 174–182 (2012). https://doi.org/10.1002/jcc 16. Lee, J., et al.: CHARMM-GUI input generator for NAMD, GROMACS, AMBER, OpenMM, and CHARMM/OpenMM simulations using the CHARMM36 additive force field. J. Chem. Theory Comput. 12(1), 405–413 (2016). https://doi.org/10.1021/acs.jctc.5b00935 17. Wu, E.L., et al.: CHARMM-GUI membrane builder toward realistic biological membrane simulations. J. Comput. Chem. 35(27), 1997–2004 (2014). https://doi.org/10.1002/jcc.23702 18. Baker, N.A., Sept, D., Joseph, S., Holst, M.J., McCammon, J.A.: Electrostatics of nanosystems: application to microtubules and the ribosome. Proc. Natl. Acad. Sci. U.S.A. 98(18), 10037–10041 (2001). https://doi.org/10.1073/pnas.181342398 19. Jurrus, E., et al.: Improvements to the APBS biomolecular solvation software suite. Protein Sci. 27(1), 112–128 (2018). https://doi.org/10.1002/pro.3280 20. Levy, R.M., Zhang, L.Y., Gallicchio, E., Felts, A.K.: On the nonpolar hydration free energy of proteins: surface area and continuum solvent models for the solute-solvent interaction energy. J. Am. Chem. Soc. 125(31), 9523–9530 (2003). https://doi.org/10.1021/ja029833a 21. Ra, L., Mb, S.: LigPlot+: multiple ligand-protein interaction diagrams for drug discovery. J. Chem. Inf. Model. 51, 2778–2786 (2011)
Collagen/Plasma-Polymerized Pyrrole Interaction: Molecular Docking and Binding Energy Calculations Teresa Gómez-Quintero, Iris Serratos-Alvarez , Rafael Godínez, and Roberto Olayo(B) Universidad Autónoma Metropolitana, Iztapalapa, 09310 Ciudad de Mexico, Mexico [email protected]
Abstract. The study of new biomaterials that can interact directly with cellular components seeks to improve the integration and functionality of the damaged organs or tissues in which they are implanted. Tissue engineering seeks materials that can provide a suitable microenvironment for proliferation, cell adhesion and differentiation; the interaction with components of the extracellular matrix is also of great interest. In this paper, the interactions generated between a collagen peptide and Plasma-Synthesized Polypyrrole (PPPy), a structure proposed by Kumar et al., were analyzed. Molecular docking analysis and computational determinations of the free energy of binding were performed. The PPPy structure used in this work has in its terminal branches an amino, nitrile and hydroxyl groups; direct interactions between the material and collagen were found, being the amino group the one that generated conformers with favorable binding free energy (Gb 0.05) was found for patients ranged in CI use time from 1 to 3 months, that reduces to 3.6 dBHL on average (p = 0.33) when
Implanted Pediatric Patient Early Audiometry
167
patient ranged in CI use time from 19 to 24 months. Table 2 shows average difference between ECR thresholds profile and audiometry along with statistical significance. Table 2. Average difference and correlation between ECR threshold profile and audiometry according to CI use time. CI use time [months]
Average difference dBHL
p
1 to 3
14.9 ± 5.6
< 0.05
4 to 6
10.3 ± 3.5
< 0.05
9.9 ± 3.6
< 0.05
10 to 12
10.6 ± 6.1
< 0.05
13 to 18
6.1 ± 2.4
< 0.05
19 to 24
3.6 ± 2.1
0.33
7 to 9
4 Discussion The basic principle behind implanted patient auditory experience lies on the electrical current level assigned to the incoming sound by cochlear implant system to elicit an auditory nerve response. Relationship between stimulation electrical current level and sound pressure level of incoming sound in modern cochlear implant systems is proportional, i.e., a soft sound corresponds to a low-level electrical current and a loud sound to a high-level electrical current. At the same time accepting that ECR represents auditory nerve response to electrical current, incoming sound minimum sound pressure level for ECR detection, ECR threshold, will be related to hearing threshold in that electrode. Therefore, by using as incoming sound a set of pip tones of variable sound pressure level and whose frequencies correspond to the band pass filter central frequency assigned to each intracochlear electrode, it is possible to obtain an objective approximation to implanted patient audiometry. According to average and standard deviation, ECR thresholds profile shows a stable behavior expressed as a mild hearing loss across patient age and CI use time, Fig. 1A and Fig. 1B. In contrast audiometry evolves from a moderately severe hearing loss with a significant standard deviation value for short age patients and scanty CI use time, Fig. 1A, to a mild hearing loss with significant standard deviation reduction when patient age and CI use time increase, Fig. 1B. Furthermore, as patient age and CI use time exceeds more than twelve months, differences between ECR thresholds profile and audiometry reduce significantly, Fig. 2, according to previous studies [13, 14]. We may say that as age and CI use time increase, sound field audiometry tends to follow ECR thresholds profile, Fig. 2. Thus, given ECR threshold objective nature, the
168
J. M. Cornejo Cruz et al.
gradual convergence over time between audiometry and ECR thresholds profile may be due to an improvement on the audiometry reliability. These results show that it is possible to obtain reliable pediatric implanted patient hearing thresholds as soon as CI is activated, reducing with this the time necessary to conduct him within normal hearing thresholds, and at the same time reducing the possibility of unsuitable electrical current stimulation scenarios.
5 Conclusion Results of this work indicate that is possible to get patient hearing thresholds through ECR, avoiding audiometry drawbacks when using with implanted pediatric patients. Additionally, individual electrode ECR threshold measure improves the whole CI fitting process by reducing electric current over or sub stimulation scenarios which impact on CI acceptance, reducing at the same time the required time to conduct patient within normal hearing thresholds, improving with this implanted patient rehabilitation expectations.
References 1. Brown, C.J., Abbas, P.J., Gantz, B.J.: Preliminary experience with neural response telemetry in the nucleus CI24 M cochlear implant. Am J. Otol 19(3), 320–327 (1998) 2. Brown, C.J., Hughes, M.L., Luk, B., Abbas, P.J., Wolaver, A., Gervais, J.: The relationship between EAP and EABR thresholds and levels used to program the nucleus 24 speech processor: data from adults. Ear Hear 21(2), 151–163 (2000) 3. Franck, K.H.: A model of a nucleus 24 cochlear implant fitting protocol based on the electrically evoked whole nerve action potential. Ear Hear. 23(Issue 1), 67S-71S (2002) 4. Franck, K.H., Norton, S.J.: Estimation of psychophysical levels using the electrically evoked compound action potential measured with the neural response telemetry capabilities of Cochlear Corporation’s CI24 M device. Ear Hear. 22(4), 289–299 (2001) 5. Hughes, M.L., et al.: A longitudinal study of electrode impedance, the electrically evoked compound action potential, and behavioral measures in nucleus 24 cochlear implant users. Ear Hear. 22, 471–486 (2001) 6. Hochmair-Desoyer, I., Schulz, E., Moser, L., Schmidt, M.: The HSM sentence test as a tool for evaluating the speech understanding in noise of cochlear implant users. Am J. Otol 18(6 Suppl), S83 (1997) 7. Zimmerling, M., Hochmair, E.S.: EAP recordings in ineraid patients - correlations with psychophysical measures and possible implications for patient fitting. Ear Hear. 23, 81–91 (2002) 8. Skinner, M.W., Holden, L.K., Holden, T.A., Demorest, M.E., Fourakis, M.S.: Speech recognition at simulated soft, conversational, and raised-to-loud vocal efforts by adults with cochlear implants. J. Acoust. Soc. Am. 101, 3766–3782 (1997) 9. Skinner, M.W., Holden, L.K., Holden, T.A., Demorest, M.E.: Comparison of two methods for selecting minimum stimulation levels used in programming the Nucleus 22 cochlear implant. J. Speech Lang Hear Res. 42, 814–828 (1999) 10. Baudhuin, J., Cadieux, J., Firszt, J.B., Reeder, R.M., Maxson, J.L.: Optimization of programming parameters in children with the advanced bionics cochlear implant. J. Am. Acad. Audiol. J. Am. Acad. Audiol. 5(23), 302–12 (2012)
Implanted Pediatric Patient Early Audiometry
169
11. Koka, K., Saoji, A.A., Litvak, L.M.: Electrocochleography in cochlear implant recipients with residual hearing: comparison with audiometric thresholds. Ear Hear. 38(3), 161–167 (2017) 12. Coulthurst, S., Nachman, A.J., Murray, M.T., Koka, K., Saoji, A.A.: Comparison of pure-tone thresholds and cochlear microphonics thresholds in pediatric cochlear implant patients. Ear Hear. 41(5), 1320–1326 (2020) 13. Quintana, A., Beltran, N., Granados, M.P., Chamlati, E., Mena, M., Cornejo, J.M.: Objective approach to audiometry in the pediatric implanted patient. In: Braidot, A., Hadad, A. (eds.) VI Latin American Congress on Biomedical Engineering CLAIB 2014; Paraná, Argentina 29, 30 & 31 October, vol. 49, pp. 707–710. IFMBE Proceedings. Springer, Cham (2016) https:// doi.org/10.1007/978-3-319-13117-7_180 14. Quintana López, A.K., Beltran Vargas, N.E., Granados Trejo, M.P., Cornejo-Cruz, J.M.: Electrical Cochlear response as an objective measure of hearing threshold and hearing performance evaluation in pediatric cochlear implant users. Mexican J. Biomed. Eng. 41(3), 72–86 (2020) 15. Cornejo, J.M., Quintana, A.K., Beltran, N.E., Granados, P.: Measuring implanted patient response to tone pips. Biomed. Eng. Online 20, 10 (2021)
Thermal Performance of a Triple Slot Antenna Considering Temperature Dependence of Thermal and Electrical Conductivity, Blood Perfusion and Tissue Metabolism Dalia Braverman-Jaiven1
and Citlalli Jessica Trujillo-Romero2(B)
1
Departamento de Estudios en Ingenier´ıa para la Innovaci´ on, Universidad Iberoamericana, Prol. Paseo de la Reforma 880, Lomas de Santa Fe, 01219 Ciudad de M´exico, Mexico 2 Divisi´ on de Investigaci´ on en Ingenier´ıa M´edica, Instituto Nacional de Rehabilitaci´ on-Luis Guillermo Ibarra Ibarra, Calz M´exico-Xochimilco 289, Coapa, Guadalupe Tlalpan, Tlalpan, 14389 Ciudad de M´exico, Mexico [email protected]
Abstract. Bone tumors account for less than 1% of all diagnosed cancer, however their morbidity and mortality are significant. Conventional treatments produce a variety of side-effects that decrease the patient’s life quality. Therefore, thermal ablation is proposed as a new treatment for bone malignancies. Commonly, computational models use constant values of tissue properties to reduce computational time. Hence, in this study the performance of a triple slot antenna to treat bone tumors, with thermal ablation, was predicted by considering the temperature dependence of several tissue properties. A parametric modeling study was implemented in COMSOL Multiphysics, based on the Finite Element Method. The Standing Wave Ratio SWR, the area of ablated bone and muscle (T > 60 ◦ C), the area of fat and skin at hyperthermia temperatures (T > 42 ◦ C), etc. were obtained. The SWR remained constant in all simulations with a value of 1.23. Skin and fat do not reach hyperthermia temperatures; moreover, less than 0.2 cm2 of muscle reach thermal ablation. The area of ablated bone varies from 3.69 cm2 to 5.2 cm2 , when thermal conductivity function goes from a positive slope to a negative one (k1 and k2 ), showing no difference when changing the functions for the other parameters.
1
Introduction
Cancer leads the list of diseases that cause death and decrease life expectancy [1]. Cancerous tumors can grow and spread to other parts of the body (metastasis) [2]. Bone tumors account for less than 1% of all diagnosed cancer but their morbidity and mortality are significant; moreover, they usually occur in children and adolescents [3].The most common treatments for bone tumors are c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 170–178, 2023. https://doi.org/10.1007/978-3-031-18256-3_18
Tissue Temperature Dependence in Thermal Ablation Therapy
171
chemotherapy and surgery, however they produce side-effects that decrease the life quality of patients, hence the need to develop different treatment options. Thermal therapy is based on the transfer of thermal energy into the body with a controllable source such as radiofrequency (RF) and microwave (MW) in protocols of elevated temperatures [4]; in thermal ablation (TA), temperatures higher than 50 ◦ C must be applied for 4–6 min. This protocol was found to produce immediate cell death [4]. This therapy has shown promising results in treating soft tissue cancer, as it provides a non-invasive and focal option for killing tumors [5–9]. TA therapy to treat bone cancer is growing in popularity. Researchers at the National Institute of Rehabilitation-LGII in Mexico City, have been developing micro-coaxial antennas designed to treat bone tumors [10–13]. The easiest and cheapest to develop are the micro-coaxial slot antennas, allowing the electromagnetic (EM) energy to be focused over a specific region, limiting the damage in other tissues. [14–16]. Although several modeling studies about TA are reported in the literature, the use of temperature dependence of tissue properties is barely used. This is mainly because more computational resources are needed to solve the equations that define the problem. Thus, constant values of tissue properties are commonly used. Therefore, to evaluate the effect of using temperature dependence of tissue properties over the antenna performance, a parametric model of a triple-slot antenna based on the finite element method (FEM) was implemented. It was previously designed and optimized by Lucero-Orozco et al. [17] to ensure the best performance by using temperature dependence of tissue properties (permittivity and conductivity) instead of constant parameters, as usual. In this case, a better temperature distribution over the bone tissue was observed. The present parametric study evaluates the effect of temperature dependence of thermal and electrical conductivity, as well as blood perfusion and tissue metabolism for each tissue involved in the model, adding more parameters to those used in previous works. The reason to include all these cases was to evaluate the effect of using different temperatures dependence functions reported in literature to describe the tissue properties. The antenna was inserted in a multi-tissue segment composed by bone, muscle, fat and skin. The model was created in COMSOL Multiphysics and the parameters evaluated to analyze the behavior of the antenna were the Standing Wave Ratio (SWR), the area of ablated tissue for bone and muscle, the area of fat and skin tissue that reaches hyperthermia, and the maximum temperature achieved during the therapy.
2 2.1
Methodology Micro-coaxial Triple-Slot Antenna
This study analyzes The triple-slot antenna (2.45 GHz). The slots are air notches that allow greater focus of energy near their location [15,17]; each slot is 2 mm long. The diameter of the antenna is that of a semi-rigid micro-coaxial cable UT-085C (2.19 mm) composed of an internal and external conductor. A polytetrafluoroethylene catheter was also included to prevent contamination of healthy tissue [17]. Figure 1a depicts the antenna geometry as well as its dimensions.
172
2.2
D. Braverman-Jaiven et al.
Antenna Modeling: Finite Element Method
The model of the antenna was implemented in COMSOL Multiphysics, a software based on the FEM. To evaluate its performance, it was introduced to a multi-tissue segment, composed by bone, muscle, fat and skin to asses the colateral damage that they may suffer during the therapy (Fig. 1b). To evaluate the temperature performance, eight measuring points or sensors with a 2 mm separation between them were located at the height of the slots, covering a region of 16 mm (See Fig. 1b). 2.2.1 Electromagnetic and Thermal Models Electromagnetic models are defined by the specific absorption rate (SAR) that determines the amount of energy absorbed by the tissues, as described in Eq. 1. Where E is the electric field generated by the antenna, σ the electrical conductivity of tissue (S/m) and ρ the tissue density [17].
Fig. 1. Antenna modeling. a) Geometry of the proposed triple-slot antenna b) 2D view of the modeling system, antenna inserted in a multi-tissue segment.
σ |E|2 (1) 2ρ Thermal models are based on the bioheat Eq. 2 [18]; where C is the heat capacity (J/KgK), ρ the blood density, k the thermal conductivity (W/m/k), Cb the heat capacity of blood, W the blood perfusion (kg/m3 /s), Tb the blood temperature, Q the heat generated by the tissues (W/m3 ), SAR the specific absorption rate described above and T the tissue temperature (37 ◦ C) [17]. SAR =
δT ) = Δ(kΔT ) + pQ + SAR − Cb W (T − Tb ) δt Table 1 shows the dielectric and thermal properties for each tissue. ρc(
(2)
Tissue Temperature Dependence in Thermal Ablation Therapy
173
Table 1. Thermal and dielectric properties of the tissues Parameter
Bone
Muscle Fat
Skin Blood
Relative permittivity ()
18.5
52.7
38
Electric conductivity (σ)
0.8050 1.74
10.8
–
0.2680 1.46 –
Thermal conductivity (k) 0.31
0.49
0.21
Metabolism [W/m3 ]
368.3
716
350
0.37 – –
Heat capacity [J/kg*K]
1300
3500
2300
3437 3639
–
Blood density [Kg/m3 ]
–
–
–
–
1050
Initial temperature [◦ C]
37
37
37
37
37
2.2.2 Evaluation of Thermal Dependence of Tissue Properties: A Parametric Study A parametric study evaluated the temperature dependence of tissue properties over the antenna performance. The parameters included were electrical conductivity (σ), thermal conductivity (k ), blood perfusion (W ) and tissue metabolism (m). To describe temperature dependence of electrical conductivity (σ), a ”piecewise” function was used. Three case-scenarios described in Table 2 were implemented [19]. Temperature dependence of thermal conductivity (k ) was described as shown in Table 2 [17]. To describe blood perfusion, the exponential functions 3, 4 were used [19]. Finally, tissue metabolism was described with a linear and exponential behaviour, as shown in Eqs. 5, 6 [20]. Table 2. Case scenarios implemented to evaluate the temperature dependence of electrical and thermal conductivity. T< 100 ◦ C Tissue properties (case scenarios)
Linear increase 1.5% ◦ C
σ1
X
σ2
X
σ3 k1
T> 100 ◦ C Linear increase 2%
Linear Decrease Decrease decrease rate 2 rate 4 1.5% ◦ C X X
X
X
X
k2
X
(T − 45)2 ) 12 (T − 45)2 ) = 0.036 + 0.036exp(− 12
Wmusc = 0.45 + 3.55exp(−
(3)
Wf at
(4)
174
D. Braverman-Jaiven et al.
mlinear = tissue constant ∗ (1 + 0.1 ∗ (T − 37))
(5)
mexponential = tissue constant ∗ exp((0.0953) ∗ (T − 37))
(6)
Several case-scenarios were implemented by combining the functions that describe the temperature dependence of tissue properties in order to evaluate the effect over the antenna performance. A mesh with 11798 elements was used. 2.3
Parameters Evaluation
To asses the effect of temperature dependence of tissue properties over the antenna performance, the Standing Wave Ratio (SWR) was evaluated. SWR describes the relationship between the applied power and the power delivered to the tissue by the micro-coaxial line, defining the impedance of the cable, as described in Eq. 7, where Γ is the reflection coefficient, that describes how much of a wave is reflected by an impedance discontinuity in the transmission medium [21]. The ideal SWR is 1, indicating that all the applied input power is transmitted to the tissue; therefore, a SWR greater than 1 indicates losses in the power. Moreover, the area of ablated bone and muscle (60 ◦ C), and the area of fat and skin that reached hyperthermia (42 ◦ C) were obtained. Finally, temperature profiles as time functions, recorded by the first (1 mm) and the last (16 mm) sensor were obtained, as well as the maximum reached temperatures. SW R =
3
1 + |Γ| 1 − |Γ|
(7)
Results and Discussion
The study was divided into two models, each swept three different functions to describe the thermal and electrical conductivity for every tissue: 1. Model 1: sweep of functions σ1 and k1 , σ1 and k2 and σ2 and k1 , shown on Table 2. 2. Model 2: sweep of functions σ2 and k2 , σ3 and k1 and σ3 and k2 , shown on Table 2. For each model, 8 simulations were computed to cover every function describing the blood perfusion and tissue metabolism (linear and exponential behaviour), giving a total of 16 simulations computed with 81 different cases. 3.1
SWR and Areas of Ablated Tissue
The SWR remained constant in all case-scenarios, with a value of 1.23. This means that temperature dependence of tissue properties does not affect the level of coupling between the microwave system and the tissue. Moreover, there are no differences between the parameters. Additionally, fatty tissue and skin never achieve temperatures higher than 42 ◦ C (hyperthermia) indicating they are not
Tissue Temperature Dependence in Thermal Ablation Therapy
175
affected during therapy. The muscle tissue that reach TA (T > 60 ◦ C) remains with values lower than 0.2 cm2 . However, the area of ablated bone tissue varies between 3.7 and 5.2 cm2 when the function that describes the temperature dependence of thermal conductivity is modified. When described with a positive slope (k1 ), the area of bone affected by TA reaches 5.2 cm2 , and with a negative slope (k2 ), it remains around 3.7 cm2 , in both cases regardless of the functions describing the other tissue properties (Fig. 2). Figure 3 shows the temperature distributions that present a significant difference. 3.2
Temperature Profiles
Figure 4 shows the temperature profiles generated around the three antenna slots, obtained by using the measuring points included in the model for the cases where the bone σ and k functions change since they are in which the results differ the most. The profiles were acquired for the first (1 mm), and the last (16 mm) sensor. The maximum reached temperature was obtained from the first sensor profile (111 ◦ C), however for most cases, it remained between 103 and 108 ◦ C. The regions near the slots on top and bottom of the antenna usually achieve the maximum temperatures, regions near the middle slot achieve the same temperature but usually take longer. By analyzing the temperature profiles from the last measuring point (16 mm), it was observed that the reached temperatures exceed 60 ◦ C considered as TA even when the distance from the slot increases. This means the antenna can generate damage up to 16 mm far from itself.
Fig. 2. Area of ablated bone tissue (60 ◦ C)
176
D. Braverman-Jaiven et al.
Fig. 3. Tissue temperature distribution with and input power of 10 W applied for 10 min. a) 3D temperature distribution for cases described with k1 , b) 2D temperature distribution for cases described with k1 , c) 3D temperature distribution for cases described with k2 , d) 2D temperature distribution for cases described with k2
Fig. 4. Temperature profiles obtained by the measuring points. a) For model 1,C1 = σ1 and k1 , C2 = σ1 and k2 and C3 = σ2 and k1 b) For model 2, C1 = σ2 and k2 , C2 = σ3 and k1 , and C3 = σ3 and k2 , S1 = measuring point in first slot, S2 = measurement point in slot 2, S3 = measurement point in slot 3
4
Conclusions
The results show that there was a significant difference between using k1 and k2 as 1.51 cm2 of ablated tissue and a 5 ◦ C in the maximum temperature differ between both functions. However, there was no significant difference when changing the electrical conductivity functions of the bone and the other adjacent tissues, as the maximum temperature varies 1 ◦ C and the area of ablated bone 0.4 cm2 . Additionally, there was no significant difference when changing the blood perfusion and tissue metabolism functions from linear to exponential. A similar study was reported by Lucero-Orozco et al.; however, this study differs mainly in the slots location. [17] reported slot location at 1 mm, 5 mm and 8 mm, while in this study the slots were located at 1 mm, 3.11 mm and 5.23 mm.
Tissue Temperature Dependence in Thermal Ablation Therapy
177
Moreover, different functions for σ and k were considered, and new parameters were introduced (blood perfusion and tissue metabolism). The main difference in the results were presented at the area of ablated bone, as [17] reports 1.09–2.2 cm2 while in this study they were 3.69–5.2 cm2 . This highlights the importance of further analyzing the antenna considering the exponential functions of σ that were not included in this study.
References 1. Sung, H., et al.: Global cancer statistics 2020: globocan estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J. Clin. 71(3), 209–249 (2021) 2. Cancer.Net. What is cancer? 3. Turner, S.P., Ferguson, J.L.: Bone cancer: diagnoses and treatment principles. Am. Family Phys. 98(4), 205–213 (2018) 4. Goldberg, S.N., Stauffer, P.R.: Introduction: thermal ablation therapy. Int. J. Hyperthermia 20(7), 671–677 (2004) 5. Camacho, J.C., Petre, E.N., Sofocleous, C.T.: Thermal ablation of metastatic colon cancer to the liver. Intervent. Radiol. 36(04), 310–318 (2019) 6. Cho, S.J., Baek, J.H., Chung, S.R., Choi, Y.J., Lee, J.H.: Thermal ablation for small papillary thyroid cancer: a systematic review. Thyroid Cancer Nodules 29(12), 1774–1783 (2019) 7. Saccomandi, P., Lapergola, A., Longo, F., Schena, E., Quero, G.: Thermal ablation of pancreatic cancer: a systematic literature review of clinical practice and preclinical studies. Int. J. Hyperthermia 35(1), 398–418 (2018) 8. Testoni, S.G.G., Healey, A.J., Dietrich, C.F., Arcidiacono, P.G.: Systematic review of endoscopy ultrasound-guided thermal ablation treatment for pancreatic cancer. Endoscopic Ultrasound 9(2), 83–100 (2020) 9. Ye, X., et al.: Expert consensus workshop report: Guidelines for thermal ablation of primary and metastatic lung tumors (2018 edition). Cancer Res. Therapeut. 14(4), 730–744 (2018) 10. Rico-Mart´ınez, G., Trujillo-Romero, C.J., Guti´errez-Mart´ınez, J.: Thermal ablation: an alternative to bone cancer. Investig. en Discapac. 7(1), 35–46 (2018) 11. Vera-Hern´ andez, A., Rico-Mart´ınez, G., Trujillo-Romero, C.J., Leija-Salas, L., Guti´errez-Mart´ınez, J.: Double slot antenna for microwave thermal ablation to treat bone tumors: Modeling and experimental evaluation. Electronics 10(7), 761 (2021) 12. Leija, L., Vera, A., Trujillo, C.J., Rico, G., Guti´errez, J.: Micro-coaxial slot antenna to treat bone tumors by thermal ablation: theoretical and experimental evaluation. IEEE Lat. Am. Trans. 16(11), 2731–2737 (2018) 13. Ram´ırez-Guzm´ an, T.J., et al.: Thermal evaluation of a micro-coaxial antenna set to treat bone tumors: design, parametric fem modeling and evaluation in multilayer phantom and ex vivo porcine tissue. Electron 10(18), 2289 (2021) 14. Romero, C.J.T., Martinez, G.R., Salas, L.L., Hernandez, A.V., Martinez, J.G.: Micro-coaxial slot antenna to treat bone tumors by thermal ablation: theoretical and experimental evaluation. IEEE Latin Am. Trans. 16(11), 2731–2737 (2018) 15. Vera-Hernandez, A., Trujillo-Romero, C.J., Rico-Martinez, G., Lejia-Salas, L., Gutierrez-Martinez, J.: Microwave ablation to treat bone tumors by using a double slot antenna: a modelling study. In: Blobal Medical Engineering Physics Exchanges (2017)
178
D. Braverman-Jaiven et al.
16. Rico-Martinez, G., Trujillo-Romero, C.J., Lejia-Salas, L., Vera-Hernandez, A., Gutierrez-Martinez, J.: Double slot antenna for microwave thermal ablation to treat bone tumors: modeling and experimental evaluation. Electronics 10(7), 761– 776 (2021) 17. Trujillo-Romero, C.J., Lucero-Orozco, N.B., Felix-Mart´ınez, G.J.: Antena microcoaxial de tres ranuras: Modelado de la dependencia t´ermica de los tejidos. Congreso Nacional de Ingenier´ıa Biom´edica 2020 (2020) 18. Pennes, H.H.: Analysis of tissue and arterial blood temperatures in the resting human forearm. J. Appl. Physiol. 85, 5–34 (1948) 19. Trujillo, M., Berjano, E.: Review of the mathematical functions used to model the temperature dependence of electrical and thermal conductivities of biological tissue in radiofrequency ablation. Int. J. Hyperthermia 29(6), 590–597 (2013). PMID: 23841882 20. Zhang, A., Zhu, Q., Ying, B., Xu, L.X.: Numerical study of the influence of water evaporation on radiofrequency ablation. ASME Summer Bioeng. Conf. SBC 13, 1–16 (2013) 21. Tang, W.: Voltage standing wave ratio definition and formula. In: Maxim Integrated (2012)
Modeling of the Interaction of Plasma-Polymerized Pyrrole with Immunoglobulin M (IgM) by Biocomputational Tools Esteban Rafael Ramírez Perez1 , Iris Natzielly Serratos2 , César Millán-Pacheco3 Salvador Tello-Solís2 , and Roberto Olayo-Valles4(B)
,
1 Posgrado en Ingeniería Biomédica, Universidad Autónoma Metropolitana-Iztapalapa,
09310 Mexico City, Mexico 2 Departamento de Química, Universidad Autónoma Metropolitana-Iztapalapa,
09310 Mexico City, Mexico 3 Facultad de Farmacia, Universidad Autónoma Metropolitana-Iztapalapa, 09310 Mexico City,
Mexico 4 Departamento de Física, Universidad Autónoma Metropolitana-Iztapalapa,
09310 Mexico City, Mexico [email protected]
Abstract. When entering the bloodstream, any exogenous biomaterial is covered by proteins, forming a coating known as protein crown. The opsonization of these biomaterials depends on the formation of the protein crown, within which immunoglobulin M (IgM) is the first antibody expressed during the primary immune response. It has been reported that plasma-polymerized pyrrole (PPPy)-coated biomaterials promote cell adhesion due to molecular interactions established between extracellular matrix proteins and amino groups (NH2 ) present on the PPPy surface. In this project, we model the interaction of a PPPy structure with a binding site of an IgM model through computational studies such as molecular docking, molecular dynamics, and binding energy. These studies allowed us to conclude that the binding between IgM and PPPy is favorable and driven by electrostatic interactions. Our results suggest that the immune system will be able to recognize PPPy-coated biomaterials, possibly unleashing an opsonic response. Our results should be considered in the design of PPPy-coated scaffolds for use in tissue engineering. Keywords: Immunoglobulin M (IgM) · Plasma polymerized pyrrole (PPPy) · Biocomputational tools
1 Introduction Upon entering the bloodstream or any other biological fluid, any exogenous object (such as a biomaterials) is covered by the proteins present in that fluid; this coating is known © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 179–187, 2023. https://doi.org/10.1007/978-3-031-18256-3_19
180
E. R. Ramírez Perez et al.
as a protein crown. Opsonization depends on the protein crown, and this response can be influenced by the coating of said biomaterial [1]. Opsonization is the process by which a pathogen is marked for ingestion and destruction by phagocytes. The immune system responds to antigens by producing cells that attack the pathogen or by producing special proteins called antibodies [2]. Each antibody molecule contains a constant region (Fc, fragment crystallizable) and a variable region (Fab, fragment antigen-binding). It is in this variable region that the antigen-antibody binding takes place to form the immunocomplexes. Once the immune complexes are formed, they can bind to the phagocytes through the Fc region of the antibodies, thus facilitating phagocytosis by macrophages and neutrophils [3]. Immunoglobulins (Ig) or antibodies are glycoproteins produced by plasma cells. There are different types of Ig (Fig. 1), depending on their location and function [4].
Fig. 1. The five different types of immunoglobulins.
Immunoglobulin M (IgM) is a pentameric protein with a molar mass of 498.25 kDa. It is involved in the primary immune response to infectious agents or antigens. Generally, IgM activates the classical pathway that leads to opsonization. IgM consists of five monomeric units, linked by their Fc regions by means of disulfide bonds in the center, with ten binding sites in its Fab region on the periphery and a J chain (essential in the polymerization of monomers) (Fig. 2). Within each binding site, the Pro436 has been shown to be important due to its ability to unleash an opsonic response [5, 6, 7].
Modeling of the Interaction of Plasma-Polymerized Pyrrole
181
Fig. 2. Immunoglobulin M (IgM) and its binding sites (highlighted in red circles) in the Fab segment.
We are interested in characterizing the interaction of immunoglobulins with biomaterials. Plasma polymerized pyrrole (PPPy) surfaces are of interest since this type of material has been used in the treatment of traumatic spinal cord injuries [8]. PPPy coatings have also been reported to promote cell adhesion and proliferation in scaffolds for tissue engineering [9]. Kumar et al. [10], synthesized and characterized PPPy films, proposing a representative chemical structure (Fig. 3). The bioactivity of PPPy is due to the molecular interactions that are established between the proteins of the extracellular matrix and amino groups present in the PPPy structure, creating an electrostatic environment that favors binding [11]. It has been shown that characterizing the interaction
Fig. 3. Chemical structure of plasma polymerized pyrrole fragment proposed by Kumar et al. [10].
182
E. R. Ramírez Perez et al.
between protein with ligands can show the mechanisms involved in biological processes to understand the nature of emerging therapies (photodynamic therapy, tissue repair, cellular proliferation, neurodegenerative and inflammatory disorders treatment, etc.). The following protocol was implemented as described in previous reports [12, 13, 14]. There are currently no reports that study the interaction of a PPPy surface with the immunoglobulins present in blood. The objective of this project is to determine the nature of the protein-ligand union between PPPy and IgM and to determine its stability. Our results will allow us to predict whether the human body can recognize this biomaterial and trigger an opsonic response.
2 Methodology 2.1 Molecular Docking Studies These studies were performed with the Autodock Vina software, this technique allowed predicting binding sites of the PPPy structure on IgM efficiently. This software helps generate “poses” at defined coordinates (in this case, its active site), to predict the position in which there would be an interaction between the PPPy structure and the target site [15]. The best generated “poses” were chosen and processed for electrostatic calculations. The IgM structure was downloaded from the Protein Data Bank (PDB Code: 6KXS) and subsequently minimized with the charmm-gui server (https://charmm-gui.org/). 2.2 Binding Energy Calculations (Gb ) The binding free energy (Gb ) can be determined from the electrostatic (Gb,elec ) and non-electrostatic (Gb,no-elec ) type contributions [16] as shown in Eq. 1 and in turn are determined as shown in Eqs. 2 and 3: Gb = Gb,elec + Gb,no−elec
(1)
Gb,elec = Gb,sol + Gb,coul
(2)
Gb,no−elec = γ ASAIgM −PPPy − ASAIgM − ASAPPPy
(3)
The electrostatic energies were determined with the Adaptive Poisson-Boltzmann (APBS) software [16]. This software implicitly considers the contribution of water to biomolecules, taking as reference the value of its dielectric constant (E), which is hugely different from the value assigned to the protein. The solvation energy (Gb,sol ) is the energy required to remove the solvent from the active site for the ligand to bind to the compound. The Coulombic energy (Gb,coul ) is the energy of interaction of the charges between IgM and the PPPy molecule in a homogeneous medium. The evaluation of non-electrostatic interactions (Gb, no−elec ) is obtained from the change in surface area accessible to the solvent for each element, caused by the conformational change of both the IgM and the structure of PPPy together and each one individually (Solvent Accessible Surface, ASA) and a parameter belonging to the surface tension (γ) [16]. Visual Molecular Dynamics (VMD) software was used [17] to calculate the area of the interface that is hidden from the solvent in the complex.
Modeling of the Interaction of Plasma-Polymerized Pyrrole
183
2.3 Molecular Dynamics The gromacs 2019.2 software was used together with charmm36 preset parameters to simulate the IgM structure in 100 ns (PDB Code: 6KXS). With the charmm-gui server (https://charmm-gui.org/), a simulation box was created maintaining a distance of 10 Å between the protein and the edge of the box, as well as a salt concentration of 0.15 M. KCl; this to use the hydrogen mass partitioning approximation, so the time interval used was 4 fs with periodic boundary conditions. The temperature was controlled with a NoseHoover type thermostat frequently every twenty steps, and the pressure was maintained with a Parrinello-Rahman type piston. For the electrostatic environment, the ParticleMesh Ewald (PME) approximation was implemented. It was decided to analyze the last 50 ns of simulation by means of clustering, taking the value of the root mean square deviation of the alpha carbons position. Structures with root mean square deviation less than 2 Å would be considered as a cluster. This grouping of structures by similarity between them (clustering) was done in gromacs 2019.2 with the gromos method. After molecular dynamics, the most representative structure of molecular dynamics was implemented and compared with the minimized structure before dynamics to observe the change in Gb over time, observing that it remained stable to verify that the interaction was maintained. That is, different poses and binding energy calculations were obtained for the structure before and after molecular dynamics.
3 Results The crystallographic structure of IgM (PDB Code: 6KXS) is shown in Fig. 4A [18], and Fig. 4B shows the detail of the binding site where the amino acids Aspartic Acid (D), Leucine (L), Proline (P) and Serine (S) are involved (active site formed by the sequence DLPSP). It is possible to see that this site is in the Fab segment of the chain (variable region). A review was made in the literature to choose the appropriate functional groups for the PPPy structure, there is previous work where NH2 , CH3 and OH groups were implemented in the highlighted sites according to Kumar et al. [10, 11]. Docking studies were conducted with Autodock Vina to generate couplings, from which were selected the three poses that generated the nearest binding by hydrogen bonds between active site of the IgM structure and the R3 group of the PPPy structure. The position of the PPPy structure near the IgM binding site can be observed (Fig. 5A); in particular, the interaction between the amino group located in the longest part of the main chain of the PPPy structure and Proline 434. The interaction maps allow us to observe this binding in detail (Fig. 5B). The binding free energies for the complexes with the nearest interactions as previously mentioned were calculated as described in Sect. 2.2 and the results are shown in Table 1. As can be seen, the greatest free energy contribution is the coulombic component, while the non-electrostatic interactions contribute to a lesser extent to the binding. Most of the complexes obtained by the molecular docking study between IgM and PPPy indicate that binding is energetically favorable with electrostatic interactions having the largest contribution.
184
E. R. Ramírez Perez et al.
Fig. 4. A) Structure of the IgM downloaded from the Protein Data Bank with PPPy next to one of the active sites in the Fab region (PDB: 6KXS). B) Amino acids (DLPSP) involved in the active site of IgM (enlarged). C) Chemical structure with its 3D equivalent of plasma polymerized pyrrole (PPPy) used for docking tests, where R1 = CH3 , R2 = OH, R3 = NH2 .
Fig. 5. A) Position of PPPy structure near the Fab region of the original IgM structure (first complex from Table 1). B) Interaction map generated in the software Ligplot+ [19] for this pose.
Regarding the structure of the dynamics, it was selected in the last 50 ns by clustering analysis. Molecular docking was done to this representative structure and the poses that established the nearest contact with the active site were processed. Subsequently, the binding energy from the first IgM-PPPy structure (before molecular dynamics) was compared to the binding energy obtained from the most representative structure obtained from molecular dynamics (Fig. 6); this to check if the Gb was stable over time, which can be seen in Table 2 where the electrostatic component is again predominant. Likewise,
Modeling of the Interaction of Plasma-Polymerized Pyrrole
185
Table 1. Binding free energy results (minimized IgM structure)
Structure
,
,
,
−
(kJ/mol)
(kJ/mol)
(kJ/mol)
(kJ/mol)
IgM-PPPy 1
37
-299
-15
-276
IgM-PPPy 2
60
-268
-20
-228
IgM-PPPy 3
44
-271
-16
-244
Mean
47 (11)
-279 (17)
-17 (3)
-249 (24)
Fig. 6. A) Position of PPPy structure near the site in the Fab region of the IgM structure generated from molecular dynamics (first complex from Table 2). B) Interaction map generated in Ligplot+.
magnitudes like those observed previously on Table 1 were obtained in the new structure, meaning that the bond was preserved over time. The difference between the mean values of Gb before and after the molecular dynamics was 18 kJ/mol, which is lower than the standard deviation. This verifies the predominant electrostatic-type interactions by the combination of functional groups chosen for the PPPy structure [11]. This can be a quantitative indicator of the initiation of an opsonic response by an antibody, as a first line against a threat or antigen.
186
E. R. Ramírez Perez et al.
Table 2. Binding free energy results (IgM structure generated after molecular dynamics)
Structure
(kJ/mol)
(kJ/mol)
(kJ/mol)
(kJ/mol)
IgM-PPPy 1
4
-274
-3
-273
IgM-PPPy 2
7
-270
-4
-267
IgM-PPPy 3
7
-264
-4
-261
Mean
6 (2)
-269 (5)
-4 (0)
-267 (6)
4 Conclusion The results of biocomputational studies show that binding of IgM and PPPy is energetically favorable due to the presence of strong mostly electrostatic interactions, confirming that the implementation the PPPy structure with the proposed residues (R2 = NH2 , R1 = CH3 and R3 = OH) would have this type of interaction. Likewise, the reaction obtained, may mean a behavior of an opsonization reaction, which typically triggers the classic response of the complement system by the immune system by containing IgM distribution. Likewise, it is important to mention that this work contemplates the subsequent analysis of the second site involved in this type of interaction located in the Fc chain of IgM) [20] which is expected to have an interaction of a similar electrostatic nature, but of greater magnitude, as it is the site that interacts with the complement system of the immune system and therefore requires a greater force of interaction to trigger a subsequent opsonic response. Experimental studies are also underway. The results will allow the evaluation of the viability of PPPy as a biomaterial implemented in tissue engineering.
References 1. Van Hong, N.: Protein corona: a new approach for nanomedicine design. Int. J. Nanomed. 12, 3137 (2017) 2. Vu, V.P.: Immunoglobulin deposition on biomolecule corona determines complement opsonization efficiency of preclinical and clinical nanoparticles. Nat. Nanotechnol. 14, 260–268 (2019) 3. Institute for Quality and Efficiency in Health Care (IQWiG): https://www.ncbi.nlm.nih.gov/ books/NBK279364/. Last accessed 23 Apr 2022 4. Riedel, S., Morse, S.A., Mietzner, S.: Jawetz Melnick & Adelbergs Medical Microbiology, 28th edn. McGraw Hill Professional (2019) 5. Perkins, S.J.: Solution structure of human and mouse immunoglobulin M by synchrotron X-ray scattering and molecular graphics modelling: a possible mechanism for complement activation. J. Mol. Biol. 221, 1345–1366 (1991)
Modeling of the Interaction of Plasma-Polymerized Pyrrole
187
6. Microbe Online: https://microbeonline.com/igm-antibody-structure-properties-functions-cli nical-significance/. Last accessed 5 May 2022 7. Wright, J.F.: C1 binding by murine IgM. The effect of a Pro-to-Ser exchange at residue 436 of the mu-chain. J. Biol. Chem. 263, 11221–11226 (1988) 8. Olayo, R.: Tissue spinal cord response in rats after implants of polypyrrole and polyethylene glycol obtained by plasma. J. Mater. Sci. Mater. Med. 19, 817–826 (2008) 9. Islas-Arteaga, N.C.: Electrospun scaffolds with surfaces modified by plasma for regeneration of articular cartilage tissue: a pilot study in rabbit. Int. J. Polym. Mater. Polym. Biomater. (2019) 10. Kumar, D.S.: Optical and electrical characterization of plasma polymerized pyrrole films. J. Appl. Phys. 93, 2705–2711 (2003) 11. Serratos, I.N.: Modeling integrin and plasma-polymerized pyrrole interactions: chemical diversity relevance for cell regeneration. Sci. Rep. 9, 1–12 (2019) 12. Serratos, I.N.: Early expression of the receptor for advanced glycation end products in a toxic model produced by 6-hydroxydopamine in the rat striatum. Chem. Biol. Interact. 249, 10–18 (2016) 13. Vicente-Escobar, J.O., García-Sánchez, M.Á., Serratos, I.N., Millán-Pacheco, C., Tello-Solís, S.R.: Binding of two tetrasulfophthalocyanines (Fe(III) and Metal-Free) to lysozyme: fluorescence spectroscopic and computational approach. J. Fluoresc. 31(3), 787–796 (2021). https:// doi.org/10.1007/s10895-021-02710-7 14. Serratos, I.N.: Modeling the interaction between quinolinate and the receptor for advanced glycation end products (RAGE): relevance for early neuropathological processes. PLoS One 10, e0120221 (2015) 15. Sharp, T.H.: Insights into IgM-mediated complement activation based on in situ structures of IgM-C1-C4b. Proc. Natl. Acad. Sci. 116, 11900–11905 (2019) 16. Baker, N.A.: Biomolecular applications of Poisson–Boltzmann methods. Rev. Comput. Chem. 21, 349–379 (2005) 17. Humphrey, W.: VMD: visual molecular dynamics. J. Mol. Graph. 14, 33–38 (1996) 18. Trott, O.: AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31, 455–461 (2010) 19. Wallace, A.C.: LIGPLOT: a program to generate schematic diagrams of protein-ligand interactions. Protein Eng. Des. Sel. 8, 127–134 (1995) 20. Czajkowsky, D.M.: IgM, FcμRs, and malarial immune evasion. J. Immunol. 184, 4597–4603 (2010)
Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis Ana Paola Leyva-Aizpuru1 , Yoshua Alberto Quezada-García2 , Graciela Ramirez-Alonso3 , Luis Carlos Hinojos-Gallardo1 , and Javier Camarillo-Cisneros1(B) 1 Laboratorio de Física Química Computacional, Facultad de Medicina y Ciencias Biomédicas,
Universidad Autónoma de Chihuahua, 31125 Chihuahua, Chihuahua, México [email protected] 2 Departamento de Traumatología y Ortopédia, Hospital General de Ciudad Juárez, Ciudad Juarez, Chihuahua 32330, México 3 Facultad de Ingeniería, Universidad Autónoma de Chihuahua, 31125 Chihuahua, Chihuahua, México
Abstract. Due to the accelerated emergence of drug-resistant bacterial strains, better techniques are needed to enable intelligent antibiotic design. In 2019, drugresistant bacterial strains were the direct cause of at least 1.27 million deaths internationally. Thus, new strategies are being developed for a faster and more costeffective process focusing on analyzing so-called older antibiotics. Nitrofurans are documented as the family of antibiotics least prone to bacterial resistance. Therefore, they are molecules from which we can learn by analyzing their structure along with their action mechanism. In this work, we use computational chemistry to calculate the minimum conformations energies, highest occupied molecular orbital, lowest unoccupied molecular orbital, and gaps energies of nitrofurans and their derivative products. Subsequently, the structure of each molecule and its charge distribution were analyzed to understand their reactivity. The charge distribution in the nitrofurans and their nitroso derivatives was concentrated in the Furan-nitro group, which explains the antibacterial properties. Keywords: Nitrofuran · Computational chemistry · HOMO-LUMO
1 Introduction Bacterial resistance is a global silent threat that claims thousands of lives each year [1]. Although the evolution of organisms goes hand in hand with natural selection, the introduction of antibiotics into clinical practice has accelerated the emergence rate of resistant strains [2]. Antibacterial resistance has been documented only a few years after the therapeutic use of new antibiotics [3]. On the other hand, developing a new drug can take 13 to 20 years and require an investment of close to a billion dollars [4]. As a support tool for new developments, computer-aided drug design has emerged. This alternative relies on computational chemistry techniques such as drug modeling to find © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 188–195, 2023. https://doi.org/10.1007/978-3-031-18256-3_20
Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis
189
the best molecular candidates before starting in vitro design [5]. A new strategy for antibiotic design is being pursued involving reintroducing previously used antibiotics but displaced as the development of new generations progresses [6]. These old antibiotics are effective against infections caused by bacteria that are multiresistant to current firstline antibiotics. It is important to raise awareness of the impact of drug resistant strains in a context where current methodologies for designing new antibiotics are not sufficient to deal with the problem. The nitrofurans family was introduced in 1940 [7]. Few strains resistant to nitrofurans have been reported compared to any other family of antibiotics [8]. Although the mechanism of action of these antibiotics has not been fully described, it is observed that they have multiple targets in bacterial cells [9]. This fact makes them good candidates to act as molecular scaffolds for the repositioning and redesigning of new antibiotics. We can state that biomedical engineering is in charge of applying the principles of engineering to analyze and solve problems in the areas of health and biology. Computer-aided drug design is an approach that can contribute to the fight against drug-resistant bacteria and drug shortages based on new methodologies that streamline the design and intelligent production of drugs. In this study, the electronic structure, the highest occupied molecular orbital (HOMO), the lowest occupied molecular orbital (LUMO), and their energy gaps were analyzed to explain the reactivity of these molecules. Understanding how the composition and structure of these molecules influence their reactivity and possible interactions with other molecules can guide the redesign of old and new antibiotics. Furthermore, by performing in silico antibiotic analysis, a methodology can be designed that applies the level of theory necessary to reproduce results with excellent and reliable accuracy. Once standardized, it will be applied to different families of antibiotics for further analysis.
2 Nitrofuran Antibiotics This family of antibiotics is formed by nitrofurazone, nitrofurantoin, furazolidone and furaltadone. Their action mechanism is based on bacterial enzymatic inhibition of several cellular processes. Being able to act on multiple targets, nitrofurans are less prone to developing resistance to bacteria than any other family of antibiotics [10]. Nitrofurans are classified as broad-spectrum antibiotics but exhibit antiparasitic and antifungal activity [11]. Nitrofurans are synthetic drugs derived from a common scaffold molecule: a furan core and a nitro group at the C2 position, which confers antibacterial properties [12]. Their chemical structure can be observed in Fig. 2. Nitrofurans need to be modified to initiate their antimicrobial activity. Bacterial nitroreductases mediate this activation via reduction of the nitro group. The mode of action has not yet been fully described, but several model organisms have been approached; The proposed mechanism of nitrofurans on E. coli states that the antibiotic goes through a stepwise 2-electron reduction of the nitro group in the C2 position, which results in nitroso and hydroxylamine derivates [13]. Once the nitrofuran is modified, it is capable of damaging bacterial DNA, causing oxidative stress, and inhibiting protein and RNA biosynthesis, as is depicted in Fig. 1.
190
A. P. Leyva-Aizpuru et al.
Fig. 1. Pictograph of E. coli bacteria. The reduction of nitrofuran is mediated by the enzyme nitroreductase type I. In the lower part are shown the antibiotic targets: DNA, RNA, and protein synthesis.
3 Computational Chemistry Computational chemistry uses atomic models and numerical simulation to study various systems based on the principles of theoretical chemistry. Techniques such as structure modeling, high-throughput screening, and quantitative structure-activity relationship analysis allow calculations of the geometric structure and physical and chemical properties of the systems of interest [14]. One of modern chemistry’s most extensive computational applications is Density Functional Theory (DFT). This level of theory calculates the energy from the electronic density of a molecule in the ground state function of each atom position [15]. DFT considers the interactions between neighboring atoms, electrons, and external fields, which gives it greater complexity than the levels of theory based on classical mechanics, such as Force-Field [16]. Compared to the experimental method, computational methods can reduce the time and cost in the initial stages of the design of antibiotics. This alternative can complement the traditional approach to drug development, which can take 13 to 20 years and require an investment of millions of dollars with no guarantee of success [17]. Several studies have been published on the study of nitrofuran´s mechanism of action using computational chemistry tools. For example, Day et al. [18] presented molecular dynamics simulations on the E. coli nitroreductase – nitrofurantoin complex. This showed that nitrofurantoin binds with the furan ring to the enzyme, not the reduced nitro group. In addition, Manin et al. [19] studied the thermodynamics of furazolidone’s, dissolution processes and nitrofurantoin crystals via DFT calculations. This approach allows studying the relationship between molecular and crystal structures and pharmaceutically relevant properties.
4 Materials and Methods Initial molecular geometries were obtained from the PubChem database [20]. The molecules used correspond to 4 nitrofurans: Nitrofurazone, Nitrofurantoin, Furazolidone and Furaltadone. Then, the first step of coarse geometric optimization was carried
Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis
191
out to find the ground state geometries. A semi-classical approach was used through Force Field methodology with a maximum of 5 000 configurations through the Tinker software [21]. Then, the five most stable structures were taken from the resulting conformational space for another round of fine geometric optimization. This time it was calculated employing DFT with the m06/lanl2dz level of theory [22, 23]. Calculation configuration was established according to coding the full-potential all-electrons of centered number bases of each atom in the FHI-aims software [24]. DFT calculations work as approximations of the wave function applied to molecular systems. For these calculations, the convergence conditions for the required accuracy set were: Convergence energy was 1x10–6 Hartree, a maximum displacement between atoms was 0.18 Å, and the maximum force was 0.003 Hartree/Bohr. Subsequently, molecules corresponding to the nitroreduction derivatives of each nitrofuran were created from the nitroreductase-mediated reaction pathway proposed by Le & Rakonjac [13]. Then, the HOMO, LUMO, and their energy gaps were calculated for each nitrofuran and each reaction step to estimate and compare the reactivity of the compounds. Total computational time was 109 h and 25 min. Calculations were carried out in 4 computers with i7 processor, eighth generation, 16 GB RAM, Nvidia GTX-1080 video card with 3584 CUDA cores.
5 Results and Discussion The optimized structures of the four nitrofurains are shown in Fig. 2 the motivation for performing a geometry optimization is the physical significance of the obtained structure: optimized structures often correspond to a substance found in nature. The process seeks to find the geometry of a particular arrangement of the atoms that represents a local or global energy minimum. The geometry of the structure can be used in various experimental and theoretical investigations [25]. The composition of the nitro group at the C2 position makes each nitrofuran different. For example, nitrofurazone has an
Fig. 2. Optimized geometry structure of nitrofurans. a) Nitrofurazone, b) Nitrofurantoin, c) Furazolidone and d) Furaltadone.
192
A. P. Leyva-Aizpuru et al.
n-methyl urea group. Furaltadone has a 3–5-nitrofuran-2-yl-methylideneamino-1, 3oxazolidin-2-one, furazolidone has a 3-methyl-1, 3-oxazolidin-2-one and nitrofurantoin has a 1-methylimidazolidine-2, 4-dione. Figure 3 shows a comparison between each bond of the optimized geometries in each nitrofuran and its derivatives. In the first reduction step, all nitro derivatives showed the highest dispersion in bond 6 between N and C2 (0.006 Å). For the second reduction step, all hydroxylamine derivatives showed the highest dispersion at bond 8 between N and O (0.023 Å), these results are in good agreement with the crystalized structures reported by Day et all with a 1.25 Å resolution [18].
Fig. 3. Comparison between de bond lengths of nitrofurans and their derivates.
Understanding the electronic properties of antibiotics can be achieved by analyzing the HOMO and LUMO of each molecule. Figure 5 shows the surfaces corresponding to the HOMO and LUMO states of the four nitrofurans and their derivates. The HOMO surface represents the charge distribution in the molecule in its ground state. When the nitrofuran molecule enters the body, it is exposed to several factors, such as increased temperature and acid pH. These factors bring it out of its ground state, since a molecule exposed to the stresses caused by external forces changes its structure and thus its energy state [26]. The LUMO surface represents the charge distribution of a molecule when it is out of its ground state. The LUMO represents the ability to obtain an electron and the HOMO represents the ability to donate an electron. The energy gap between the HOMO and LUMO energies is the lowest energy necessary for electronic excitation. Figure 4 shows a diagram for a better understanding of these concepts. According to the diagram in Fig. 5, there is a tendency for the charge distribution to migrate to the furan-nitro group side of the molecules from the HOMO to the LUMO surfaces.
Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis
193
Fig. 4. Diagram of the transition from HOMO to LUMO.
Fig. 5. HOMO and LUMO values along with the reaction pathway for each nitrofuran and the derivatives.
Figure 5 shows the reaction pathway for the reduction of nitrofurans by the nitroreductase. The first reduction step results in the loss of 2 electrons and the formation of a nitroso product. A second reduction step removes another 2 electrons giving a hydroxylamine product. The energy gap between the HOMO-LUMO orbitals of the nitroso derivatives has a slight variation, max = 865 eV, in comparison with the nitrofuran before reduction. On the other hand, the energy gap between the HOMO-LUMO orbitals of the hydroxylamine product has a max variation of = 1.785 eV compared with the nitrofuran before reduction. According to the computational results of Khan et al., the
194
A. P. Leyva-Aizpuru et al.
energy gap reported in this study has good approximation to the energy gap reported in this study [27].
6 Conclusion From the computational physicochemical characterization of nitrofurans and their derivatives, we have found the most reactive surfaces by excess charges on the molecular sites reported with antibiotic activity. Since the results obtained had a good approximation to the experimental reports, we can study the change of charge distribution along with the mechanism of action at the molecular level. Understanding such changes in each nitrofuran molecule and their derivatives makes it possible to support the intelligent design of effective antibiotics against drug-resistant bacterial strains.
References 1. Centers for Disease Control and Prevention: «Antibiotic Resistance Threats Report,» de Atlanta: Centers For Disease Control And Prevention, Atlanta (2019) 2. Hasan, C., Dutta, D., Nguyen, N.: Revisiting Antibiotic Resistance: Mechanistic Foundations to Evolutionary Outlook. Antibiotics 11(40), 23 (2021) 3. Errecalde, J.: Uso de antimicrobianos en animales de consumo. Incidencia del desarrollo de resistencias en salud pública. Universidad Nacional de la Plata, La plata (2004) 4. Macalino, S.J., Gosu, V., Hong, S., Choi, S.: Role of computer-aided drug design in modern drug Discovery. Archives of Pharmacal Research 38(9), 1686–1701 (2015) 5. Willems, H., De Cesco, S., Svensson, F.: Computational Chemistry on a budget: supporting drug discovery with limited resources. Journal of Medicinal Chemistry 63(18), 10158–10169 (April 2020) 6. Theuretzbacher, U., et al.: Reviving old antibiotics. J Antimicrob Chemother 70(8), 2177– 2181 (2015) 7. Rezaei, S., Akbari, S., Rahmani, F., Varnousfaderani, S., Gomroki, S., Jafarzadeh, E.: Nitrofurans as Potent Antibacterial Agents: A Systematic Review of Literature. Int. J. Adv. Biologi. Biomed. Res. 10(2), 126–138 (2022) 8. Le, V., Davies, I., Moon, C., Biggs, W.D.P., Rakonjac, J.: Novel 5-nitrofuran-activating reductase in Escherichia coli. Antimicrob Agents Chemother 63(11), 868–919 (2019) 9. Zuma, N., Aucamp, J., N´Da, D.: An update on derivatisation and repurposing of clinical nitrofuran drugs. European Journal of Pharmaceutical Sciences 140, 13 (2019) 10. Brunton, L., Chabner, B., Knollmann, B.: Goodman & Gilman´s The Pharmacological Basis of Therapeutics. Mc Graw Hill, New York (2020) 11. Murugasu-OEI, B., Dick, T.: Bactericidal activity of nitrofurans against growing and dormant Mycobacterium bovis BCG. The Journal of antimicrobial Chemotherapy 46(6), 917–920 (2000) 12. Bot, C., et al.: Evaluating 5-Nitrofurans as Trypanocidal Agents. Antimicrobial Agents and Chemotherapy 57(4), 1638–1647 (2013) 13. Le, V., Rakonjac, J.: Nitrofurans: Revival of an “old” drug class in the fight against antibiotic resistance. Plos Pathogens 17(7), 7–8 (2021) 14. Willems, H., De Cesco, S., Svensson, F.: Computational chemistry on a budget: supporting drug discovery with limited resources. Journal of Medicinal Chemistry 63(18), 10158–10169 (2020)
Nitrofuran Antibiotics and Their Derivatives: A Computational Chemistry Analysis
195
15. Magos, G., Lorenzana, M.: Las fases en el desarrollo de nuevos medicamentos. Revista de la Facultad de Medicina de la UNAM 52(6), 260–264 (2009 Diciembre) 16. Deriabina, A., Ledesma, J.O., González, E., Herrera, J.N., Poltev, V.: Desarrollo de un campo de fuerzas de mecanica molecular ‘para la interaccion de Na+ con agua. Revista Mexicana de Fisica 52(1), 74–79 (2006 Febrero) 17. Wouters, O., Mcnkee, M., Luyten, J.: Estimated research and development investment needed to bring a new medicine to market, 2009–2018. Journal of the American Medical Association 323(9), 844–853 (2020) 18. Day, M., et al.: The structures of E. coli NfsA bound to the antibiotic nitrofurantoin; to 1,4-benzoquinone and to FMN. Biochemical Journal 478(13), 2601–2617 (2021) 19. Manin, A., Drozd, K., Voronin, A., Churakov, A., Perlovich, G.: A combined experimental and theoretical study of nitrofuran antibiotics: crystal structures, DFT computations, sublimation and solution thermodynamics. Molecules 26(11), 20 (2021) 20. Kim, S., et al.: Pubchem, PubChem in 2021: new data content and improved web interfaces. Nucleic Acids Res. (2021). [En línea]. Available: https://pubchem.ncbi.nlm.nih.gov/. [Último acceso: 2022] 21. Rackers, J., et al.: Tinker 8: Software Tools for Molecular Design. Journal of Chemcial Theory and Computation 14(10), 5273–5289 (2018 February) 22. Chiodo, S., Russo, N., Sicilia, E.: LANL2DZ basis sets recontracted in the framework of density functional theory. The Journal of Chemical Physics 125(10), 104107–104115 (2006) 23. Zhao, Y., Truhlar, D.: The M06 suite of density functionals for main group thermochemistry, thermochemical kinetics, noncovalent interactions, excited states, and transition elements: two new functionals and systematic testing of four M06-class functionals. Theoretical Chemistry Accounts 120(1), 215–241 (2008) 24. Blum, V., et al.: Ab initio molecular simulations with numeric atom-centered orbitals. Computer Physics Communications 180(11), 2175–2196 (2009) 25. Schlegel, H.: Geometry optimization. Wiley Interdisciplinary Reviews: Computational Molecular Science 1(5), 790–809 (2011) 26. Jensen, F.: Molecular Properties, de Introduction to Computational Chemistry, pp. 315–349. John Wiley & Sons Ltd, West Sussex (2007) 27. Tandon, P., Khan, E.: Structural and Reactivity Analyses of Nitrofurantoin–4dimethylaminopyridine Using Spectroscopic and DFT. Crystals 9(413), 14 (2019) 28. Fariña, N.: Resistencia bacteriana: un problema de salud pública mundial de difícil solución. Memórias del Instituto de Investigaciones en Ciencias de la Salud 14(1), 4–5 (2016) 29. Tathe, A., Gupta, V., Sekar, N.: Synthesis and combined experimental and computational investigations on spectroscopic and photophysical properties of red emitting 3-styryl coumarins. Dyes and Pigments 119(1), 49–55 (2015) 30. Elhassanny, R.V.D.-A.A.A.E.: Drugs Targeting Bacterial DNA, de Brody’s Human Pharmacology, pp. 508–516. Elsevier Inc., Philadelphia (2019) 31. Saha, M., Sarkar, A.: Review on multiple facets of drug resistance: a rising challenge in the 21st century. Jorunal of Xenobiotics 11, 197–214 (2021)
Simulating the Ca2+ -cAMP Crosstalk and Its Role in Pancreatic Cells Hugo Enrique Romero-Campos1(B) , Geneviève Dupont2 and Virginia González-Vélez1
,
1 Department of Basic Sciences, Universidad Autónoma Metropolitana - Azcapotzalco,
02200 México City, México [email protected] 2 Unit of Theoretical Chronobiology, Université Libre de Bruxelles, 1050 Brussels, Belgium
Abstract. Ca2+ and cAMP are the most extended intracellular second messengers playing as transducers of physiological activity in the form of oscillations. In pancreatic α- and β-cells they are directly related to glucagon and insulin secretion, respectively, which are the main hormones involved in blood glucose regulation. The crosstalk between these messengers is very complex and of great relevance to infer abnormalities in both hormones’ secretion. In this work, we simulate the interaction between the Ca2+ and the cAMP pathways when considering either positive or negative effects of one pathway on the other one. As long as Ca2+ inhibits cAMP synthesis, out-of-phase oscillations were reached, regardless of positive or negative effect of cAMP on Ca2+ dynamics. In contrast, when Ca2+ enhances cAMP synthesis, oscillations were always in phase. We thus conclude that Ca2+ effect on cAMP pathway can switch these phase relationships, which agrees with recent experimental observations in pancreatic β-cells. Keywords: cAMP · Ca2+ · Oscillations · Islet cells · Glucose regulation
1 Introduction Ca2+ and cAMP are ubiquitous second messengers in mammalian cells, and they participate in many diverse intracellular vital processes such as secretion, growth, and apoptosis [1]. In pancreatic islets, it has been shown that glucose increases cAMP levels, and that this messenger can amplify insulin and glucagon secretion. cAMP signaling is primarily modulated by the Gs -Coupled Protein Receptor (Gs ) which activates the enzyme converting ATP into cAMP, the Adenylyl Cyclase (AC), as well as its use in other pathways [2]. On the other hand, insulin and glucagon release are both Ca2+ -dependent processes [3]. Ca2+ signaling within pancreatic cells is primarily determined by Ca2+ entrance through Voltage Dependent Calcium Channels (VDCC), Ca2+ transport and buffering, and Ca2+ release from the Endoplasmic Reticulum (ER) [4].
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 196–203, 2023. https://doi.org/10.1007/978-3-031-18256-3_21
Simulating the Ca2+ -cAMP Crosstalk and Its Role in Pancreatic Cells
197
Both messengers interact in many ways inside pancreatic cells where they are involved in the intrinsic response to glucose. Indeed, the rhythm of secretion is associated with Ca2+ and cAMP oscillations in pancreatic islets [5]. Then, the study of observed cellular responses may be explained only when considering their complex interactions. A first approximation of the Ca2+ -cAMP signaling was proposed considering two general forms of interaction [6]. One form, characterized by cAMP leading to cytosolic Ca2+ increase and then Ca2+ acting as cAMP synthesis inhibitor. Another form, described as cAMP leading to cytosolic Ca2+ decreases and then Ca2+ acting as cAMP synthesis activator. In that work, authors identified the first way of interaction in pancreatic β-cells and they proposed it to be the responsible for slow oscillations. They also found that the phase relationship between Ca2+ and cAMP oscillations relied on the number of intermediates linking them, and on their functional interactions. A more recent model describing the crosstalk between Ca2+ and cAMP signaling pathways was proposed in [7]. Siso-Nadal et al. model (SNM) [7] considers the effect of external stimuli over two type G-protein-coupled receptors (GPCR). The Gq type receptor corresponds to the beginning of the Ca2+ pathway. This leads to Ca2+ induced Ca2+ release (CICR) that relies on IP3 receptors (IP3 Rs). Ca2+ released from Endoplasmic Reticulum exerts two effects: stimulation of Phospholipase C enzyme (PLC), in positive feedback; and inhibition of IP3 Rs, in negative feedback. This is a well-established system able to produce intracellular Ca2+ oscillations in many cell types, including pancreatic β-cells [8]. It is worth noticing that a key component in this pathway is the protein kinase C enzyme (PKC). It is activated by Diacylglycerol (DAG) and Ca2+ and can phosphorylate serine and threonine residues of target proteins involved in many diverse signaling pathways [9]. On the other hand, the Gs type receptor begins the cAMP pathway. Once an agonist is linked to this receptor, it activates adenylyl cyclase enzyme (AC). AC converts ATP into cAMP, which in turn activates protein kinase A enzyme (PKA) [10]. And this kinase phosphorylates serine and threonine residues of several target proteins involved in regulation of ion channels, gene expression, metabolism, cytoskeleton dynamics and exocytotic processes [11]. In summary, SNM model considered that the communication between both pathways is through the action of the protein kinases C and A (PKC and PKA). PKC is capable of activating or inhibiting cAMP synthesis by directly influencing AC activity, while PKA sensitizes or desensitizes the protein lipase C (PLC) which indirectly determines Ca2+ release from ER. While not rigorous in quantitative terms, SNM model is useful to test the relative positive or negative interaction between the pathways, and to discuss the implications of the phase relationship of both oscillations in pancreatic cells. Therefore, in this work, we base on this model to discuss the possible interactions between Ca2+ and cAMP pathways, and how they could exist in pancreatic cells under physiological conditions and according to recent reported experimental oscillations.
2 Methods Ca2+ -cAMP system underlies on a great variety of interactions, however, SNM describes the crosstalk between Ca2+ and cAMP pathways in a simplified manner by considering the effect of PKC over AC, as well as the action of PKA over PLC. To do this,
198
H. E. Romero-Campos et al.
they proposed a homogeneous model conformed by 7 ordinary differential equations in the variables: IP3 , Cac2+ , DAG, PKC, cAMP, PKA and θR , where the first six are concentrations and θR is the fraction of active IP3 Rs (see Ref. [7]). Parameter values and equations are taken as in the original article [7], except cs and cq . Values taken for these two parameters are specified below. In SNM, the relative activity of each pathway is imposed by the values of the parameters PLC and AC. Increasing the values of these parameters results in higher Ca2+ and cAMP concentrations. Indeed, PLC acts as a bifurcation parameter to enter in and out from the oscillatory regime in the Ca2+ pathway (see Fig. 2 in [7]). The effect of Ca2+ pathway on cAMP is modeled as dependent on PKC. In the differential equation for cAMP (Eq. 9 from [7]): dcAMP PKC dPKA = vm · AC 1 + cq · − bm · cAMP − 2 · (1) dt Kq + PKC dt The parameter cq determine the enhancement (when positive) or inhibitory (when negative) effect of PKC on cAMP synthesis. On the other hand, the effect of cAMP pathway on Ca2+ dynamics is modeled by making the affinity constant of Ca2+ to PLC dependent on PKA (Eq. 8 from [7]): Kc1 =
∗ Kc1
+ cs
PKA Ks + PKA
4 (2)
The parameter cs determine the sensitizing (when negative) or desensitizing (when positive) of PLC. In this work, we extend the analysis reported in [7] by considering not just the mutual inhibition and mutual activation between both pathways, but the cases in which Ca2+ inhibits cAMP pathway while cAMP activates Ca2+ pathway and vice versa. These two latter cases are specially interesting since the interaction between both messengers is so complex that they could act as inhibitors or activators or even change over time [6]. The four cases in the present study with their corresponding cq and cs values are summarized in Table 1. The system of seven differential equations was numerically integrated in Mathematica 13.1.1.0 for Windows using the “StiffnessSwitching” method with “AccuracyGoal” and “PrecisionGoal” equal to 10 [12]. Temporal evolution of Ca2+ and cAMP are shown in normalized quantities aiming to highlight the qualitative properties of the system response for the cases under study. We focus on the phase relationship between Ca2+ and cAMP oscillations and discuss the effect of the relative activity of Ca2+ pathway (PLC parameter) for the four cases summarized in Table 1. All Ca2+ and cAMP dynamics in Results are stable oscillations reached after long simulation times, starting from the same initial conditions (AC = 0.5 and PLC = 0.1). Oscillatory dynamics is activated by changing the PLC value as indicated in each case, while AC value is kept equal to 0.5.
Simulating the Ca2+ -cAMP Crosstalk and Its Role in Pancreatic Cells
199
Table 1. Parameters values for the Ca2+ -cAMP interactions under study. Case
Parameter values
Description
i
cq = −1, cs = 1
Mutual inhibition
ii
cq = −1, cs = −1
Ca2+ inhibits cAMP and cAMP activates Ca2+
iii
cq = 1, cs = 1
Ca2+ activates cAMP and cAMP inhibits Ca2+
iv
cq = 1, cs = −1
Mutual activation
3 Results We started simulating the inhibitory effect of Ca2+ over cAMP (cases i and ii in Table 1). As can be seen in Fig. 1, when Ca2+ has an inhibitory effect over cAMP synthesis, a change in the cAMP effect over PLC from inhibitory (Figs. 1A and 1B) to activating (Figs. 1C and 1D) does not modify the oscillatory phase relationship between both oscillations. In both cases, Ca2+ increases as cAMP decreases (out-of-phase response) and PLC has an effect on the oscillatory frequency. For cAMP inhibiting the Ca2+ pathway, the frequency can change from 0.01 Hz (Fig. 1A) to 0.035 Hz (Fig. 1B) for PLC values of 1 and 1.8, respectively, whereas for cAMP activating Ca2+ , the frequency can change from 0.04 Hz (Fig. 1C) to 0.079 Hz (Fig. 1D) for PLC values of 0.3 and 0.8, respectively.
Fig. 1. Out-of-phase response of Ca2+ -cAMP system when Ca2+ inhibiting cAMP pathway (cq = −1). Panels A and B correspond to Case i (cs = 1). Panels C and D correspond to Case ii (cs = −1). Values of PLC are 1, 1.8, 0.3 and 0.8, for A, B, C and D.
200
H. E. Romero-Campos et al.
If Ca2+ has an activating effect over cAMP synthesis (Fig. 2), the phase relation is preserved either cAMP inhibits (Figs. 2A and 2B) or activates (Figs. 2C and 2D) the Ca2+ pathway. However, in these two cases the response is in-phase, which means that as Ca2+ increases, cAMP does so. Regarding oscillatory frequency, as before, when PLC increases the frequency increases too. When cAMP inhibits Ca2+ , the frequency varies from 0.008 Hz (Fig. 2A) to 0.098 Hz (Fig. 2B) for PLC values of 1 and 3.2, respectively, whereas if cAMP activates Ca2+ , frequency goes from 0.019 Hz (Fig. 2C) to 0.077 Hz (Fig. 2D) for PLC values of 0.14 and 0.21, respectively. It is worth noticing that mutual activation is the only case leading to complex Ca2+ oscillations. A very interesting finding by analyzing Ca2+ effect on cAMP pathway is the possibility of switching the phase response. When cAMP acts either as inhibitor (cs = 1) or as activator (cs = −1), a change in the effect of Ca2+ , from inhibitor to activator, switches the phase response from out-of-phase to in-phase (Fig. 3).
Fig. 2. In-phase response of Ca2+ -cAMP system when Ca2+ activating cAMP pathway (cq = 1). Panels A and B correspond to Case iii (cs = 1). Panels C and D correspond to Case iv (cs = −1). Values of PLC parameter are 1, 3.2, 0.14 and 0.21, for A, B, C and D.
Simulating the Ca2+ -cAMP Crosstalk and Its Role in Pancreatic Cells
201
Fig. 3. Switch of phase between Ca2+ and cAMP oscillations due to a change in Ca2+ effect on cAMP pathway. An out-of-phase response is developed during the first 5 min of simulation when Ca2+ has an inhibitory effect on cAMP pathway (cq = −1). An in-phase response is observed when Ca2+ activates cAMP pathway (cq = 1). PLC = 1.0.
4 Discussion It is well known that oscillations of intracellular messengers such as Ca2+ and cAMP are tightly related to the cellular response to physiological stimuli [1]. Therefore, a deep understanding of the mechanisms and biochemical pathways behind oscillatory behavior of both messengers is crucial. Here we extend the analysis of a previous model of Ca2+ cAMP interaction [7] adding the possibility that feedback could be combined (positivenegative and negative-positive). As discussed some time ago [6], these combinations are physiologically meaningful and could be appearing in some particular cells [1, 13]. The observed out-of-phase response in our simulations, when Ca2+ has an inhibitory effect over cAMP pathway, fully agrees with previous modeling reports in Aplysia R15 neuron [14] and in pancreatic β-cells [15]. In these works, that behavior is explained either by a decreased activity of AC (i.e., cAMP synthesis) or by an increased activity of PDE (i.e., cAMP degradation). Moreover, the ability of the Ca2+ pathway to switch from one type of interaction to another (inhibition versus activation) could explain the appearance of completely different oscillations (out-of-phase versus in-phase) in real conditions. This kind of phase switching has been observed in cells with abnormal activity of the AC enzyme [16]. In particular, in pancreatic β-cells, cAMP oscillations could be in-phase or out-of-phase with Ca2+ oscillations depending on the intracellular region where they are measured, that is, the submembrane compartment or the cytosol [16]. Frequency of oscillations is also a key feature for cellular responses. It has been discussed that efficiency of these messengers depends on their dynamics, having local impact for short pulses or widespread impact for long-lasting oscillations. For example, in β-cells, frequency of cAMP oscillations is fully linked to pulsatile insulin secretion, so abnormalities in the cAMP pathway are associated with glucose regulation impairment [17]. Reported values for cAMP oscillations in β-cells are between 0.003 − 0.016 Hz [18, 19], which completely agree with our simulations. Based on our modeling approach, frequencies are modulated by PLC, which is a direct measure of the Ca2+ pathway activity.
202
H. E. Romero-Campos et al.
5 Conclusions In this work we studied the crosstalk between Ca2+ and cAMP pathways through dynamic simulations. We analyze four ways of interactions between both pathways, that is, the possible combinations of activating or inactivating effects from one on the other. Based on our results, changes of cAMP effect on Ca2+ pathway strongly modifies the frequency of both messengers’ oscillations, while the phase relationship persists. In contrast, changes of Ca2+ effect on cAMP pathway can switch the phase relationship, which agrees with recent experimental observations in pancreatic β-cells. As insulin and glucagon secretion rely on Ca2+ and cAMP dynamics, deep comprehension of their crosstalk is key to understand the dysregulations leading to abnormal secretion in pathological conditions. Acknowledgment. H.E.R.C. thanks CONACyT for his doctoral scholarship.
References 1. Sanchez-Collado, J., Lopez, J.J., Jardin, I., Salido, G.M., Rosado, J.A.: Cross-Talk Between the Adenylyl Cyclase/cAMP Pathway and Ca2+ Homeostasis. In: Pedersen, S.H.F. (ed.) Reviews of Physiology, Biochemistry and Pharmacology. RPBP, vol. 179, pp. 73–116. Springer, Cham (2020). https://doi.org/10.1007/112_2020_55 2. Stožer, A., et al.: The role of camp in beta cell stimulus–secretion and intercellular coupling. Cells 10(7), 1–26 (2021). https://doi.org/10.3390/cells10071658 3. González-Vélez, V., Dupont, G., Gil, A., González, A., Quesada, I.: Model for glucagon secretion by pancreatic α-cells. PLoS ONE 7(3), e32282 (2012). https://doi.org/10.1371/jou rnal.pone.0032282 4. Dupont, G., Falcke, M., Kirk, V., Sneyd, J.: Models of Calcium Signalling, 1st ed. Springer International Publishing, Switzerland (2016) 5. Tengholm, A.: Cyclic AMP dynamics in the pancreatic β-cell. Ups. J. Med. Sci. 117(4), 355–369 (2012). https://doi.org/10.3109/03009734.2012.724732 6. Rapp, P.E., Berridge, M.J.: Oscillations in calcium-cyclic AMP control loops form the basis of pacemaker activity and other high frequency biological rhythms. J. Theor. Biol. 66(3), 497–525 (1977). https://doi.org/10.1016/0022-5193(77)90299-5 7. Siso-Nadal, F., Fox, J.J., Laporte, S.A., Hébert, T.E., Swain, P.S.: Cross-talk between signaling pathways can generate robust oscillations in calcium and cAMP. PLoS ONE 7(10), e7189 (2009). https://doi.org/10.1371/journal.pone.0007189 8. Félix-Martínez, G.J., Godínez-Fernández, J.R.: Mathematical models of electrical activity of the pancreatic β-cell: A physiological review. Islets 6(3), e949195 (2014). https://doi.org/10. 4161/19382014.2014.949195 9. Webb, B.L.J., Hirst, S.J., Giembycz, M.A.: Protein kinase C isoenzymes: A review of their structure, regulation and role in regulating airways smooth muscle tone and mitogenesis. Br. J. Pharmacol. 130(7), 1433–1452 (2000). https://doi.org/10.1038/sj.bjp.0703452 10. Morrison, H.: Protein kinase A. Enzym. Act. Sites their React. Mech. 173–178 (2021 Jan). https://doi.org/10.1016/B978-0-12-821067-3.00029-5 11. Crul, T., Maléth, J.: Endoplasmic reticulum-plasma membrane contact sites as an organizing principle for compartmentalized calcium and camp signaling. Int. J. Mol. Sci. 22(9) (2021). https://doi.org/10.3390/ijms22094703 12. Trott, M.: The mathematica guidebook for numerics. Springer Science+Business Media Inc, Illinois (2006)
Simulating the Ca2+ -cAMP Crosstalk and Its Role in Pancreatic Cells
203
13. Ahuja, M., Chung, W.Y., Lin, W.Y., McNally, B.A., Muallem, S.: Ca2+ signaling in exocrine cells. Cold Spring Harb. Perspect. Biol. 12(5) (2020). https://doi.org/10.1101/cshperspect. a035279 14. Yu, X., Byrne, J.H., Baxter, D.A.: Modeling Interactions between electrical activity and second-messenger cascades in aplysia neuron R15. J. Neurophysiol. 91(5), 2297–2311 (2004). https://doi.org/10.1152/jn.00787.2003 15. Fridlyand, L.E., Harbeck, M.C., Roe, M.W., Philipson, L.H.: Regulation of cAMP dynamics by Ca2+ and G protein-coupled receptors in the pancreatic β-cell: A computational approach. Am. J. Physiol. - Cell Physiol. 293, C1924–C1933 (2007). https://doi.org/10.1152/ajpcell. 00555.2006 16. Tenner, B., et al.: Spatially compartmentalized phase regulation in the Ca2+ -cAMP-PKAoscillatory circuit. Biophys. J. 118(3), 175a (2020). https://doi.org/10.7554/eLife.55013. Feb. 17. Tengholm, A.A., Gylfe, E.: cAMP signalling in insulin and glucagon secretion. Diabetes, Obes. Metab. 19(S1), 42–53 (2017). https://doi.org/10.1111/dom.12993 18. Tian, G., Sandler, S., Gylfe, E., Tengholm, A.: Glucose- and hormone-induced cAMP oscillations in α- and β-cells within intact pancreatic islets. Diabetes 60, 1535–1543 (2011). https:// doi.org/10.2337/db10-1087 19. Landa, L.E., et al.: Interplay of Ca2+ and cAMP signaling in the insulin-secreting MIN6 β-cell line. J. Biol. Chem. 280(35), 31294–31302 (2005). https://doi.org/10.1074/jbc.M505657200
Simulating the Loss of β-cell Mass in a Human Pancreatic Islet: Structural and Functional Implications Sergio Ruiz-Santiago1 , Jos´e Rafael God´ınez-Fern´andez1 , and Gerardo Jorge F´elix-Mart´ınez1,2(B) 1
Departmento de Ingenier´ıa El´ectrica, Universidad Aut´ onoma Metropolitana, Iztapalapa, M´exico City 09310, M´exico 2 Consejo Nacional de Ciencia y Tecnolog´ıa (CONACYT), Investigadoras e Investigadores por M´exico, Benito Ju´ arez, 03940 M´exico City, M´exico [email protected]
Abstract. Type II diabetes (T2D) is a disorder defined by an impaired insulin secretion and insulin resistance. Throughout the progression of the disease, β-cells are lost due to the high-demanding environment produced by a prolonged hyperglycemic state. Moreover, T2D has been associated to structural alterations of the architecture of pancreatic islets, which might interfere negatively in the gap-junctional communication between β-cells. In this work, aiming to evaluate the effects of the loss of β-cell mass on connectivity metrics and on the synchronization of the cells’ electrical signals, we performed computational simulations of the network formed by β-cells in a human islet. Our results indicate that the loss of 15 and 30% of β-cells of a human islet would have a negative impact on the connectivity, integration and efficiency of the network, with a marked negative effect on the overall synchronization of the electrical signals of the islet β-cells. Keywords: Human pancreatic islets Connectivity
1
· Electrical activity · Network ·
Introduction
Insulin, a hormone secreted by pancreatic β-cells, is the only hormone responsible for lowering blood glucose levels by promoting the uptake of glucose by muscle, adipose cells and liver. β-cells, located in the pancreatic islets, secrete insulin in response to glucose stimulation by producing an electrical pattern that leads to the exocytosis of insulin following an increase in intracellular calcium [13]. In the islets, neighbor β-cells can be electrically communicated by gapjunctions, channels formed between two β-cells, that allow the exchange of ions, messengers and other molecules [15]. The electrical communication between βcells has been functionally associated to a synchronized, more efficient response to a glucose challenge [14]. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 204–211, 2023. https://doi.org/10.1007/978-3-031-18256-3_22
Simulating the Loss of β-cells in a Human Pancreatic Islet
205
Type 2 diabetes (T2D) is a metabolic disease characterized by a prolonged hyperglycemic state along with a deficient response to the insulin signal by the target tissues [17]. In addition, it has been shown that, at advances stages, T2D can be accompanied by a certain degree of β-cell loss [16], although the structural and functional impacts are still a matter of intensive research. As a complement to the experimental work, computational models have been used in the last decades to study the electrical and secretory properties of the pancreatic β-cells [5]. More recently, models have evolved into multicellular models of pancreatic islets with the main objective of elucidating the functional role of intraislet communication between islet cells (see for instance refs. [6–8]). Other models have aimed to gain a better understanding of the functional implications of the loss of β-cells [2] and the loss of coordination [3] between the α, β and δ-cells. In addition, network science has proven to be extremely useful to dissect the connectivity properties of the pancreatic islets and the functional and structural networks formed by the islet cells (reviewed recently in [4]). Based on these previous works, here we use computational modeling to evaluate the structural and functional effects of the loss of β-cell mass. To achieve this goal, we firstly reconstructed a human pancreatic islet to obtain the size, location and connections between β-cells within the islet. Then, based on the contacts between β-cells identified, we generated the corresponding network to quantitatively evaluate the impact on the connectivity properties of the network of β-cells. Finally, we performed simulations of the electrical behavior of the islets’ β-cells to associate the structural changes of the islet with the corresponding functional response.
2 2.1
Methods Reconstruction of Pancreatic Islets
The human islet analyzed in this work was reconstructed in IsletLab [1] based on the experimental data provided by Hoang et al. [2]. In short, the reconstruction algorithm used, described in detail in [20], uses the cells’ nuclei positions along with the identity of each cell (α, β or δ) obtained experimentally, in an iterative optimization algorithm to determine the optimal size and position of each cell in the islet until an islet composed of non-overlapped cells is obtained. As a result, the reconstruction algorithm provides with the optimized size and coordinates of each cell in the islet, along with the connectivity properties obtained from the analysis of the final reconstructed islet. Since our objective is to analyze the connectivity between β cells within the islet, only the β-cells of the reconstructed islet were further considered. The finalized reconstructed architecture and the contacts between β-cells identified are shown in Fig. 1A (left and middle panels, respectively). Note that only one control islet was used for this work.
206
2.2
S. Ruiz-Santiago et al.
Loss of β-cells
The loss of β-cell mass was simulated by removing 15 and 30% of cells from the control islet by using a random number generator to select the cells to be removed. Five replicates were generated for each percentage of β-cells removed. As a result, the cell-to-cell contacts associated to the removed cells were also removed. The percentages of β-cells lost were selected to simulate the early stages of the disease, given that greater percentages between 40 and 60% of β-cells can be lost at more advanced stages [9]. Visualization of the islet architectures with 15 and 30% loss of β-cells, and the corresponding contacts between cells, are shown in the left and middle panels of Fig. 1 (B and C). Note that the results associated to the perturbed islets are presented as mean ± SD throughout the article. 2.3
Islet Connectivity Networks
Islet networks were generated from the connectivity matrices for both the control architecture and the altered islets (15 and 30% β-cell loss) in Wolfram Mathematica 13 [10]. Network metrics (density, efficiency, average degree, clustering coefficient, diameter) were calculated for the three architectures analyzed. As described in a previous work [11], average degree is the average number of contacts of each cell in the islet; density reflects the network’s connectedness; the average clustering coefficient measures the interconnection of each cell with its neighbors; global efficiency quantitatively describes the integration properties of the network and the diameter is an indicator of the size of the network. 2.4
Modelling the Electrical Behavior of the Human β-cell
The electrical behavior of each β-cell during glucose stimulation (i.e. firing of action potentials) included in the islet was modelled using an adapted version of a previous model [12] in which the changes in the cell membrane potential are calculated by solving the following differential equation: 1 dVmi =− ILi + IT i + IP Qi +IN ai + IKvi + IBKi dt Cmi + IKAT P i + IERGi + ILeaki +
Iij
(1)
j
where i is the cell number, IL , IT and IP Q are the Ca2+ current produced by the L, T and P/Q-type Ca2+ channels, respectively; IN a is the Na+ current, IKv , IBK , IKAT P and IERG are K + currents through the voltage dependent, Ca2+ activated, ATP-dependent and ether-a-go-go channels, respectively, and ILeak is a leak current. j Iij represent the coupling currents between the i cell and the j neighbor cells to which it is connected via gap-junctions. All the currents in Eq. 1 were modelled using a Hodgkin-Huxley type model, described in general as:
Simulating the Loss of β-cells in a Human Pancreatic Islet
207
Ixi = gx mx hx (Vmi − Vr )
(2)
where x indicates the type of current (L, T, P/Q, Na, etc.), mx and hx are the activation and inactivation functions, Vm is the cell membrane potential and Vr is the reversal potential of the corresponding ion type (r = Ca2+ , Na+ , K+ ). Coupling currents (Iij ) were modelled as: Iij = gij (Vmi − Vmj ),
(3)
where gij is the coupling conductance between the connected β-cells and Vmi and Vmj are the membrane potentials of cells i and j, respectively. The interested reader can consult the details of the model of electrical activity, including additional equations and parameters, in the original work on the subject [12]. The coupling conductance and the initial value of the cells’ membrane potentials were assigned randomly to include a certain degree of heterogeneity to the model. Parameters used in the simulations of the electrical activity of the islet β-cells are shown in Table 1. Table 1. Parameters used for the simulation of the electrical behavior of the islet β-cells. a Capacitance values were obtained from the islet reconstruction process. b Random values generated from a uniform distribution in the indicated range. Parameter Value
Parameter Value
Parameter Value
Parameter Value
gP Q
0.73 nS gERG
1.98 nS gKv
9.9 nS
Cm
Variablea
gT
1.29 nS gKAT P
0.15 nS gN a
0.4 nS
gij
0.1–0.3 nS
gL
1.32 nS gKCa
1.49 nS gLeak
0.11 nS Vm (t = 0) –75–65 mV
2.5
b b
Computational Aspects
Islet reconstruction was performed in IsletLab [1] in a PC with 16 Intel Core i7 processors (3.8 GHz) and 64 GB RAM memory running Ubuntu 18.04. Functional simulations (electrical activity), postprocessing, network analysis and visualizations were performed in Wolfram Mathematica 13 [10].
3
Results
As shown in Table 2, the control islet was composed by 257 β-cells with an overall count of contacts between β-cells of 217. The loss of 15 and 30% of βcells represented a loss of 38 and 77 cells from the islet, thus reducing the number of cell-to-cell contacts by 62 (24%) and 114 (53%), respectively. The effects of the loss of β-cell mass (and the corresponding contacts between β-cells) on the properties of the networks formed by the islet β-cells are shown graphically in the right column of Fig. 1 and described numerically in Table 3. Firstly, the number of components (or subnetworks) increased from 58 for the
208
S. Ruiz-Santiago et al.
Fig. 1. Islet architectures (left column), cell-to-cell contacts (middle column) and connectivity networks (right column) analyzed. Top row: Control case (100% β-cells). Middle row: 15% β-cell loss. Bottom row: 30% β-cell loss.
control islet to 73 and 79 for the cases of 15 and 30% loss, respectively. In practice, this means that the loss of contacts between β-cells produced a greater number of cell groups, which at the end resulted in a more segregated islet. In spite of this, the density of the networks only decreased marginally, which can be explained by the fact that the actual number of links in the islet network is considerably lower than the total number of possible links (as if all the cells were connected to each other). In contrast, the clustering coefficient decreased from 0.117 (control) to 0.107 (–8.5%) and 0.059 (–49.6%) for the 15 and 30% loss cases, respectively. Similarly, the efficiency of the networks decreased from 0.91 for the control case to 0.84 (–7.7%) and 0.77 (–15.4%) for the 15 and 30% loss cases, respectively. On the other hand, the size of the network, represented by the diameter of the network, decreased from 33 (control) to 18.5 (–44%) for the 15% loss case and 10.8 (–67.3%) for the 30% loss case. In summary, these results indicate that 1) each cell is less connected to its neighborhood, 2) the size
Simulating the Loss of β-cells in a Human Pancreatic Islet
209
Table 2. Characteristics of the simulated islets. Islet
Number of cells Contacts between β-cells
Control
257
217
15% loss 219
155 ± 4.3
30% loss 180
103.4 ± 4.8
Table 3. Network metrics for the islets simulated. Islet
Components Density
Clustering
Efficiency
Diameter
Control
58
0.117
0.91
33
0.0066
15% loss 73.3 ± 5.1
0.0065 ± 0.00018 0.107 ± 0.022 0.84 ± 0.02 18.5 ± 6.2
30% loss 79.4 ± 4.4
0.0064 ± 0.00029 0.059 ± 0.044 0.77 ± 0.04 10.8 ± 1.5
of the networks decreased considerably and, as a consequence, 3) the integration of the network was negatively impacted by the loss of β-cells. In order to evaluate the effects of the changes of the islet architecture and the connectivity properties, functional simulations of the electrical behavior of the islet β-cells were performed. The electrical signals of all the cells are shown in Fig. 2 (left column). A correlation analysis was conducted to quantitatively determine the impact of the loss of β-cell mass, and the corresponding loss of connection between β-cells (as indicated by the network metrics described above) on the synchronization of β-cells within the islet. Graphically, the correlation matrices are shown in the right column of Fig. 2. Quantitatively, the control islet had 235 highly correlated electrical signals (>0.9), while the loss of 15 and 30% β-cells reduced this number to 221 (–6%) and 138 (–41.3%), respectively, as described in detail in Table 4. Table 4. Highly correlated electrical signals in the simulated islets. Islet
Correlation (>0.9)
Control
235
15% loss 220.5 ± 43.3 30% loss 138.6 ± 26.9
Previous experimental works have been mostly based on functional networks derived from correlation analyses of Ca2+ signals (e.g. references [4,18,19]). While these works have demonstrated the usefulness of network science for the study of pancreatic islets, they have not been able yet to consider the signals generated by all the β-cells in the islet, thus restricting the analyses to a small proportion of visible cells. For this reason, here we have used a full reconstructed
210
S. Ruiz-Santiago et al.
Fig. 2. Left column: electrical signals of the islet β-cells. Note that the changes in membrane potential are presented as a change in color. Right column: Correlation matrices of the electrical signals of the islet β-cells. Top: control. Middle: 15% β-cell loss. Bottom: 30% β-cell loss.
islet to analyze the network formed by the β-cells in close contact and to simulate the electrical activity of all the cells simultaneously, which allowed us to analyze the relation between the functional and structural properties of the islet β-cell population.
4
Conclusions
In this work, we analyzed the impact of the loss of β-cell mass on the connectivity properties of a human pancreatic islet and evaluated the functional impact via computational simulations of the electrical pattern of the islet β-cells. According to our results, even the loss of a relatively small percentage of β-cells could have a negative impact on the communication of each cell with its neighbors. In addition, the loss of links between β-cells also produced a more segregated network, with the overall effect of a decrease on both the size and the integration of the network. From the functional viewpoint, the number of highly correlated electrical signals also decreased considerably. In conclusion, according to our results, the loss of β-cell mass commonly observed during early stages of type 2 diabetes could lead to a decreased synchronization of the electrical response of the β-cells, which could lead to an impaired and suboptimal insulin secretory response.
Simulating the Loss of β-cells in a Human Pancreatic Islet
211
References 1. F´elix-Mart´ınez, G.: IsletLab: an application to reconstruct and analyze islet architectures. Islets 14, 36–39 (2022) 2. Hoang, D., Hara, M., Jo, J.: Design principles of pancreatic islets: glucosedependent coordination of hormone pulses. PloS One 11, e0152446 (2016) 3. Hogan, J., Peercy, B.: Flipping the switch on the hub cell: Islet desynchronization through cell silencing. PloS One 16, e0248974 (2021) 4. Stoˇzer, A., et al.: From Isles of K¨ onigsberg to islets of langerhans: examining the function of the endocrine pancreas through network science. Front. Endocrinol. 13 (2022) 5. F´elix-Mart´ınez, G., God´ınez-Fern´ andez, J.: Mathematical models of electrical activity of the pancreatic β-cell: a physiological review. Islets 6, e949195 (2014) 6. Dwulet, J., Briggs, J., Benninger, R.: Small subpopulations of β-cells do not drive islet oscillatory [Ca2+] dynamics via gap junction communication. PLoS Comput. Biol. 17, e1008948 (2021) 7. Saadati, M., Jamali, Y.: The effects of beta-cell mass and function, intercellular coupling, and islet synchrony on Ca2+ dynamics. Sci. Rep. 11, 1–14 (2021) 8. Lei, C., Kellard, J., Hara, M., Johnson, J., Rodriguez, B., Briant, L.: Beta-cell hubs maintain Ca2+ oscillations in human and mouse islet simulations. Islets 10, 151–167 (2018) 9. Weir, G., Gaglia, J., Bonner-Weir, S.: Inadequate β-cell mass is essential for the pathogenesis of type 2 diabetes. Lancet Diab. Endocrinol. 8, 249–256 (2020) 10. Wolfram Mathematica, Version 13.0.0, Champaign, IL (2021). https://www. wolfram.com/mathematica 11. F´elix-Mart´ınez, G., God´ınez-Fern´ andez, J.: Comparative analysis of reconstructed architectures from mice and human islets. Islets 14, 23–35 (2022) 12. F´elix-Mart´ınez, G., God´ınez-Fern´ andez, J.: Modeling the spatiotemporal distribution of Ca2+ during action potential firing in human pancreatic β-cells. Biomed. Phys. Eng. Exp. 3, 025020 (2017) 13. Rorsman, P., Ashcroft, F.: Pancreatic β-cell electrical activity and insulin secretion: of mice and men. Physiol. Rev. 98, 117–214 (2018) 14. Farnsworth, N., Benninger, R.: New insights into the role of connexins in pancreatic islet function and diabetes. FEBS Lett. 588, 1278–1287 (2014) 15. Moreno, A., Berthoud, V., P´erez-Palacios, G., P´erez-Armendariz, E.: Biophysical evidence that connexin-36 forms functional gap junction channels between pancreatic mouse β-cells. Am. J. Physiol.-Endocrinol. Metab. 288, E948–E956 (2005) 16. Chen, C., Cohrs, C., Stertmann, J., Bozsak, R., Speier, S.: Human beta cell mass and function in diabetes: recent advances in knowledge and technologies to understand disease pathogenesis. Molec. Metab. 6, 943–957 (2017) 17. Kalin, M., Goncalves, M., John-Kalarickal, J., Fonseca, V.: Pathogenesis of type 2 diabetes mellitus. In: Principles Of Diabetes Mellitus, pp. 1–11. Springer, Heidelberg (2017). https://doi.org/10.1007/978-3-319-20797-1 13-2 18. Stoˇzer, A., et al.: Functional connectivity in islets of Langerhans from mouse pancreas tissue slices. PLoS Comput. Biol. 9, e1002923 (2013) 19. Salem, V., et al.: Leader β-cells coordinate Ca2+ dynamics across pancreatic islets in vivo. Nat. Metab. 1, 615–629 (2019) 20. F´elix-Mart´ınez, G., Mata, A.N., God´ınez-Fern´ andez, J.: Reconstructing human pancreatic islet architectures using computational optimization. Islets 12, 121–133 (2020)
Role of Endogenous Ca2+ Buffering and the Readily Releasable Pool on Fast Secretion in Auditory Inner Hair Cells Crystal Azucena Valverde-Alonzo1 , Gerardo Jorge F´elix-Mart´ınez1,3 , Virginia Gonz´ alez-Velez2(B) , and Amparo Gil4 1
3
Departmento de Ingenier´ıa El´ectrica, Universidad Aut´ onoma Metropolitana, Iztapalapa, 09310 M´exico City, Mexico 2 Departamento de Ciencias B´ asicas, Universidad Aut´ onoma Metropolitana, Azcapotzalco, 02200 M´exico City, Mexico [email protected] Consejo Nacional de Ciencia y Tecnolog´ıa, Investigadoras e Investigadores por M´exico, Benito Ju´ arez, 03940 M´exico City, Mexico 4 Departamento de Matem´ atica Aplicada y Ciencias de la Computaci´ on, Universidad de Cantabria, 39005 Santander, Spain
Abstract. Sensory hair cells, located at the cochlea, convert sound into a depolarizing stimulus that, as a response to an increase in the intracellular Ca2+ concentration, triggers the release of glutamate. Experimental observations have shown that depending on their location in the inner ear, sensory hair cells respond differently to sounds of different frequencies. The origin of this behavior is still a matter of debate but, given the importance of the dynamics of intracellular Ca2+ for the exocytotic response, it has been hypothesized that the availability of endogenous Ca2+ buffers at the active zone, along with the size of the readily releasable pool of glutamate vesicles, could be associated to the frequency-dependent tuning of the exocytotic response. Here, we implemented a computational model of the active zone of a sensory hair cell with the main objective of evaluating the effects of the endogenous Ca2+ buffers and the readily releasable pool on the fast exocytosis of glutamate. Keywords: Sensory hair cells Calcium buffers
1
· Secretion · Computational model ·
Introduction
Transduction of sound waves to electrical signals starts at the sensory hair cells of the inner ear, where glutamate is secreted into the synaptic cleft from the presynaptic active zones as a result of the opening of voltage-dependent Ca2+ channels (VDCC) and the corresponding increase in intracellular Ca2+ . Glutamate then activate the postsynaptic auditory fibers, finally enabling the brain to transmit sound in the form of electrical impulses. Since damage to inner hair c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 212–218, 2023. https://doi.org/10.1007/978-3-031-18256-3_23
Role of Ca2 + Buffers and RPF on Fast Secretion
213
cells, produced either by environmental or genetic factors, has been linked to hearing loss [7], it is extremely important to gain a better understanding of the cellular mechanisms involved, in order to identify plausible therapeutic agents capable of restoring or maintaining the auditory function. An active zone is then the area where VDCC and primed vesicles are organized to release glutamate in a well-controlled dynamics such that sound information could be preserved. On average, a sensory hair cell have between 10 and 30 active zones [5] (schematized in Fig. 1), each containing about 90 Ca2+ channels and 32 glutamate vesicles [6]. Functionally, vesicles in the active zone are tightly attached to the cell membrane forming the so called readily releasable pool (RRP), from where they can be released in a Ca2+ -dependent manner when needed. Inside the RRP there is a small group of vesicles located so close to VDCC that Ca2+ entering through channels may immediately fuse them to release the glutamate; they are known as the Immediately Releasable Pool (IRP). In the last decades, computational models have been widely used to study secretion in a great variety of cell types, including the inner hair cells [11,12]. However, previous models of inner hair cells have focused on studying different aspects of the mechanical to electrical and neural transduction processes using computational models based on ordinary differential equations, thus neglecting important aspects such as the spatio-temporal distribution of Ca2+ in the active secretion zones, the effects of Ca2+ handling mechanisms (e.g. Ca2+ buffers) or a detailed description of the secretion process. While other works have addressed the relation between Ca2+ influx, exogenous and endogenous Ca2+ buffering, the coupling distance between Ca2+ ionic channels and exocytosis sensor in presynaptic terminals [13,14], to our knowledge, detailed simulations of the Ca2+ dependent secretion process in inner hair cells have not yet been performed. Given the relevance of Ca2+ for the secretory response of the sensory hair cells, it is reasonable to hypothesize that fast secretion heavily relies on the availability of endogenous Ca2+ buffering and on the size and spatial organization of the readily releasable pool within the active zones. In this work, we studied the effects of endogenous Ca2+ buffering and the geometric distribution of the RRP on the fast secretion of glutamate from sensory hair cells. We used a threedimensional computational model of an active zone, explicitly considering the number and distribution of VDCC and glutamate vesicles.
2
Methods
Secretion of glutamate vesicles and the buffered diffusion of Ca2+ was modeled following the Monte Carlo algorithm described in [1,2]. The main difference with standard differential methods is that our algorithm uses the number of ions and buffers as fundamental variables instead of concentrations (ours is a microscopic computational scheme). The active zone of a sensory hair cell was simulated as a cylindrical domain of radius 0.3 µm and 0.5 µm height, assuming that the base of the cylinder is the cell membrane where ion channels are located. The cylindrical domain was divided into cubic compartments of volume (Δl)3 ,
214
C. A. Valverde-Alonzo et al.
with a spatial resolution Δl = 20 nm. As shown schematically in Fig. 1, the model includes voltage-dependent Ca2+ channels, endogenous Ca2+ buffers and the secretion of glutamate vesicles. Ninety Ca2+ channels were distributed randomly throughout the cylinder base. Thirty-two RRP vesicles were distributed throughout the active zone either randomly (case 1: RRP with 0% of IRP) or locating 10% of the 32 vesicles (3 vesicles) co-localized to the VDCC (case 2: RRP with a 10% of IRP). This percentage of vesicles co-localized to the VDCC was estimated from extensive preliminary simulations in which a wide range of percentages was evaluated (not shown). The results obtained indicated that 10% was the percentage of co-localized vesicles that allowed us to better reproduce the experimental observations. The gating of the Ca2+ channels is described by the following kinetic scheme: O
k1 k−1
C
k2 k−2
I.
(1)
where O, C, I refer to the Open, Closed and Inactive states assumed for the voltage activation of each VDCC. Only the Open state can introduce Ca2+ into the active zone. Ca2+ buffering was simulated as a first order kinetic reaction: B + Ca2+
k+
BCa2+ ,
k−
(2)
where B is the Buffer molecule, and BCa2+ represent a buffer molecule bound to a Ca2+ ion. Glutamate secretion was simulated by the following kinetic scheme: B0
5ω 1k− b0
B1
4ω 2k− b1
B2 B5
3ω 3k− b2 γ
B3
2ω 4k− b3
B4
1ω 5k− b4
B5 (3) (4)
F
where Bi is the stage of the glutamate vesicle with i Ca bounded ions, ω = k+ Ca2+ , k+ and k− represent the forward and backward binding rates, respectively, b is the cooperativity factor, γ is the secretion rate and F represents the vesicle fusion state. A common experimental method to study fast secretion dynamics in different neuron-like cells, such as hair cells, is to stimulate them with voltage pulses in a voltage-clamp configuration. These pulses are intended to depolarize the cell in a similar way as physiological stimuli do, in order to measure vesicle fusions. It is also a common practice to variate the duration and the amplitude of the pulses to test the response of the cell. In accordance to experimental protocols, we performed simulations of fast secretion due to 5, 10 and 20 ms voltage pulses (five replicates per case). As a result, the concentration of Ca2+ at the active zone, free and bound Ca2+ buffers (exogenous and endogenous) and the cumulative secretion, are obtained. At each 2+
Role of Ca2 + Buffers and RPF on Fast Secretion
215
Fig. 1. Scheme of the active zone of a sensory hair cell
time step of the simulation, the state of the Ca2+ ion channels (open, closed or inactivated) is determined stochastically. When a channel opens, the algorithm places an extra number of Ca2+ ions (given by the channel current) below the cell membrane, at the corresponding position where the open channel is located. Then, Ca2+ ions and mobile buffer molecules move from one compartment of the grid to another compartment due to diffusion, which is modeled as a random walk process. The timescale for the random walk is given in terms of the spatial resolution and the coefficient diffusion of Ca2+ (Δt = (Δl)2 /4DCa2+ ) and this is used as the time step of the simulation. During each diffusional step in the Monte Carlo algorithm, the equations describing the binding of Ca2+ ions to buffers and glutamate vesicles are solved using a probabilistic interpretation of the kinetic reactions. More details about the stochastic implementation of the equations are given in [1,2]. The parameters used in our simulations are shown in Table 1. Parameters of the diffusion parameters for both Ca2+ and the endogenous and exogenous Ca2+ buffers were adopted from [1,2,4] without modification. Table 1. Parameter values adopted from [1, 2, 4] Parameter Value
k2
Description 0.7161 exp 0.1183(Vm − 32) Transition rate (O to C, in ms−1 ) 0.00037 exp − 0.0385(Vm − 32) Transition rate (C to I in ms−1 ) 2.48 Cai /Cao Transition rate (C to I in ms−1 )
k−2
0.302
k1 k−1
k+
Transition rate (I to C ms−1 ) −1
27.6 μM s −1
Forward binding rate (Ca2+ buffers) Backward binding rate (Ca2+ buffers)
k−
2150 s
b
0.4
Cooperativity factor
γ
1695 s−1
Secretion rate
216
3
C. A. Valverde-Alonzo et al.
Results and Discussion
Simulations of short depolarizing pulses (5 and 10 ms) suggest that the readily releasable pool is required by sensory hair cells to be able to release glutamate in response to short stimuli. This is shown in Fig. 2A and B, where it can be seen that when vesicles were not explicitly co-localized to the VDCC (RRP with a 0% of IRP), Ca2+ entering due to the stimuli were not capable of triggering the fusion of any glutamate vesicle. In contrast, secretion was triggered only when 10% of the vesicles were located very close to the Ca2+ channels (RRP with a 10% of IRP). Under these conditions, the 10 ms depolarizing produced a slightly higher secretory response than the 5 ms stimulus. Notice that this behavior remains even if the amount of endogenous buffering changes from 4 to 5.5 mM. Short stimuli (5 and 10 ms) allowed the Ca2+ concentration at the active zone to reach between ∼8 and 10 μM (Fig. 3A), an increase clearly limited by the bounding of Ca2+ to buffer molecules, as illustrated in Fig. 3B.
Fig. 2. Accumulated secretion of sensory hair cells varying the endogenous buffer concentration (4, 4.5, 5 and 5.5 mM) and the percentage of vesicles from the RRP forming the IRP (0 and 10%) in response to (A) 5, (B) 10 and (C) 20 ms voltage pulses from -90 to 0 mV (n = 5 in all cases). Bar plots with error bars show the corresponding means and standard deviations.
According to our simulations, longer pulses (20 ms) produced a much higher secretion in comparison to the 5 and 10 ms stimuli, as illustrated in Fig. 2C, where the accumulated secretion even reached 100% (32 vesicles) when 4 mM of Ca2+ endogenous buffers were included in the simulations whether vesicles were located at the readily releasable pool or not. Higher concentrations of Ca2+ buffer (4.5, 5 and 5.5 mM) reduced the secretory response by ∼25, 70 and 90%, respectively. Interestingly, the fraction of the readily releasable pool forming the IRP seems to be more relevant as the concentrations of Ca2+ buffer increases, as can be seen for the cases of 5 and 5.5 mM, where it can be appreciated that having vesicles at the IRP (RRP with a 10% of IRP) allowed the active zone to produce a considerably higher concentration in both cases in comparison to the simulations performed without vesicles at the IRP (RRP with 0% of IRP). In contrast to the shorter pulses, the Ca2+ concentration at the active zone raised much higher in response to a 20 ms stimulus, as depicted in Fig. 3A, where
Role of Ca2 + Buffers and RPF on Fast Secretion
217
a more pronounced increase occurs when the bounded buffer molecules reach a maximum value (∼4.5 mM, Fig. 3B). The late pronounced increase in the Ca2+ concentration of the active zone triggers the release of glutamate vesicles, as shown in Fig. 3C. These results suggest that fast glutamate secretion from active zones of sensory hair cells requires a prolonged and steeper increase in Ca2+ , which can be only achieved once the Ca2+ buffering capacity is reduced.
Fig. 3. A. Ca2+ concentration at the active zone and (B) Bound Ca2+ concentration (due to the endogenous buffer) for the 5, 10 and 20 ms stimuli. C. Temporal evolution of the accumulated secretion for the 20 ms stimulus.
Experimental observations by Johnson et al. [3] have shown that, depending on their location in the cochlea, secretion from sensory hair cells is heterogeneous: cells located at the apical, middle and basal regions of the cochlea respond to low-frequency (∼0.35 kHz), middle-frequency (∼2.5 kHz) and high-frequency (∼30 kHz) sounds, respectively. The origin of this heterogeneous behavior is still unknown, but our results suggest that the endogenous Ca2+ buffers and/or the composition of the readily releasable pool might be involved. This is in close agreement with previous reports about differences in endogenous buffering along the cochlea [3,8,9], that is also related to the vulnerability to hearing loss due to noise or age [10].
4
Conclusions
In this work, our simulations allowed us to conclude that 1) an immediately releasable pool is required for the fast secretion of glutamate in response to short stimuli ( 0, α(q, τ ) presents greater variations along τ . Significantly higher coefficients were observed for scales 44 ≤ τ ≤ 68 with 1 ≤ q ≤ 2 during daytime sleep. Additionally, the variations of α(q, τ ) were more pronounced during nighttime sleep for 2 < q ≤ 5. The correlation in nighttime sleep decreased to values close to 0.6 at scales 128 ≤ τ ≤ 256, while, in daytime
328
R. Delgado-Aranda et al.
Fig. 1. Average surfaces of MMF-DFA coefficients, α(q, τ ), during (a) daytime and (c) nighttime sleep. The α(q, τ ) coefficients for the (b) daytime and (d) nighttime surrogate sleep data. The color map represents in yellow the significant differences between daytime vs nighttime sleep of the (e) original and (f) surrogate data with a p < 0.05. (Color figure online)
sleep, it reduced to around 0.8. However, no significant differences were found, suggesting that despite the change in sleep schedule, the autonomic nervous system adapts by maintaining the long-term correlation of the IBI signal in large fluctuations. Figures 1b and d showed the fractal structures of the surrogate data, which are slightly similar to the dynamics of the original IBI signal for q > 0. However, they appear to be less dependent on q. Also, Fig. 1f shows no statistical difference when comparing the daytime and nighttime sleep scale coefficients of the surrogate data. The results suggest that statistical differences observed during day and night time sleep are due to an alteration in the correlation of the IBI signal. The α(q, τ ) coefficients obtained from the original IBI were also compared with respect to the surrogate data (graphs are not shown). Surrogate data of nighttime sleep had 38% of the scales statistically different from the original time
Multiscale-Multifractal Assessment
329
series, while in the case of the daytime sleep a 29% was found. These results are expected because the surrogate signals conserve the 2nd order statistics but not the multifractal components [7]. The results suggest that on specific scales, the HRV of shift workers during daytime sleep has significantly higher α(q, τ ) than during nighttime sleep. These alterations in the IBI correlation could be due to the altered sleep macrostructure and decreased total daytime sleep time observed in shift workers. Sleep duration and quality are associated with the metabolic functioning of the organism [12]. In addition, because shift work causes circadian rhythm disturbance, there is a hormonal disturbance that could influence the regulation of physiological processes during sleep [12]. These alterations in the organism can lead to a decrease in sleep quality and therefore cause alterations in the cardiovascular system. Finally, although the organism attempts to adapt to the changes, maintain homeostasis, and achieve adequate sleep, continuous alterations might increase the risk of cardiovascular disease.
4
Conclusions
This work compared patterns of heart rate complexity of shift workers during the daytime and nighttime sleep as assessed by the multifractal multiscale DFA approach. The MMF-DFA of IBI showed that the α(q, τ ) depends on the time scale and the different component fractals. In some daytime sleep scales, a significant increase in the correlation of small and large fluctuations was observed. The high long-term correlation during daytime sleep could be associated with the alteration of the sleep-wake cycle due to the irregular sleep schedule. Also, modifications in the cardiovascular system due to changes in sleep schedule could increase the risk of cardiovascular disease. However, interpretation of the physiological phenomena associated with HRV occurring in this range of scales could be relevant in the assessment of the impact of shift work on workers’ health. In addition, comparing the multiscale dynamics of HRV of healthy shift workers with respect to a control group that sleeps only during the night could help to understand the heart rate dynamics in shift workers. Thus, evaluating the multifractal multiscale surface of scaling coefficients during sleep could be a tool to improve the evaluation of HRV alteration due to shift work.
References 1. Castiglioni, P., Di Rienzo, M., Faini, A.: Self-similarity and detrended fluctuation analysis of cardiovascular signals. In: Barbieri, R., Scilingo, E.P., Valenza, G. (eds.) Complexity and Nonlinearity in Cardiovascular Signals, pp. 197–232. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-58709-7 7 2. Castiglioni, P., Faini, A.: A fast DFA algorithm for multifractal multiscale analysis of physiological time series. Front. Physiol. 10, 115 (2019). https://doi.org/10. 3389/fphys.2019.00115
330
R. Delgado-Aranda et al.
3. Castiglioni, P., Lazzeroni, D., Coruzzi, P., Faini, A.: Multifractal-multiscale analysis of cardiovascular signals: a DFA-based characterization of blood pressure and heart-rate complexity by gender. Complexity 2018, 1–14 (2018). https://doi.org/ 10.1155/2018/4801924 4. Castiglioni, P., Omboni, S., Parati, G., Faini, A.: Day and night changes of cardiovascular complexity: a multi-fractal multi-scale analysis. Entropy 22(4), 462 (2020). https://doi.org/10.3390/e22040462 5. Castiglioni, P., Parati, G., Civijian, A., Quintin, L., Rienzo, M.: Local scale exponents of blood pressure and heart rate variability by detrended fluctuation analysis: effects of posture, exercise, and aging. IEEE Trans. Biomed. Eng. 56(3), 675–684 (2009). https://doi.org/10.1109/TBME.2008.2005949 6. Castiglioni, P., Parati, G., Di Rienzo, M., Carabalona, R., Cividjian, A., Quintin, L.: Scale exponents of blood pressure and heart rate during autonomic blockade as assessed by detrended fluctuation analysis. J. Physiol. 589(2), 355–369 (2011). https://doi.org/10.1113/jphysiol.2010.196428 7. Castiglioni, P., Parati, G., Faini, A.: Multifractal and multiscale detrended fluctuation analysis of cardiovascular signals: how the estimation bias affects shortterm coefficients and a way to mitigate this error. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC), pp. 257–260 (2021). https://doi.org/10.1109/EMBC46164.2021.9629623 8. Chung, M.H., Kuo, T., Hsu, N., Chu, H., Chou, K.R., Yang, C.: Recovery after three-shift work: relation to sleep-related cardiac neuronal regulation in nurses. Ind. Health 50(1) (2012). https://doi.org/10.2486/indhealth.MS1305 9. Gu, F., et al.: Total and cause-specific mortality of U.S. nurses working rotating night shifts. Am. J. Prev. Med. 48(3), 241–252 (2015). https://doi.org/10.1016/j. amepre.2014.10.018 10. Guerrero, G., Kortelainen, J., Palacios-Hern´ andez, E., Tenhunen, M., Bianchi, A., M´endez, M.: Evaluation of pressure bed sensor for automatic SAHS screening. IEEE Trans. Instrum. Meas. 64(7), 1935–1943 (2015). https://doi.org/10.1109/ TIM.2014.2366976 11. Hulsegge, G., et al.: Shift work is associated with reduced heart rate variability among men but not women. Int. J. Cardiol. 258, 109–114 (2018). https://doi.org/ 10.1016/j.ijcard.2018.01.089 12. Jehan, S., et al.: Shift work and sleep: medical implications and management. Sleep Med. Disord. 1(2) (2017). https://doi.org/10.15406/smdij.2017.01.00008 13. Kantelhardt, J., Zschiegner, S., Koscielny-Bunde, E., Havlin, S., Bunde, A., Stanley, H.: Multifractal detrended fluctuation analysis of nonstationary time series. Phys. A: Stat. Mech. Appl. 316(1), 87–114 (2002). https://doi.org/10.1016/S03784371(02)01383-3 ˙ 14. Kokosi´ nska, D., Gieraltowski, J.J., Zebrowski, J.J., Orlowska-Baranowska, E., Baranowski, R.: Heart rate variability, multifractal multiscale patterns and their assessment criteria. Physiol. Meas. 39(11), 114,010 (2018). https://doi.org/10. 1088/1361-6579/aae86d 15. Kortelainen, J., van Gils, M., P¨ arkk¨ a, J.: Multichannel bed pressure sensor for sleep monitoring. In: 2012 Computing in Cardiology, vol. 39, pp. 313–316. IEEE (2012) 16. Makowiec, D., Gala, R., Dudkowska, A., Rynkiewicz, A., Marcin., Z.: Long-range dependencies in heart rate signals-revisited. Phys. A: Stat. Mech. Appl. 369(2), 632–644 (2006). https://doi.org/10.1016/j.physa.2006.02.038
Multiscale-Multifractal Assessment
331
17. Peng, C.K., Havlin, S., Stanley, H., Goldberger, A.: Quantification of scaling exponents and crossover phenomena in nonstationary heartbeat time series. Chaos 5(1), 82–87 (1995). https://doi.org/10.1063/1.166141 18. Taylor, J., Carr, D., Myers, C., Eckberg, D.: Mechanisms underlying very-lowfrequency RR-interval oscillations in humans. Circulation 98(6), 547–555 (1998). https://doi.org/10.1161/01.CIR.98.6.547 19. Torquati, L., Mielke, G., Brown, W., Kolbe-Alexander, T.: Shift work and the risk of cardiovascular disease. A systematic review and meta-analysis including dose-response relationship. Scand. J. Work Environ. Health 44(3), 229–238 (2018). https://www.jstor.org/stable/26567001 20. Vyas, M., et al.: Shift work and vascular events: systematic review and meta-analysis. BMJ 345, e4800 (2012). https://doi.org/10.1136/bmj.e4800pmid: 22835925 21. Wessel, N., et al.: Nonlinear analysis of complex phenomena in cardiological data. Herzschr Elektrophys 11, 159–173 (2000). https://doi.org/10.1007/s003990070035
EEG Connectivity Analysis in a Motor Imagery Task C´esar Covantes-Osuna , Omar Paredes , Diana Yaneli De la Mora , Hugo V´elez-P´erez , and Rebeca Romo-V´azquez(B) Departamento de Bioingenieria Traslacional, Centro Unviersitario de Ciencias Exactas E Ingenierias, Universidad de Guadalajara, 44430 Guadalajara, Jalisco, Mexico [email protected] Abstract. EEG recordings have been used to study the spatio-temporal dynamics of brain processes. In motion intention, these analyses have detected EEG patterns present before the movement begins. This dynamics can be mapped using network models based on global functional brain connectivity, estimated from temporal connectivity matrices of all electrodes. However, this temporal matrix representation leads to dense networks containing redundant information. Brain-computer interfaces (BCI) are systems using brain patterns as control signals. Therefore, the simplification of calculations in devices or applications that need brain information with the least number of electrodes, without losing significant information, identifying the most significant channels associated with motion intention, becomes an important task. The aim of this work is to propose a methodology to select EEG electrodes containing significant information during the motor intention. Thus, a single-layer model was fitted to the BNCI Horizon 2020 motor imagery task database (2a of BCI Competition IV), including EEG recordings of four different motor imagery tasks: left hand, right hand, feet, and tongue. The correlation index was computed to estimate the pairwise correlation between EEG channels in α1 (8–10 Hz), α2 (10–13 Hz), and β1 (13–18 Hz) bands. To simplify, the adjacency matrices were thresholded. Then, a graph analysis was conducted using the degree and the eigenvector centrality as graph metrics. Later, a statistical analysis was made to identify the most significant channels associated with motion intention. Our results show that the electrodes with significant differences are located in the fronto-central area (C2, C3, FC1, Fz). These results agree with other works but also estimate the electrodes associated with the motor imagery tasks. Keywords: Correlation · EEG metrics · Motion intention
1
· Functional connectivity · Graph
Introduction
The spatio-temporal dynamics of the brain can be mapped by various techniques such as electroencephalography (EEG), which is the most widely employed scanning technique due to its accessibility and that records over-the-scalp brain electrical activity at a millisecond resolution [1]. The brain dynamics is commonly c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 332–341, 2023. https://doi.org/10.1007/978-3-031-18256-3_37
EEG Connectivity Analysis in a Motor Imagery Task
333
represented through the construction of network models based on brain connectivity. In particular, the functional connectivity can be estimated in time by methods such as correlation, and, in the frequency domain, by the coherence, and phase synchronization methods, besides other approaches, to find the connection between the neural populations. These functional interactions result from the temporal synchronization of activity between areas of the brain, both local and distant. Thus, the brain networks are constructed by regions distributed throughout brain space, but functionally connected to process information [2]. Functional connectivity reflects the statistical dependencies between close and distant regions of the information processed by different neural populations. The most common method to estimate functional connectivity between two signals in time is Pearson’s coefficient. To explore the whole brain’s functional connectivity, a global connectivity matrix is built from the correlation coefficients between all electrodes on the scalp at each instant. This temporal matrix representation could be an effective way to visualize large and/or dense networks [3,4]. However, these kinds of networks contain redundant information. Some devices, applications, software, or systems, like Brain-Computer Interfaces (BCIs), use EEG recordings to acquire brain activity. BCIs could be considered systems that record, quantify, and classify brain patterns for use as control signals in external devices to the human body [5]. Although the research, development, and technologies of the BCIs have grown considerably, different aspects such as speed, accuracy, precision, consistency, security, reliability, and practicality, still need to be improved [6]. Simplifying calculations in devices or applications that require recording brain information with the least number of electrodes, without losing significant information, and identifying the most significant channels associated with movement intention, also becomes an important task. In the other hand, EEG recordings have been used to study brain processes of motion, cognition, and sleep [7], as well as their disorders. In motion, imaginary and real movements have been extensively explored. Particularly, in motion intention, the analysis of brain signals has allowed the detection of the EEG patterns present before the movement begins. Thus, the motion intention has been associated with α (8–12 Hz) and β (13–30 Hz) rhythms [8–10]. Motion intention has been also related to a decrease in amplitude before the movement starts in the EEG signals (Movement Related Cortical Potential, MRCP) [11] and with an implementation of Event-Related Desynchronization (ERD), which is observed as a desynchronization in α and β bands before the movement starts [12]. MRCP refers to all potentials related to the planning and execution of movement, starts ∼2–1.5 s before the start of movement in EEG recordings, and is commonly recorded at F3, Fz, F4, C3, Cz, C4, P3, Pz, and P4 electrodes [13]. Additionally, other works have been focused to identify the cortical zones involved in the imaginary movement [6,14–16]. For its part, the ERD signal appears at the same time as the MRCP as a result of the desynchronization of the contralateral cortical areas during the movement intention.
334
C. Covantes-Osuna et al.
A desynchronized EEG means that a neural network or neuronal circuitry works in relative independence [17]. In the above context, this study focused on analyzing, in a classical singlelayer model, the dynamics of EEG recordings for a motor imagery dataset. Pearson’s coefficient was computed to estimate the pairwise correlation between EEG channels. Later, the correlation adjacency matrices in α (α1 (8–10 Hz) and α2 (10–13 Hz)) and β (β1 13–18 Hz) bands were estimated. To simplify the analysis, the networks were reduced to a non-directed binary form through the use of thresholding criteria [18]. This threshold allows for retaining the most significant connections presented in the network. Finally, a statistical analysis was aimed to identify the most significant channels associated with motion intention.
2 2.1
Materials and Methods Database
To explore patterns of motor imagery, the BNCI Horizon 2020 database (2a of BCI Competition IV) [19] was fetched. This dataset includes two EEG recordings of nine subjects acquired on two different days (18 EEG recordings in total). Each participant completed a sequence of four different motor imagery tasks: left hand, right hand, feet, and tongue, each labeled as class 1–4, respectively. The EEG signals were acquiered at a fs = 250 Hz and filtered between 0.5 100 Hz. The electrodes recorded were Fc1, Fc2, Fc3, Fc4, Fcz, Fz, C1, C2, C3, C4, C5, C6, Cz, Cp1, Cp2, Cp3, Cp4, Cpz, P1, P2, Pz, and POz. 2.2
Experimental Paradigm
The scheme paradigm is illustrated in Fig. 1. First, the participants were placed in front of a screen. Each test began with a black screen, where the subjects focused their sight (t = 0 s). Two seconds later (t = 2 s), the image of an arrow appeared pointing right, left, up, or down (one for each class) and remained on the display for 1.25 s. The subjects then completed the corresponding motor imaging task until the arrow image vanished at t = 6 s. Finally, a short break was indicated before starting the following trial. The time window considered in this study corresponds to the motion intention (t = 2.5–4.5 s) [20–23] of the experimental paradigm analyzed (red rectangle of Fig. 1). 2.3
Data Selection
Figure 2 illustrates a schematic flowchart of the methodology applied to the time windows selected. Here, Fig. 2a corresponds to the EEG recordings of the four motor intention classes from the paradigm. Six runs separated by a brief pause integrated every single EEG recording. Subjects executed 12 trials of each class (48 trials total) in each run. A total of 288 two-second EEG windows each (72 trials of each class) were generated.
EEG Connectivity Analysis in a Motor Imagery Task
335
Fig. 1. In red, the motor intention time window location (t = 2.5–4.5 s (Color figure online)) of the experimental paradigm.
2.4
Preprocessing
For all the 288 EEG windows, a Common Average Reference (CAR) filter (Eq. 1) was used to reduce the spatial interference (Fig. 2b). ViCAR = ViCR −
1 CR V N j=1 j
(1)
where ViCR is the potential between the reference electrode and the i-th electrode, N represents the total number of electrodes. Later, the EEG signals were filtered through the wavebands to be analyzed (α1 , α2 , and β1 ). 2.5
Adjacency Matrices Estimation
The adjacency matrices were estimated using the correlation index as connectivity estimator. Since correlation index values range from −1 to 1, in this work the absolute value of correlation was used. Correlation index is defined as (Eq. 2): n( xy) − ( x)( y) (2) rxy = [n x2 − ( x)2 ][n y 2 − ( y)2 ] where x and y represent two signals (channels), rxy is the correlation matrix in time domain.We averaged the adjacency matrix of the 72 windows (of each waveband) of each class to get four two-second average EEG windows. This proceeding was completed for the 18 records. (Figure 2c). Finally, a mean threshold was applied to eliminate stray connections (Fig. 2d).
336
C. Covantes-Osuna et al.
Fig. 2. Schematic flowchart of the methodology. Here, (a) corresponds to the EEG recordings of the four motor intention classes from the paradigm, (b) represents the data preprocessing step, while (c - d) are associated with the thresholded adjacency matrices estimation from the correlation index, in (e) graph metrics are computed, and finally, the statistical analysis is presented in (f).
2.6
Graph Analysis
After applying the threshold, unweighted non-directed graphs were obtained on which graph metrics were estimated. The employed graph metrics implemented for this work consisted in the degree and the eigenvector centrality. Degree is the metric that evaluates the number of neighbors for each node (electrode); whereas centrality refers to connectivity between neighbors of the evaluated node (Fig. 2e). 2.7
Statistical Analysis
Then, a MANOVA test was carried out to verify if there were differences between the metrics of each electrophysiological sub-band (Fig. 2f). Such analysis was conducted to explore if any change in brain inter-connectivity occurs during motor intention due to a modulation shift in the individual sub-bands. Next, to measure the movements effect in the MANOVA analysis, a partial eta squared test was computed, and to visualize such movement effect a linear discriminant analysis (LDA) was implemented. After, multiple one-way ANOVAs were estimated to determine which electrodes presented significant differences.
EEG Connectivity Analysis in a Motor Imagery Task
3
337
Results and Discussion
Significant differences (p = 2.2 × 10−16 ) between the four movements were found in the three MANOVAs corresponding to each electrophysiological sub-band. By applying the partial eta test, it was found that they all obtained values above 0.14 (see Table 1). This indicates that certain single movements are highly influential in the data variance [24]. The variance influence is shown in Fig. 3, which displays the LDAs for degree and eigenvector centrality. Table 1. Partial eta squared for the two graph metrics in each electrophysiological sub-band. Sub-band Degree Eigenvector alpha I
0.75
0.75
alpha II
0.85
0.84
beta I
0.79
0.79
From Fig. 3a, it is observed that in α2 and β1 sub-bands, the outermost movement corresponds to the right hand. As the evaluated metric in this figure (Fig. 3a) is the degree, this may suggest the existence of an electrode interconnectivity that varies between these two sub-bands, where the μ rhythm is located. Similar behavior is also observed in the eigenvector centrality metrics (Fig. 3b), recalling that this metric refers to the neighbor connectivity among the electrodes. Both of these arguments strengthen the hypothesis that the brain dynamics during movements, most noticeably on the right hand, are changing. This could be due to a possible right laterality of the participants. However, it is not possible to affirm this because the authors of the database used do not specify this data. A further noteworthy trend in both metrics occurs for the left hand and tongue in all three sub-bands. For both motions we can visually infer that the centroids of both movements are differentiable among them. The latter is relevant since it enables further analysis to predict the motor imagery. Although for feet, such behavior is not as evident as in other movements.
Fig. 3. Linear Discriminant Analysis (LDA) representation for the four evaluated movement.
(b) Eigenvector Centrality.
(a) Degree.
338 C. Covantes-Osuna et al.
EEG Connectivity Analysis in a Motor Imagery Task
339
In order to propose the candidate electrodes containing significant information during the motor intention task, specifically those that can control the inter-movement dynamics, a series of ANOVAs were performed. The electrodes with significant difference for the three analyzed sub-bands are listed in Table 2. From these table, it is possible to observe that the electrodes that exhibit significant differences in both metrics and throughout all the sub-bands are located at the fronto-central area (C2, C3, FC1, Fz). These results agree with others works that have reported that these areas are associated with the motor imagery movements [25]. Table 2. List of electrodes with significant differences in both metrics and by sub-band. In bold, the common electrodes are indicated. Sub-band alpha I alpha II beta I
4
Electrodes C2 C3 CP1 FC1 FC3 Fz P1 P2 Pz C2 C3 C4 FC1 CP2 FC3 FC2 FCz Fz P1 P2 C1 C2 C3 C4 CP1 FC1 CP2 CP4 FC2 FCz Fz POz
Conclusion
Our analysis shows that the electrodes in the fronto-central region, specifically the electrodes (C2, C3, FC1, Fz) are those that containing significant information during the motor intention task. These findings are consistent with those found in the literature. From the optics of simplifying calculations or applications that need a minimum of electrodes, these studies are important since they allow the identification of the most significant electrodes during the different brain processes, thus eliminating electrodes with redundant information. Finally, for future work the implementation with another brain connectivity estimation will be consider. Acknowledgments. The authors wish to thank the Consejo Nacional de Ciencia y Tecnologa (CONACyT) for the support received through the scholarship grant CVU311539, CVU713526, and CVU1103073.
Conflict of Interest. The authors declare that they have no conflict of interest.
References 1. Subasi, A., Ercelebi, E.: Classification of EEG signals using neural network and logistic regression. Comput. Methods Program. Biomed. 78(2), 87–99 (2005) 2. Hassan, M., Wendling, F.: Electroencephalography source connectivity: toward high time/space resolution brain networks. arXiv preprint arXiv:1801.02549 (2018)
340
C. Covantes-Osuna et al.
3. Alper, B., Bach, B., Riche, N.H., Isenberg, T., Fekete, J.-D.: Weighted graph comparison techniques for brain connectivity analysis. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 483–492. ACM (2013) 4. Yi, J.S., Elmqvist, N., Lee, S.: Timematrix: Analyzing temporal social networks using interactive matrix-based visualizations. Intl. Journal of Human-Computer Interaction, 26(11–12), 1031–1051 (2010) 5. Wolpaw, J., Wolpaw, E.W.: Brain-computer interfaces: principles and practice. OUP USA (2012) 6. Shih, J.J., Krusienski, D.J., Wolpaw, J.R.: Brain-computer interfaces in medicine. In Mayo Clinic Proceedings, vol. 87, pp. 268–279. Elsevier (2012) 7. Saper, C.B., Cano, G., Scammell, T.E.: Homeostatic, circadian, and emotional regulation of sleep. J. Comp. Neurol. 493(1), 92–98 (2005) 8. Nicolas-Alonso, L.F., Gomez-Gil, J.: Brain computer interfaces, a review. Sensors 12(2), 1211–1279 (2012) 9. Kilavik, B.E., Zaepffel, M., Brovelli, A., MacKay, W.A., Riehle, A.: The ups and downs of beta oscillations in sensorimotor cortex. Exp. Neurol. 245, 15–26 (2013) 10. Hasan, S.M.S., Siddiquee, M.R., Atri, R., Ramon, R., Marquez, J.S., Bai, O.: Prediction of gait intention from pre-movement EEG signals: a feasibility study. J. NeuroEng. Rehab. 17(1), 1–16 (2020) 11. Bai, O., et al.: Prediction of human voluntary movement before it occurs. Clin. Neurophysiol. 122(2), 364–372 (2011) 12. Pfurtscheller, G., Brunner, C., Schl¨ ogl, A., Lopes Da Silva, FH.: Mu rhythm (de) synchronization and EEG single-trial classification of different motor imagery tasks. NeuroImage, 31(1), 153–159 (2006) 13. Shakeel, A., Navid, M.S., Anwar, M.N., Mazhar, S., Jochumsen, M., Niazi, I.K.: A review of techniques for detection of movement intention using movement-related cortical potentials. Comput.Math. Methods Med. (2015) 14. Pfurtscheller, G., Lopes da Silva, F.H.: Event-related EEG/MEG synchronization and desynchronization: basic principles. Clinical Neurophysiol. 110(11), 1842–1857 (1999) 15. Stanˇca ´k Jr, A., Feige, B., L¨ ucking, C.H., Kristeva-Feige, R.: Oscillatory cortical activity and movement-related potentials in proximal and distal movements. Clinical Neurophysiol. 111(4), 636–650 (2000) 16. Bobrova, E.V., Reshetnikova, V.V., Frolov, A.A., Gerasimenko, Y.P.: Use of imaginary lower limb movements to control brain-computer interface systems. Neuroscience and Behavioral Physiology 50(5), 585–592 (2020) 17. Pfurtscheller, G.: Functional brain imaging based on erd/ers. Vision. Res. 41(10– 11), 1257–1260 (2001) 18. Rubinov, M., Sporns, O.: Complex network measures of brain connectivity: uses and interpretations. Neuroimage 52(3), 1059–1069 (2010) 19. Brunner, C., Leeb, R., M¨ uller-Putz, G., Schl¨ ogl, A., Pfurtscheller, G.: Bci competition 2008-graz data set a. Institute for Knowledge Discovery (Laboratory of Brain-Computer Interfaces), Graz Univ. Technol. 16, 1–6 (2008) 20. Ang, K.K., Chin, Z.Y., Wang, C., Guan, C., Zhang, H.: Filter bank common spatial pattern algorithm on BCI competition iv datasets 2a and 2b. Front. Neurosci. 6 39 (2012) 21. Kam, T.-E., Suk, H.-I., Lee, S.-W.: Non-homogeneous spatial filter optimization for electroencephalogram (EEG)-based motor imagery classification. Neurocomputing 108, 58–68 (2013)
EEG Connectivity Analysis in a Motor Imagery Task
341
22. Gaur, P., Pachori, R.B., Wang, H., Prasad, G.: A multivariate empirical mode decomposition based filtering for subject independent BCI. In: 2016 27th Irish Signals and Systems Conference (ISSC), pp. 1–7. IEEE (2016) 23. Gaur, P., Pachori, R.B., Wang, H., Prasad, G.: A multi-class EEG-based BCI classification using multivariate empirical mode decomposition based filtering and riemannian geometry. Expert Syst. Appl. 95, 201–211 (2018) 24. Lakens, D.: Calculating and reporting effect sizes to facilitate cumulative science: a practical primer for t-tests and ANOVAs. Front. Psychol. 4, 863 (2013) 25. Rathee, D., Cecotti, H., Prasad, G.: Estimation of effective fronto-parietal connectivity during motor imagery using partial granger causality analysis. In: 2016 Int. Joint Conf. Neural Netw.(IJCNN), pp. 2055–2062. IEEE (2016)
Brain Mapping: Location of the Words Through EEG Omar Cano-Garcia1
, María Hernández-Rizo1(B) , Lorena López-Medina1 and J. Alejandro Morales2
,
1 Centro Universitario de Ciencias Exactas e Ingenierías, Universidad de Guadalajara,
44430 Guadalajara, Jalisco, Mexico [email protected] 2 Departamento de Bioingeniería Traslacional, Centro Universitario de Ciencias Exactas e Ingenierías, Universidad de Guadalajara, 44430 Guadalajara, Jalisco, Mexico
Abstract. At all times, the brain is capable of integrating and processing multiple sensory inputs, so knowing how listening comprehension is represented in the brain is essential in areas such as linguistics and neuroscience. In the present work, an electroencephalographic (EEG) recording was performed on 5 subjects to identify the brain response to auditory stimuli. The auditory stimuli correspond to 3 short stories in order to find the comprehension of the words within a narrative. Processing of words occurs in a range of 300 to 400 ms after they are heard. After preprocessing the signals, a time-frequency analysis was performed to find cerebral auditory activation in two cerebral parameters of the gamma rhythm (35, 45 Hz). The comprehension was determined by the two maximum values of the spectral density and we located the specific electrode that corresponds to the said word. Word comprehension was distributed over different areas of the cortex, therefore we did not find a specific location in the brain for listening comprehension. Keywords: EEG · Listening comprehension · Time-frequency analysis
1 Introduction Over the years, different areas of the brain have been associated with language processing [1]. Among them is Wernicke’s area, located in the temporal lobe [2]. More recent studies have as a premise that a function cannot be localized in a single area of the brain [3]. It has been determined that language processing occurs in different regions of the cortex [1]. Knowing how language is processed and understood along with its comprehension has become a research task in many disciplines such as linguistics, cognition, neuroscience, language engineering, among others [4]. At all times, the brain can integrate and process multiple sensory inputs, so it is important to know how the brain responses are manifested [5]. Huth et al. [6] prove the existence of a semantic system in the cortex by mapping the meaning of language using functional magnetic resonance image (fMRI), the study © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 342–351, 2023. https://doi.org/10.1007/978-3-031-18256-3_38
Brain Mapping: Location of the Words Through EEG
343
subjects listened to stories as auditory stimuli. The results showed a symmetry of activity in both hemispheres, so it can be determined that semantic processing takes place in different regions of the brain. Broderick et al. [7] were able to identify the electrical activity of the brain, associated with the semantic processing of language comprehension through EEG recordings. There are other research works that carry out brain mapping with auditory stimuli, these differ because they use a certain number of words, such is the case of the work carried out by Alcaraz et al. [8] who studied comprehension of a number of abstract and concrete words using EEG. The goal of this work is to locate the listening comprehension of words within a semantic context. We used EEG because we were interested in its high temporal resolution in contrast to fMRI since we seek to have a detailed analysis per word [9]. By understanding how the comprehension of the human language is manifested, we can design brain circuits to create artificial intelligence systems [10]. In addition, using EEG signals it is possible to predict the neural activation of the language comprehension and in this way improve the natural language processing (NLP) models applied in machine translation, speech processing, sentiment analysis and relation detection [10–12].
2 Methods 2.1 Participants The study involved 5 subjects (4 male, 1 female), all college students aged between 21 and 22. 5 EEG recordings were acquired. Each participant gave their approval to participate in the study through an informed consent, likewise, they referred not to suffer from any vision, hearing or neurological problem. 2.2 Stimuli 3 different stories were used as auditory stimuli, which were taken from conferences belonging to the Technology, Entertainment and Design (TED) the narrated stories use daily words, so the purpose was that the subjects know the meaning of the words they hear. The total duration of the stimulus makes up an audio of 16 min with 2 s. At the end of each story, there were 7 s of rest (see Fig. 1). The stimuli were reproduced with a sampling frequency of 44100 Hz and through two speakers placed 60 cm from the subject’s sides. The study was conducted in a room
Fig. 1. Experimental design.
344
O. Cano-Garcia et al.
isolated from external noise where the subject was seated at a distance of 1 m in front of a monitor observing a fixation cross. The instruction given to the subjects was to observe the fixation cross while listening to the stories carefully, they were asked to avoid sharp movements. 2.3 Segmentation and Relation of Words A textual corpus was obtained by transcribing the auditory stimulus, obtaining 2,305 words. Articles, links and prepositions were removed from the text. All those that were repeated throughout the corpus were taken as a single word. All the words with close meanings (such as mom-mother) were interpreted as one. From the complete list of words those that were similar were manually removed. We determined that plurals and singulars were a single word (child/children), resulting in 292 words that will be taken to analyze. All word processing was done using Python with the NLTK (Natural Language Toolkit) library. The word2vec model for language processing consists of a word embedding space built using the co-occurrence of these within a text [13]. The 292 selected words formed a word embedding, with this our own word2vec model was trained and pragmatic features were extracted from it. A semantic model of words was developed using the Python WordNet library. We construct a matrix of 292 × 25, where the rows are the selected words, and the columns are made up of 20 pragmatically related words obtained from word2vec and 5 semantically related words obtained from WordNet. With this pragmatic and semantic relationship matrix, we can make a correlation between both categories in order to know the relationship that all the words have with each other. 2.4 EEG Recording The electroencephalographic signal was acquired at 256 Hz through a g.HIAMP 256 with an equidistant arrangement of 32 active electrodes according to the 10–20 system with reference in the right and left mastoid (see Fig. 2). Two digital filters, a bandpass filter ranging from 0.1 Hz to 100 Hz, and a notch filter at 60 Hz, were applied to remove line noise.
Fig. 2. Distribution of the 32 electrodes.
Brain Mapping: Location of the Words Through EEG
345
2.5 Signal Pre-processing and Artifact Removal The artifacts in the EEG signal correspond to electrophysiological signals external to the brain, such as heart rate, blinking or muscle movements. Independent Component Analysis (ICA) performs a decomposition of the EEG signal into Independent Components (ICs) [14], using EEGLAB (Delorme and Makeig 2004) a MATLAB (The Mathworks, R2022a) toolbox, ICA was applied and with ICLabel we identified and removed the artifact of blinking from the signal (see Fig. 3 and Fig. 4) [15]. We used Second-Order Blind Identification (SOBI) because it is the most efficient ICA method, obtaining similar results to other algorithms in a shorter execution time [16]. We implemented the Wavelet Transform (WT) for noise removal. According to Al-Qazzaz et al. [17], Daubechies, Coiflet and Symlet are the best option to perform noise removal on an EEG signal. 25 tests were performed with the Mother Wavelets: Daubechies (db1-db10), Symlet (sym1-sym10) and Coiflet (coif1-coif5). We selected
Fig. 3. Independent components of an EEG recording. (a) Independent components calculated with SOBI. (b) Blink-related independent component (IC2).
Fig. 4. Application of SOBI to remove ocular artifact from an EEG record. A 5 s fragment of the original signal is shown in blue and the new signal without artifacts in red.
346
O. Cano-Garcia et al.
the Mother Daubechies in her order 2 for its similarity to the QRS complex in order to remove the heart rate from the signal [18]. 2.6 Signal Processing and Analysis Within the electroencephalographic signals there are brain rhythms, associated with different activities, these rhythms have a specific frequency that varies according to the activity being performed [19, 20]. Gamma waves are found at oscillations of 30– 90 Hz [19]. Gamma brain rhythms have shown to be associated with general mental activities, as well as attention and comprehension [19–21]. There is evidence indicating that word processing occurs in a range of 300–400 ms after a word is finished [22, 23]. Time-frequency analyses are important for the detection of responses in a temporal range of EEG signals [24]. A time-frequency analysis was performed on the recordings, producing two frequency parameters (35, 45 Hz) to find the cerebral gamma rhythm. We looked for these frequencies 400 ms right after each auditory stimulus and each word was listened. Different authors have evidenced a high inaccuracy when searching for frequencies at a specific point within a very long signal [24], therefore, we divided the signals into 2-s segments (epochs). Through the Brainstorm software (Tadel et al. 2011) [25], the spectral density of each epoch was obtained together with a temporal reference to locate the frequency parameters (35, 45 Hz) at a point in the time which represents when the subject processes a word.
3 Results and Discussion The frequency spectrum of the signal was obtained to find the two maximum values in each channel for each analyzed word, based on the two established frequency parameters and the moment in which the subject heard the word. From these values, we determined in which channel the comprehension of each word was located. The comprehension of a word can be distributed in different channels, therefore, two maximum values corresponding to two different channels are selected. Although channel activation varies between subjects, it can be concluded that word comprehension occurs in both cerebral hemispheres, as well as in different areas of the cortex (see Fig. 5). The data of a subject were taken to make a three-dimensional representation to graphically observe the approximate location of the comprehension of the words with respect to the electrodes used. In the 3D model, all the analyzed words are represented with different colors and located around the electrode where the greatest activation was detected (see Fig. 6).
Brain Mapping: Location of the Words Through EEG
347
Fig. 5. Number of words according to the channel in which they generated a response.
By analyzing the groups of words corresponding to each channel we observed that there is a certain pragmatic and semantic relationship between the words located in the same channel, that is, words within the same context or with a similar meaning. As can be seen in Fig. 5 and Fig. 6, the channels where a greater number of words were found were located in the occipital area and in the right frontoparietal area. On the other hand, no very significant activity was found in the motor cortex, which coincided in the results of all the subjects. From the channel distribution of the words analyzed and the variation between subjects, it cannot be concluded that listening comprehension within a semantic context occurs only in a specific area of the brain.
348
O. Cano-Garcia et al.
Fig. 6. 3D mapping of responses to words in the brain (a) Front view. (b) Rear view. (c) Right lateral view. (d) Left lateral view.
Some authors have mentioned that all words are stored as a setlist in the form of a dictionary, where it is believed that they are located in the left middle temporal gyrus through a semantic network [7, 26]. It has been mentioned that the processing of language and words are distributed over different regions of the cortex [6]. Our results show a partial homogeneous distribution between the two hemispheres of the brain, for which we obtained almost the same number of words located per hemisphere (see Fig. 7), some type of lateralization can be ruled out. Not all channels presented the same number of words understood, the main channels were the following: FP1, F7, T7, CP5, P7, PO3, O1 in the case of the left hemisphere and FP2, F8, T8, P8, PO4 for the right hemisphere, while the PZ and OZ channels presented the highest number of words as reference channels. We would like to mention that the channels PO3, O1, FP2, PO4, T7 and P7 had the highest number of words found. It could be argued that there may be some reference points for words such as a dictionary [7, 26] but with a distribution over different areas of the cortex [6].
Brain Mapping: Location of the Words Through EEG
349
Fig. 7. Number of words located per channel of all subjects (a) Left hemisphere. (b) Right hemisphere. (c) Reference channels.
4 Conclusion It was long thought that language processing and comprehension occurred in Wernicke’s are, in our study we followed the premise that an area in the brain is not necessarily associated with a specific function. After obtaining the analysis of five electroencephalographic recordings in different subjects, we were able to determine that the comprehension of the words is not located in a specific area of the brain, but it takes place in different regions of the brain signals associated with listening comprehension (see results). With the results obtained thanks to the 3D model, it is intended to determine a categorization of words according to their pragmatics, semantics and brain activity. In the near future, we will seek to make more records to different individuals to find possible patterns that lead to the generalization of listening comprehension in order to classify all the words used to determine which words are related pragmatically and semantically, Declarations. Consent to Participate Written. informed consent was obtained from all individual participants included in the study. Ethics Approval. This study was performed in line with the principles of the Decla-ration of Helsinki, the Belmont Report, and the Nuremberg Code Consent for Publication. All participants included in the study provided informed consent to the publishing of their de-identified study data.
350
O. Cano-Garcia et al.
References 1. Ardila, A., Bernal, B., Rosselli, M.: How localized are language brain areas? A review of Brodmann areas involvement in oral language. Arch. Clin. Neuropsychol. 31(1), 112–122 (2015) 2. Funciones de la corteza cerebral: áreas de Wernicke, Broca y de asociación límbica. https:// www.elsevier.com/es-es/connect/medicina/edu-funciones-de-la-corteza-cerebral#:~:text=% C3%81rea%20de%20Wernicke%3A%20es%20el,circunvoluci%C3%B3n%20superior% 20del%20l%C3%B3bulo%20temporal. Accessed 08 June 2022 3. Binder, J.: The Wernicke area. Neurology 85(24), 2170–2175 (2015) 4. Shafaei-Bajestan, E., Harald Baayen, R.: Wide learning for auditory comprehension. Interspeech, 966–970 (2018) 5. Naue, N., et al.: Auditory event-related response in visual cortex modulates subsequent visual responses in humans. J. Neurosci. 31(21), 7729–7736 (2011) 6. Huth, A., Heer, W., Griffiths, T., Theunissen, F., Gallant, J.: Natural speech reveals the semantic maps that tile human cerebral cortex. Nature 532(7600), 453–458 (2016) 7. Broderick, M., Anderson, A., Lalor, E.: Semantic context enhances the early aunditory encoding of natural speech. J. Neurosci. 39(38), 7564–7575 (2019) 8. Alcaraz, V., Granados, D., Hernández, T., Carrasco, L.: Potenciales cerebrales relacionados a eventos (PRE) y mapeo cerebral durante la presentación de palabras abstractas y concretas. Revista Neuropsicología Latinoamerica 9(3), 7–20 (2017) 9. Da Silva, F.: EEG and MEG: relevance to neuroscience. Neuron 80(5), 1112–1128 (2013) 10. Chowdhary, K.R.: Introducing artificial intelligence. In: Chowdhary, K.R. (ed.) Fundamentals of Artificial Intelligence, pp. 1–23. Springer, New Delhi (2020). https://doi.org/10.1007/97881-322-3972-7_1 11. Bulat, L., Clark, S., Shutovabrahimian, E.: Speaking, seeing, understanding: correlating semantic models with conceptual representation in the brain. In: Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, single volume, Copenhagen, pp. 1081–1091. Association for Computational Linguistics (2017) 12. Hollenstein, N., et al.: Decoding EEG brain activity for multi-modal natural language processing. Front. Hum. Neurosc., 15 (2021) 13. Mikolov, T., Zweig, G.: Context dependent recurrent neural network language model. In: IEEE Spoken Language Technology Workshop (SLT 2012), pp. 234–239. Institute of Electrical and Electronics Engineers, Miami (2012) 14. Mahajan, R., Morshed, B.: Unsupervised eye blink artifact denoising of EEG data with modified multiscale sample entropy, kurtosis and wavelet-ICA. IEEE J. Biomed. Health Inform. 19(1), 158–165 (2015) 15. Delorme, A., Makeig, S.: EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J. Neurosci. Methods 134(1), 9–21 (2004) 16. Sahonero-Alvarez, G.: A comparison of SOBI, FastICA, JADE and infomax algorithms. In: Callaos, N., Horne, J., Sánchez, B., Savoie, M., Tremante, A. (eds.) IMCIC 2017, Proceedings of the 8th International Multi-Conference on Complexity, Informatics and Cybernetics, Single Volume, pp. 17–22. International Institute of Informatics and Systemics, Orlando (2017) 17. Al-Qazzaz, N., Hamid, S., Islam, M., Escudero, J.: Selection of Mother Wavelet Functions for Multi.Channel EEG Signal Analysis during a Working Memory Task. Sensors 15(11), 29015–29035 (2015) 18. Jiang, J., Chao, C., Chiu, M., Lee, R., Tseng, C., Lin, R.: An automatic analysis method for detecting and eliminating ECG artifacts in EEG. Comput. Biol. Med. 37(11), 1600–1671 (2007)
Brain Mapping: Location of the Words Through EEG
351
19. Buzsáki, G., Watson, B.: Brain rhythms and neural syntax: implications for efficient coding of cognitive content and neuropsychiatric disease. Dialogues Clin. Neurosci. 14(4), 345–367 (2012) 20. Brust, H., et al.: Desincronización/sincronización de ritmos EEG laterales en la habituación a la fotoestimulación en adultos. Rev. Invest. Clin. 65(5), 436–444 (2013) 21. Ebrahimian, M., Razegui, M., Zamani, A., Bagheri, Z., Rastegar, K., Motealleh, A.: Does high frequency transcutaneous electrical nerve stimulation (TENS) affect EEG gamma band activity? J. Biomed. Phys. Eng. 8(3), 271–280 (2018) 22. Delogu, F., Brouwer, H., Crocker, M.: Event-related potentials index lexical retrieval (N400) and integration (P600) during language comprehension. Brain Cognit. 135, 103569 (2019) 23. Wlotko, E., Lee, C., Federmeier, K.: Language of the aging brain: event-related potential studies of comprehension in older adults. Lang. Linguist. Compass 4(8), 623–638 (2010) 24. Harpale, V., Bairagi, V.: Time and frequency domain analysis of EEG signals for seizure detection: a review. In: 2016 International Conference on Microelectronics, Computing and Communications (MicroCom), pp. 1–6. Institute of Electrical and Electronics Engineers, Durgapur (2016) 25. Tadel, F., Baillet, S., Mosher, J., Pantazis, D., Leahy, R.: Brainstorm: a user-friendly application for MEG/EEG analysis. Comput. Intell. Neurosci. 2011 (2011) 26. Anderson, A., et al.: Predicting neural activity patterns associated with sentences using a neurobiologically motivated model of semantic representation. Cereb. Cortex 27(9), 4379– 4395 (2017)
Processing of Biomedical Images
Artifacts Generated by the 3D Rotation of a Freely-Swimming Human Sperm in the Measurement of Intracellular Ca2+ Andrés Bribiesca-Sánchez , Fernando Montoya , Ana Laura González-Cota , Paul Hernández-Herrera , Alberto Darszon , and Gabriel Corkidi(B) Instituto de Biotecnología, Universidad Nacional Autónoma de México, 62200 Cuernavaca, Morelos, México [email protected]
Abstract. Intracellular calcium [Ca2+ ]i is key to many sperm functions including swimming, locating the egg, and fertilizing it. However, the 3D motion of the cell complicates measuring the [Ca2+ ]i of freely-swimming cells, mainly because a sperm rotates on its axis of movement as it swims (head and flagellum rolling), which may add spurious fluctuations (artifacts) to the fluorescence images. Here we implement an experimental device and software that for the first time allow quantifying the effect of the head spin on the [Ca2+ ]i signal. The device is a 3D + t multi-focal-plane optical microscope that permits simultaneous recording of bright-field and fluorescence images. The scripts we developed allow measurement of the head’s spin time series from the bright-field images and the [Ca2+ ]i signal from the fluorescence images. We found that the fluorescence signal of cells labeled with a [Ca2+ ]i -insensitive dye significantly resemble the rotation signal, while cells with a calcium-sensitive dye have a high-frequency harmonic related to the beat of the flagellum. Keywords: 3D sperm motility · Sperm rotation · Intracellular calcium · 3D microscopy · Digital image processing
1 Introduction Sexual reproduction is a complex phenomenon in which a spermatozoon must detect the female gamete, swim towards it and fuse with it, overcoming every obstacle along the way. To accomplish its goal, the cell propels itself combining the fast beat of its tail (~25 Hz in humans) and a slower (~5 Hz) corkscrew-like rotation along its movement axis [9]. The intracellular calcium ([Ca2+ ]i ) of a sperm is a key signaling element that allows it to locate the egg, fertilize it, and self-regulate its swim [1–3]. The recently discovered correlation between [Ca2+ ]i signaling patterns and the success of in vitro fertilization has further encouraged its analysis [15], however, the fast and three-dimensional nature of sperm swimming and the physiological processes originating it constantly demand microscopy and computer vision innovations. To track the cell’s [Ca2+ ]i concentration, it must be loaded with a fluorescent Ca2+ -sensitive dye, but the light that a camera can © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 355–362, 2023. https://doi.org/10.1007/978-3-031-18256-3_39
356
A. Bribiesca-Sánchez et al.
detect at a high enough sampling frequency for 3D fluorescence acquisition is very low, making it very difficult to segment. On the other hand, because the sperm’s head does not have rotational symmetry, there are oscillations in the fluorescence signals which are due to the cell’s physical movement rather than its [Ca2+ ]i signal. Since the effect that the 3D rotation has on the fluorescence signal has not been effectively identified and eliminated to date, the relationship between the [Ca2+ ]i concentration and the behavior of a freely-swimming sperm remains ill-defined. Isolating the calcium signal is essential to understand the regulation of sperm function by this ion, which might lead to the development of effective treatments for infertility, species preservation, as well as male-gamete targeted contraceptives. The analysis of sperm motility and the biochemical processes related to it have been mostly confined to two dimensions due to the challenges of recording three-dimensional data of the flagellar beat. In the early 90s, Suarez et al. [4] found that there are rapid oscillations of intracellular Ca2+ which are correlated with the flagellar beat of mouse sperm. They dyed the sperm with a radiometric fluorescent dye that was sensitive to [Ca2+ ]i (indo-1) and recorded individual cells. Crucially, the cell’s head was tethered to a glass substrate, suppressing its ability to rotate. The experimental setup consisted of a stroboscope equipped with two high-speed CCD cameras: one for bright-field observations, used to analyze the beat of the flagellum, and another one to record fluorescence images. Samples were set in an observation chamber small enough to constrain their movement to the focal depth of the optics. Due to the difficulty of simultaneously monitoring the beat of the flagellum and its biological signals, it took more than 20 years for Suarez’s observations to be confirmed in other species. In 2017, we found that these oscillations are also present in human sperm using a 3D + t microscopy system that is based on an inverted optical microscope with a piezoelectric device attached to its objective that makes it vibrate on the z-axis at a set frequency and amplitude so that instead of recording a single focal plane, images are recorded at different heights [5, 6]. The regions of the flagellum focused on each plane were enhanced using a one-class classification algorithm [7], then the 3D flagellum centerline at each time was segmented using a fast-marching algorithm that selects the best minimal path [8]. Some samples were loaded with fluo-4, which is sensitive to [Ca2+ ]i , and the rest were stained with calcein, which is not. Hence, the latter cells serve as a control to detect artifacts produced by the 3D flagellar beat of the spermatozoon, even though the sperm’s head was attached to a coverslip. The development of hardware and software for the analysis of [Ca2+ ]i of freely-swimming spermatozoa is necessary since they portray a more accurate representation of a cell that can fertilize an egg. In this work we present a novel experimental device and software that allow for the first time to quantify the effect of a free-swimming sperm’s head rotation in the measurement of its fluorescence. The experimental setup is an updated version of the one we introduced for multi-focal plane acquisition [6], the key difference being that the device can now record bright-field and fluorescence images simultaneously, which permits supplementing the biological signal with the 3D movement of the sperm and identify components in the [Ca2+ ]i measurements that are induced by the cell’s rotation. We recorded individual freely-swimming human spermatozoa dyed with calcein as a
Artifacts Generated by the 3D Rotation of a Freely-Swimming Human Sperm
357
control to measure the fluctuations their rotation introduces on the fluorescence measurement and with fluo-8 to analyze the [Ca2+ ]i . We found matching harmonics in the fluorescence and rotation signals of calcein and fluo-8-loaded gametes. The developed instrument and software will likely allow us to remove the rotation artifact and monitor the [Ca2+ ]i signal of freely-swimming human sperm.
2 Materials and Methods 2.1 Sperm Preparations Cells were obtained from healthy donors. Highly motile sperm were separated from the rest by the swim-up technique in HTF medium and processed as described in [9]. Half of the samples were loaded with calcein and the rest with fluo-8. Thus, the former served as a baseline measurement of any fluctuations not originating from the sperm’s [Ca2+ ]i . 2.2 Experimental Setup The experimental device consists of an inverted microscope (Olympus IX71) equipped with a 49011 Fluo cube Filter (Chroma Technology Corporation, USA) and a highintensity excitation LED M490D2 (Thorlabs, USA). A piezoelectric device PI-725 (Physik Instruments, USA) supports the objective (Olympus UPlanSApo 100x/1.4 na oil objective) and makes it vibrate vertically while images of the sample are being recorded. The piezo-objective device was controlled by an E-517 display interface that fed it an 80 Hz ramp through an E-501 servo-controller through an E-505 high-current amplifier. The experimental device was set on a TMC optic table to remove external vibrations. Two high-speed cameras were used simultaneously to record 8-bit grayscale images of the sample. The first one was a MEMRECAM Q1v (Nac, USA) capable of recording up to 8000 fps (640 × 480 px). This camera was connected to an unfiltered microscope port to record bright-field images. Concurrently, a CamRecord 5000 x2 BW (Optronis, Deutchland) capable of recording up to 5000 fps (512 × 512 px) was used to record the microscope’s fluorescence output. Due to the small number of photons captured at 5000 fps by this camera a C9016-04 image intensifier with a C4412-01 booster unit (Hamamatsu, Japan) was attached to it. During each experiment, an individual cell freely-swimming in an imaging chamber was recorded for 3–4 s until the camera’s RAM was filled (up to 27,000 images). The objective oscillated at 80 Hz and 20 µm while the cameras recorded at their respective maximum sampling frequency, yielding 50 bright-field focal planes with a mean zresolution of 0.4 µm and 31 fluorescence planes with a 0.64 µm average resolution. The device was controlled by software we designed in C# which synchronizes the cameras and piezoelectric device. It configures the piezoelectric device through a USB 3.0 interface, while the cameras are synchronized and triggered via a digital/analogic converter (NI 6211-USB, National Instruments). Through this converter, the software also digitizes the camera trigger, their synchronization signals, and the piezoelectric signal -which is directly proportional to the objective’s z-coordinate- to store the height at which each image was recorded. The software saves two text files with the heights of
358
A. Bribiesca-Sánchez et al.
the sequence of TIF images recorded in each camera. The experiments were conducted on an Intel Core (TM) i7-6700 CPU @ 3.4 GHz, 32 Gb RAM processor (Intel Corporation, USA). Figure 1 shows a diagram of the experimental setup.
Fig. 1. Block diagram of experimental setup.
2.3 Image Processing The images recorded with each camera were reorganized into a 4D (3D+t) TIF hyperstack using homemade software that reconstructs the focal planes using their respective zcoordinate text file (see [10]). This format allows us to observe and apply filters to the images along the z and t dimensions independently. Detection and Analysis of the Sperm’s Head Rotation A subset of focal planes containing the sperm’s head over the experiment’s duration was projected onto one using ImageJ2’s average z-projection tool [11] to simplify its segmentation. A 3 × 3 gaussian blur and a contrast enhancement were applied to reduce background noise and close small holes in the cell. Finally, the image was binarized using
Artifacts Generated by the 3D Rotation of a Freely-Swimming Human Sperm
359
a threshold intensity of ~50 to segment the cell’s flattened head. A Python 3.7 script was developed to analyze the cell’s rotation by fitting an ellipse to each binary image using Scikit-image [12] so that the major axis passes from its neck to the acrosomal cap, thus its minor axis can be used to measure the adjacent side of the sperm. Figure 2a and 2b show a bright-field image of a spermatozoon with its head in a horizontal position (parallel to xy-plane) and the binary image obtained from it with the computed ellipse’s semi-axes. Figure 2c and 2d show the same cell and 0.07 s later, when the head has rotated 90° (perpendicular to the xy-plane). The obtained time series was smoothened using a moving average with a window of size 3 and its power spectrum was computed using NumPy’s FFT module [13].
Fig. 2. Segmentation and measurement of the rotating sperm head. (a–b) Bright field image of a sperm whose head is horizontal and its segmentation. (c–d) Bright-field-image and segmentation of the same gamete 0.07 s later, when the sperm has rotated 90° so the head is vertical. The green dot marks the ellipse center, and the red lines are its major and minor hemi-axes.
Measurement of Head Fluorescence Fluorescence images were processed using ImageJ2 [11] tools as follows: First, the hyperstack was projected along the z-axis (Z Project with average intensity) to generate a 2D reconstruction of the multi-focal-plane fluorescent head. Then, the image was smoothened using a gaussian 2 × 2 kernel. The threshold was adjusted to delete pixels with a brightness value below 70, removing background noise. Finally, we used the Particle Analyzer tool [14] to obtain the head’s average brightness over time. The signal was smoothened through a moving average, then a least mean square 5th degree polynomial was fit and subtracted from it to reduce low frequency noise. Their power spectrum was computed using the magnitude of NumPy’s FFT module [13] to obtain the signals dominant frequencies.
3 Results The full rotation of a freely-swimming spermatozoon recorded simultaneously in fluorescence and bright-field microscopy is displayed in Fig. 3. Figure 4 shows the rotation and fluorescence measurements of freely-swimming calcein-loaded sperm. With a rotation frequency of around 3.1 Hz. Figure 4a shows the head rotation signal of a
360
A. Bribiesca-Sánchez et al.
calcein-dyed sperm and Fig. 4b shows its power spectrum. This signal has a secondary harmonic matching the rotation frequency and a main harmonic that doubles it, which is to be expected since the length of the fitted ellipse’s minor axis is the same size after the sperm turns 180°. Figure 4c and 4d show the fluorescence time series and power spectrum of the same cell. Because calcein is insensitive to [Ca2+ ]i any fluctuations are due to optical artifacts. Crucially, its main frequential components match those of the rotation signal.
Fig. 3. Sequence of images containing a human spermatozoon’s full rotation elapsing 0.31 s. Bright-field (top row) and fluorescence (bottom row) images. The blue curves are an overlay enclosing the region where fluorescence was measured.
Fig. 4. Calcein-dyed sperm signals. (a) Rotation signal over a 1.2 s time interval, (b) power spectrum of the rotation signal, (c) fluorescence time series, and (d) power spectrum of fluorescence signal.
Figure 5 shows the rotation and fluorescence signals of a sperm loaded with fluo-8 that has a rotation frequency of 4.4 Hz. Figure 5a shows the sperm’s rotation signal, and Fig. 5b contains its power spectrum which has two dominant harmonics, matching the
Artifacts Generated by the 3D Rotation of a Freely-Swimming Human Sperm
361
frequency of the sperm head’s full spin and half spin. Because this cell is tinted with a [Ca2+ ]i -sensitive dye, the fluorescence signal (Fig. 5c) is comprised of the [Ca2+ ]i signal and the rotation artifact. Because of this, its power spectrum (Fig. 5d) shares harmonic frequencies with the rotation signal at 4.4 and 8.8 Hz, however there is a high frequency harmonic that is synchronized with the beat of the flagellum (25 Hz) that seems to originate from its [Ca2+ ]i signal since it is not present in the fluorescence signals of calcein-loaded gametes.
Fig. 5. Signals and power spectra of a cell loaded with fluo-8. (a) Rotation signal over a 1.2 s time interval, (b) power spectrum of the rotation signal, (c) fluorescence time series, and (d) power spectrum of fluorescence signal.
4 Conclusions Intracellular calcium [Ca2+ ]i is an essential part of a sperm’s signaling pathways that regulates its ability to move, detect other cells, and fertilize the egg [1–3]. The 3D motility dynamics of the cell have hindered our ability to study [Ca2+ ]i and its effect on freelyswimming gametes. The setup and software we devised allowed us to detect a fluctuation in the fluorescence signal that is produced by the cell’s 3D rotation rather than its inner physicochemical dynamics. Furthermore, these tools permitted the detection of a highfrequency [Ca2+ ]i wave in the sperm head that seems to be mostly biological, related to the flagellar beat. These tools will likely be helpful to understand the internal mechanisms that allow sperm to fertilize an egg and contribute to the development of a platform for semen sample quality assessment incorporating [Ca2+ ]i signal features crucial for the detection and treatment of male subfertility, the development of motility-targeted contraceptives, and animal breeding.
362
A. Bribiesca-Sánchez et al.
Acknowledgements. We gratefully acknowledge financial support from CONACYT (PhD scholarship to ABS) and DGAPA PAPIIT IN105222.
References 1. Torrezan-Nitao, E., Brown, S.G., Mata-Martínez, E., Treviño, C.L., Barratt, C., Publicover, S.: [Ca2+]i oscillations in human sperm are triggered in the flagellum by membrane potentialsensitive activity of CatSper. Hum. Reprod. 36(2), 293–304 (2021). https://doi.org/10.1093/ humrep/deaa302 2. Mata-Martínez, E., et al.: Role of calcium oscillations in sperm physiology. BioSystems 209, 104524 (2021). https://doi.org/10.1016/j.biosystems.2021.104524 3. Darszon, A., Nishigaki, T., Beltrán, C., Treviño, C.L.: Calcium channels in the development, maturation, and function of spermatozoa. Physiol. Rev. 91(4), 1305–1355 (2011). https://doi. org/10.1152/physrev.00028.2010 4. Suarez, S.S., Varosi, S.M., Dai, X.: Intracellular calcium increases with hyperactivation in intact, moving hamster sperm and oscillates with the flagellar beat cycle. Proc. Natl. Acad. Sci. U.S.A. 90(10), 4660–4664 (1993). https://doi.org/10.1073/pnas.90.10.4660 5. Corkidi, G., et al.: Are there intracellular Ca2+ oscillations correlated with flagellar beating in human sperm? A three vs. two-dimensional analysis. Mol. Hum. Reprod. 23(9), 583–593 (2017). https://doi.org/10.1093/molehr/gax039 6. Silva-Villalobos, F., Pimentel, J.A., Darszon, A., Corkidi, G.: Imaging of the 3D dynamics of flagellar beating in human sperm. In: 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2014, pp. 190–193. Institute of Electrical and Electronics Engineers Inc. (2014). https://doi.org/10.1109/EMBC.2014.694 3561 7. Hernandez-Herrera, P., Papadakis, M., Kakadiaris, I.A.: Multi-scale segmentation of neurons based on one-class classification. J. Neurosci. Methods 266, 94–106 (2016). https://doi.org/ 10.1016/j.jneumeth.2016.03.019 8. Hernandez-Herrera, P., Montoya, F., Rendon-Mancha, J.M., Darszon, A., Corkidi, G.: 3D+t human sperm flagellum tracing in low SNR fluorescence images. IEEE Trans. Med. Imaging. 10, 2236–2247 (2018). http://www.ncbi.nlm.nih.gov/pubmed/29993713 9. Corkidi, G., Hernandez-Herrera, P., Montoya, F., Gadelha, H., Darszon, A.: Long-term segmentation-free assessment of head-flagellum movement and intracellular calcium in swimming human sperm. J. Cell Sci. 134(3) (2021). https://doi.org/10.1242/jcs.250654 10. Bribiesca, A., Montoya, F., Hernández, P., Ramos, E., Corkidi, G.: Device for experimental characterization of the 4D flow inside an evaporating sessile water droplet. Rev. Sci. Instrum. 91(1), 16101 (2020). https://doi.org/10.1063/1.5126112 11. Rueden, C.T., et al.: ImageJ2: ImageJ for the next generation of scientific image data. BMC Bioinform. 18(1) (2017). https://doi.org/10.1186/s12859-017-1934-z 12. Van Der Walt, S., Schönberger, J.L., Nunez-Iglesias, J., et al.: Scikit-image: image processing in python. PeerJ. 2014(1) (2014). https://doi.org/10.7717/peerj.453 13. Harris, C.R., et al.: Array programming with NumPy. Nature 585, 357–362 (2020). https:// doi.org/10.1038/s41586-020-2649-2 14. Sbalzarini, I.F., Koumoutsakos, P.: Feature point tracking and trajectory analysis for video imaging in cell biology. J. Struct. Biol. 151(2), 182–195 (2005). https://doi.org/10.1016/j.jsb. 2005.06.002 15. Kelly, M.C., et al.: Single-cell analysis of [Ca2+]i signalling in sub-fertile men: characteristics and relation to fertilization outcome. Hum. Reprod. 33(6), 1023–1033 (2018). https://doi.org/ 10.1093/HUMREP/DEY096
Morphological Temporal Analysis in Subjects with Alzheimer’s Disease by Brain Graph Descriptors Laura Gonzalez–Meza1(B) , Jesus Siqueiros–Garcia2 , Nidiyare Hevia–Montiel2 , Jos´e Javier Reyes–Lagos1 , and Jorge Perez–Gonzalez2 1
2
Facultad de Medicina, Universidad Aut´ onoma Del Estado de M´exico, Toluca de Lerdo, 50180 Estado de M´exico, M´exico [email protected] Unidad Acad´emica del Instituto de Investigaciones en Matem´ aticas Aplicadas y en Sistemas en el Estado de Yucat´ an, Universidad Nacional Aut´ onoma de M´exico, 97302 M´erida, Yucat´ an, M´exico
Abstract. In this paper we present a new approach to morphological analysis of magnetic resonance brain images based on graphs. The construction of the graphs is based on the segmentation of brain subregions and measurement of two invariant metrics: discrete compactness and discrete tortuosity. The aim is to construct a graph for a subject at different time instants and quantify the brain changes using network characteristics. When a temporal graph analysis is applied to control subjects, patients with mild cognitive impairment and patients with Alzheimer’s disease (AD), decreases in average degree (up to 25%) and average closeness (up to 28%) were found. Preliminary results suggest that the proposed analysis may be useful in the study and follow-up of patients with AD, as a diagnostic tool or as features in an automatic classification strategy. Keywords: Brain magnetic resonance imaging · Discrete compactness · Discrete tortuosity · Graph analysis
1
Introduction
Structural Magnetic Resonance Imaging (MRI) is one of the most accepted techniques both in research and in clinical routine, for clinical diagnosis of Alzheimer’s disease (AD) and treatment of brain atrophy by morphometric measurements. It allows observing tissue alterations by providing static anatomical information [1]. However, it is necessary to propose new methodologies for the study of interactions between brain substructures, as well as for the analysis of brain morphological changes and their connection between regions. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 363–370, 2023. https://doi.org/10.1007/978-3-031-18256-3_40
364
L. Gonzalez–Meza et al.
In the current literature there are a variety of studies that have investigated the usefulness of the graph theory method for the evaluation of neural networks at different stages of AD [2]. Cortical networks have been constructed from gray matter volumes obtained from MRI of all brain regions where the highest clustering coefficient and highest absolute path length are found in AD, indicating possibly less optimal cortical network organization [3]. It has also been mentioned that gray matter graphs can provide a quantitative description of brain morphological alterations, which may be useful in clinical practice [4]. In contrast to the works described above, this paper presents the temporal analysis based on graphs of Healthy Controls (HC), Mild Cognitive Impairment (MCI) and AD subjects. The proposed analysis is based on segmentation and extraction of morphological features for each brain substructure using T1–MRI. For each subject analyzed, an individual graph is obtained from the morphological relationships of each substructure. The objective is to construct a graph for each subject at different time instants and quantify the changes using network characteristics.
2 2.1
Methodology T1–MRI Database
A set of T1-weighted MR images of HC, MCI and AD patients were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu) [5]. All images were acquired on 1.5 T equipment with 1 mm3 of spatial resolution. To observe the temporal changes in brain tissue, a total of 9 subjects (3 from each population) were analyzed. Each subject has a total of 4 T1–MRI studies acquired at different time instants (periods of 6 months to one year between one acquisition and the next). The characteristics of each group are described in Table 1. The proposed analysis is a temporal study for each individual subject and not by group. Table 1. Characteristics of the study subjects. Characteristics Age (range in years) Gender (F/M) MMSE (rank) CDR (rank) Time between acquisitions (months) Number of acquisitions per subject
HC(N = 3) 75.1−89 1/2 30−26 0−0.5 11.8 ± 2.9 4
MCI (N = 3) 62.3−76.8 2/1 28−5 0.5−1 7.5 ± 3.3 4
AD (N = 3) 70.1−93 2/1 29−5 0.5−2 6.7 ± 4.1 4
HC:Healthy Controls; MCI: Mild Cognitive Impairment; AD:Alzheimer’s disease. MMSE: Mini-Mental State Examination; CDR: Clinical Dementia Rating.
Morphological Temporal Analysis by Brain Graph Descriptors
2.2
365
Automatic Brain Segmentation and Morphological Descriptors
Images were processed with FreeSurfer software, version 6.0 [6]. The labeling of the regions of interest was performed based on the Desikan-Killiany Atlas. For this work a total of 28 subcortical structures and 66 cortical regions were segmented. Once the brain substructures were segmented, the Discrete Compactness (DC) [7] and Discrete Tortuosity (DT) [8] were calculated. These metrics have already been used in previous work for the study and automatic classification of subjects with AD with good results [9,10]. Therefore, these metrics have been proposed for the construction of morphological brain graphs. 2.3
Morphological Brain Networks
The graphs were constructed based on the morphological relationship of each brain substructure. For each MRI study, two graphs were obtained (one based on DC and the other on DT). For each subject, a total of 94 substructures were measured; therefore, a vector of morphological metrics can be constructed as follows MK = {M1 , M2 , M3 , . . . , Mk } with k = 94. Now, the objective is to construct an adjacency matrix containing the morphological relationships of all the analyzed substructures (where each brain region represents a node). To do this, an auxiliary vector has been proposed MK = MJ = {M1 , M2 , M3 , . . . , Mj } with j = 94. Finally, the following equation has been proposed to compute the adjacency matrix (AM): ⎞ ⎛ 1 a12 a13 · · · a1k ⎜a21 1 a23 · · · a2k ⎟ ⎟ ⎜ ⎟ ⎜ (1) AM = 1 − |MK − MJ | = ⎜a31 a32 1 · · · a3k ⎟ , ⎟ ⎜ .. .. . . .. ⎠ ⎝ . . . . aj1 aj2 aj3 · · · 1 where a represents the edges (elements) of the adjacency matrix with size 94 × 94. Each element of the matrix has a range from 0 to 1, where 1 represents morphologically equal substructures and 0 totally different. As mentioned above, for each T1-MRI study, two adjacency matrices were obtained (one for DT and the other for DC). 2.4
Graph Descriptors and Statistical Analysis
A representation of the data as an undirected graph was made based on the adjacency matrix obtained. For each network/graph, the average degree, betweenness and closeness were obtained. For this work, each segmented brain substructure is considered as a node. To quantify the most representative connections of each network, a threshold of 98% was empirically selected. As described in Sect. 2.1, each subject has 4 MRI acquisitions at different time intervals. To observe the temporal changes
366
L. Gonzalez–Meza et al.
of each subject, the average degree, betweenness and closeness for each of their 4 brain studies have been calculated. This analysis was done for the DC and DT metrics. Finally, to evaluate the temporal changes of each subject and each graph metric obtained, a trend analysis was performed using the Mann-Kendall test (p ≤ 0.05).
3
Results and Discussion
Figure 1 shows a representative example of the segmentation of a subject with MCI. The 94 segmented brain substructures are presented in different colors. Figure 2 shows a representative example of the difference between a network of a 70.1 year old subject with AD and the same subject with 71.7 years of age. The red color shows the DC network and the blue color shows the DT network. To show the main connections, a threshold of 98% was set for both examples.
Fig. 1. Representative example of brain segmentation in an (a) axial, (b) saggital and (c) coronal view. Each color represents a segmented substructure.
According to the changes shown in Figs. 2a–b, some of the structures that presented major morphological alterations were the left and right inferior lateral ventricle, the right caudate, the left cuneus, the right superior parietal cortex and the right fusiform gyrus, which is consistent with published literature [11–13]. Figure 3 shows the temporal changes for the metrics degree, betweenness and closeness obtained from the graphs of a representative HC subject (in red), a patient with MCI (in blue) and another with AD (in green) for 4 time instants. In general, it can be observed that the degree and closeness metrics show a decrease; in contrast, the betweenness metric shows an increase. This same trend was observed for the rest of the subjects analyzed, however the Mann-Kendall test showed that there were no significant differences for the proposed metrics in any of the nine subjects studied. Nevertheless, the degree metric showed a significance of up to p ≤ 0.08 for both DC and DT analysis.
Morphological Temporal Analysis by Brain Graph Descriptors
367
Fig. 2. Representative examples of a brain network obtained with Discrete Compactness (a) and Discrete Tortuosity (b) metrics. The networks shown were obtained with a difference between the same subject with AD (Alzheimer’s Disease) at an age of 70.1 years and 71.7 years (to show temporal brain changes). CSF: Cerebrospinal fluid; Ctx: Cortex; lh: left hemisphere; rh: right hemisphere.
368
L. Gonzalez–Meza et al.
Fig. 3. Representative results for one subject from each population according to the graphs obtained from the DC (Discrete Compactness, first row) and DT (Discrete Tortuosity, second row). (a) and (d) are the plots of the degree characteristic; (b) and (e) are for the betweenness descriptor, and (c) and (f) correspond to the closeness index. The abscissa axis represents the four time instants when the MRIs (Magnetic Resonance Imaging) were acquired.
The results obtained are consistent with the alterations reported in brain networks proposed by other authors [2,4]. Although no statistically significant differences were obtained in the temporal analysis, the results show temporal changes which may help to quantify the degree of deterioration during the disease or help to predict the patient’s evolution.
4
Conclusions
In this paper, we present a new approach for the analysis of morphological changes of brain substructures based on network characteristics. The construction of the graphs is based on the segmentation of brain subregions using T1-MRI and measurement of two invariant metrics: discrete compactness and discrete tortuosity. When analyzing the temporal changes in a network (in a period of 1 to 2 years) of control subjects, patients with MCI and subjects with AD, a decrease in degree and closeness was found. Preliminary results are promising, however, it is necessary to increase the number of subjects analyzed, to evaluate more characteristics of the graphs and to evaluate different thresholds to select the most representative connections.
Morphological Temporal Analysis by Brain Graph Descriptors
369
In conclusion, the results suggest that the proposed approach can be useful to extract brain morphological indicators based on graph analysis, which can be useful in the study and evolution of AD. Acknowledgments. This work was supported by UNAM-PAPIIT programs: IT101422, IA104622, and CONACYT CF-2019–217367. Authors also wish to acknowledge technical assistance of Adrian Duran Chavesti for server management used for data processing.
Conflict of Interest. The authors declare that they have no conflict of interest.
References 1. del Pilar Rueda, A., Enr´ıquez, L.F.: Una revisi´ on de t´ecnicas b´ asicas de neuroimagen para el diagn´ ostico de enfermedades neurodegenerativas. Biosalud 17(2), 59–90 (2018) 2. He, Y., Chen, Z., Evans, A.: Structural insights into aberrant topological patterns of large-scale cortical networks in alzheimer’s disease. J. Neurosci. 28(18), 4756– 4766 (2008). https://doi.org/10.1523/JNEUROSCI.0141-08.2008 3. Yao, Z., Zhang, Y., Lin, L., Zhou, Y., Xu, C., Jiang, T., Initiative, A.D.N.: Abnormal cortical networks in mild cognitive impairment and alzheimer’s disease. PLoS Comput. Biol. 6(11), e1001,006 (2010). https://doi.org/10.1371/journal. pcbi.1001006 4. Tijms, B.M., et al.: Single-subject grey matter graphs in alzheimer’s disease. PloS one 8(3), e58,921 (2013). https://doi.org/10.1371/journal.pone.0058921 5. Petersen, R.C., et al.: Alzheimer’s disease neuroimaging initiative (adni). Neurology 74(3), 201–209 (2010). https://doi.org/10.1212/WNL. 0b013e3181cb3e25,http://n.neurology.org/content/74/3/201 6. Fischl, B., et al.: Automatically parcellating the human cerebral cortex. Cereb. Cortex 14(1), 11–22 (2004). https://doi.org/10.1093/cercor/bhg087 7. Perez-Gonzalez, J.L., Yanez-Suarez, O., Bribiesca, E., Cos´ıo, F.A., Jim´enez, J.R., Medina-Ba˜ nuelos, V.: Description and classification of normal and pathological aging processes based on brain magnetic resonance imaging morphology measures. J. Med. Imag. 1(3), 1–7 (2014). https://doi.org/10.1117/1.JMI.1.3.034002 8. Bribiesca, E.: A measure of tortuosity for enclosing surfaces of voxel-based objects. SN Comput. Sci. 2(3), 1–11 (2021). https://doi.org/10.1007/s42979-021-00565-0 9. Barbar´ a-Morales, E., P´erez-Gonz´ alez, J., Rojas-Saavedra, K.C., Medina-Ba˜ nuelos, V.: Evaluation of brain tortuosity measurement for the automatic multimodal classification of subjects with alzheimer’s disease. Comput. Intell. Neurosci. (2020) ´ 10. Perez-Gonzalez, J., Jim´enez-Angeles, L., Saavedra, K.R., Morales, E.B., MedinaBa˜ nuelos, V.: Mild cognitive impairment classification using combined structural and diffusion imaging biomarkers. Phys. Med. Amp Biol. 66(15), 155,010 (2021). https://doi.org/10.1088/1361-6560/ac0e77 11. Kumar, A., Sidhu, J., Goyal, A., Tsao, J.W.: Alzheimer Disease. StatPearls Publishing, Treasure Island (FL) (2020). ‘europepmc.org/books/NBK499922’
370
L. Gonzalez–Meza et al.
12. Guti´errez-Robledo, L.M., Arrieta-Cruz, I.: Demencias en m´exico: la necesidad de un plan de acci´ on. Gac. Med. Mex. 151(5), 667–673 (2015) 13. Sperling, R.A., et al.: Toward defining the preclinical stages of alzheimer’s disease: recommendations from the national institute on aging-alzheimer’s association workgroups on diagnostic guidelines for alzheimer’s disease. Alzheimer’s Dementia 7(3), 280–292 (2011)
PET Image Reconstruction Using a GRU-Convolutional Network Jose Mejia(B) , Boris Mederos , Leticia Ortega-M´ aynez , Nelly Gordillo , and Lidia Hortencia Rasc´on-Madrigal Departamento de Ingenier´ıa El´ectrica y Computaci´ on, Universidad Aut´ onoma de Ciudad Ju´ arez, 32310 Chihuahua, M´exico [email protected]
Abstract. Positron emission tomography is widely used for tumor detection and treatment monitoring in oncology. However, the quality of the images depends, among other factors, on the amount of radiopharmaceutical ingested by the patient. In this sense, the quality suffers degradation because there is a limit on the amount of radiation the patient can tolerate. Because of this, image reconstruction algorithms are required to generate images of adequate quality even if the amount of radiopharmaceutical to produce the image is small. In this study, a reconstruction algorithm is proposed based on deep learning using a GRU recurrent network which is expected to model the series of projections produced by the PET scanner as an input sequence to the recurrent network and is capable of reconstructing an image even with low amounts of the radiopharmaceutical. In comparisons using image quality metrics, our proposal achieves a SIMM of 0.95, outperforming other state-of-theart methods. Additionally, tests were performed for the evaluation of the task of lesion detection; the proposed method obtained a better contrast of the lesion with a value of 0.54 when using the weber contrast metric, very similar to the ground truth contrast of 0.55.
Keywords: PET
1
· Image reconstruction · GRU · Deep learning
Introduction
Positron emission tomography (PET) is widely used in clinical practice; the primary purpose of this imaging modality is to provide information about the metabolism inside the body and to detect and follow up treatment in oncology [1]. In this medical imaging modality, photons are detected caused by the iteration of the positrons of a radiopharmaceutical ingested or injected into the patient. However, due to the radiation limit that the patient can tolerate, the use of an elevated amount of radiotracer substance will expose patients to excessive radiation, harming their health. For this reason, it is necessary to design image reconstruction methods capable of obtaining quality images, using the minimum possible radioactive doses injected into the patient. Therefore, when the dose is c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 371–381, 2023. https://doi.org/10.1007/978-3-031-18256-3_41
372
J. Mejia et al.
reduced, the radiation registered by the detecting system is lower, meaning that the data collected will not be enough to reconstruct a quality image, introducing artifacts and noise. In different development phases, several reconstruction algorithms have been proposed to improve the reconstructed image quality. In the first phase, the analytical methods were extensively used; although efficient, appropriate sampling is required, examples of such methods are the use of Fourier methods and the back-projection (FBP) algorithms. The second phase of the algorithms used mainly were iterative methods, e.g., the expectation maximization and algebraic methods; such methods’ development were considered for modeling the scanner’s statistical and physical characteristics and imaging process. The third phase also takes into account the existing data and can learn from the data; most of these methods are based on deep learning; the strength of these methods is that from low count poor quality data, images can be reconstructed via models learned with similar data, and the limitations are related to the fact that the computation is not efficient and requires an extensive training data set [2]. In recent years, accumulated scientific work has shown that reconstruction methods based on the deep learning framework have improved image reconstruction quality qualitatively and quantitatively. Deep learning applications have increased because of the new advances in high-performance computational techniques, internet services such as cloud computing systems, and specialized hardware such as faster graphical processing units, which have increased the deployment of deep learning applications in various fields of medical imagenology including image reconstruction. However, these techniques’ high computational cost is an inconvenience, requiring a large number of training samples. The challenge of insufficient training datasets is generally faced using automatic generation of samples using generative techniques; and by the use of transfer learning, which focuses on the use of knowledge gained in a similar task and then solving the current problem by applying the former knowledge. The Deep learning-based methods have revealed that their reconstructed images present reduced noise levels, lower partial volume effects, less ringing, and well-defined edges, improving the general aspects of the resulting image. Several experimental results have been reported within the scientific literature. In [3] a deep learning network method was introduced to overcome the lack of an automated means for the optimization and computational expense; however, the limitation was that synthetic data was used instead of real patient data. The proposed method in [4] aims at reconstructing low counts of dynamic PET images. Such images have a high signal-to-noise ratio; they proposed to blindly decompose an image matrix into spatial and temporal elements; the authors show that their method performs better than conventional methods and can extract spatial factors representing homogeneous tissues. In another work [5], a stacked sparse auto-encoder for dynamic PET imaging was proposed; here, the encoding layers extract features while the reconstruction is done in the decoding layers. In [6] a model based on Wasserstein generative adversarial network is proposed. For training, the loss function combines a perceptual loss with Wasserstein distance.
PET Reconstruction
373
Also, in [7] they use a deep neural network in a supervised learning task searching for mapping between sensor and image domain to emerge. Other scientific work presented in [8] proposed the use of CNN for processing post-reconstruction images to reduce reconstruction artifacts while keeping the resolution recovery; results reveal lower noise and reduced ringing in the images treated with their method. This study proposes a reconstruction algorithm based on deep learning using a gated recurrent unit (GRU) and CNNs. The PET scanner’s projections series is modeled as an input sequence to the recurrent network. It is expected that the GRU internal state obtains a representation of a pre-reconstructed image, which is further elaborated by comparing it with an initial raw estimate of the final image and filtering through convolutional layers to finally obtain a reconstructed quality image. Thus in this work, we try to answer the following research question, is it possible to use the GRU as a reconstruction algorithm for images obtained from patients with a limited amount of radiopharmaceutical ingestion? We found that the proposed method works satisfactorily even when low-count raw data is used as input.
2 2.1
Theory The Sinogram
During a scan, the set of projections at different angles obtained by a scanner, is the Radon transform, R{f (x, y)}, of the scanned object f (x, y), where x and y are spatial coordinates [9]. This raw data is organized in the so-called sinogram, s(θ, d), where θ is the projection angle and d the distance of the projection line to the plane’s origin xy. For a given angle θ0 and a varying distance d, s(θ0 , d) is a set of projections at that angle. To reconstruct an image from this raw data, the inverse Radon transform R−1 {s(θ, d)}, can be implemented numerically using a fast algorithm to reconstruct the image of f from the data s (for example, using FBP). 2.2
Recurrent Networks
Recurrent neural networks (RNNs) are artificial neural networks capable of processing sequential or time series data. These are used in deep learning architectures as layers dealing with temporal problems, such as natural language processing and speech recognition. Among the RNNs, the Gated Recurrent Unit [10] has been shown to outperform other popular RNNs implementations such as the LSTM. In general, a RNN consists of a vector hidden state h and an output y(h), the RNN process an input sequence x = (x1 , ..., xN ). The hidden state is updated at each step n of the input sequence by hn = f (xn , hn−1 )
(1)
374
J. Mejia et al.
Fig. 1. Internal components of the GRU cell.
where f (·) depends on the type of RNN. In the case of the GRU, generally y(h) = h, for sequence to sequence modeling or y(h) = hN for a vector representation of the entire input sequence. The state hn is updated as ˜n hn = z hn−1 + (1 − z) h
(2)
where is element-wise (Hadamard) multiplication, z is the output of the ˜ is a state that depends of a reset gate r as follows update gate, and h ˜ n = φ(Wx x + Ux (r hn−1 )) h
(3)
and the gates are given by z = σ(Wz x + Uz hn−1 ),
(4)
r = σ(Wr x + Ur hn−1 )
(5)
where Wind and Uind are weight matrices which are learned. From equation (3) it can be seen that when r, the reset gate, is close to zero, the state only depends of the input x, thus resetting the state. From equation (2), it can be seen that control of how much information from the previous hidden state will translate to the new hidden state is done via a convex combination of the update gate. Figure 1 shows a graphical representation of the GRU. The most critical hyperparameters of this cell are the dimensionality of the output space; the σ(·) activation function, generally a sigmoid function; and the φ(·) activation function, where the hyperbolic tangent is frequently used.
3 3.1
Methods Proposed Method
This section describes the architecture proposed for PET image reconstruction from sinograms. The architecture consists of recurrent and convolutional layers,
PET Reconstruction
375
Fig. 2. PET scanner and sinogram generation.
as illustrated in Fig. 2. The input corresponds to sinogram, s(θ, d) and an initial reconstruction estimate of the image fˆ0 , in this case, the estimate is from the filtered back projection algorithm. The input sinogram is directed to the recurrent layer, given by a GRU, projections of the sinogram are taken as an input sequence, that is x = (s(θ1 , d), s(θ2 , d), ..., s(θN , d)), where N is the total number of projection angles. The intuition is that the GRU final state, hn summarizes the input sequence of projections as a pre-reconstructed image, which is further elaborated by comparing it with an initial raw estimate of the final image and filtering through convolutional layers. After processing the sinogram, the GRU output is fed to two convolutional layers without pooling; the first convolutional layer has 128 filters of size (20×20), the second layer has one filter of size (20×20), and its output is combined with the estimate fˆ0 using a subtraction layer. Subsequently, the output of the subtract layer is fed to a convolutional with 64 filters of size (2 × 2); finally, the output consists of a convolutional layer of one filter of size (2 × 2). Unlike [11], projections of the sinogram were taken instead of back-projected images because an individual projection, which is a one-dimensional vector, contains fewer elements than a back-projected, which has the size of the reconstructed image. This saves memory and reduces the computational cost. The peak signal-to-noise ratio (PSNR) was used as a loss function for training using the Adam optimizer and a batch size of 17; the number of epochs was around 500. Also, 80% of the data was used for training and 20% for testing; no validation data was used during training. The training was done on the colab (colab.research.google.com) environment using a Tesla P100-PCIE-16GB GPU and 13.6 gigabytes of RAM.
376
3.2
J. Mejia et al.
Phantom Simulation
The experiments to validate the proposed method were aimed at quantitatively evaluating the reconstruction for lesion detectability. For this end, a mouse phantom was used based on the digimouse project [12,13]. This phantom provides a detailed structure of a real mouse’s anatomy and was created by using computer tomography and cryosection images from a 28 g mouse, resulting in volumetric data of size 80 × 992 × 208. Also, segmented anatomical structures are provided, and more characteristics and details of its construction are in [12,13]. Since the original digimouse phantom is from a healthy individual, a lesion in the lung was simulated by adding a small sphere of 1 mm in radius. For PET simulation, the lesion was filled with an activity equivalent relative lesion-to-background radioactivity ratio of 4 : 1. The simulation of a PET scanner was done using the Simset software, which uses Monte Carlo algorithms to model the physics and instrumentation used in PET imaging [14]. A total of 100 simulations with low counts (20E 6 ) were carried out, each simulation consisting of 5 slices to give a total of 500 images; in each simulation, the phantom was rotated randomly for [−10, 10] degrees, and randomly moved horizontally and vertically in an interval of [−10, 10] voxels. As ground truth (GT) information, a simulation with high counts (160E 6 ) and reconstructed with the Expectation maximization algorithm was used. The simulations were carried out in an Intel Core i7 computer with 32 gigabytes of RAM and a Linux operating system.
4
Results
In this section are presented the numerical results to verify the proposed architecture. For comparisons, the following methods were used, FBP with ramp filter, the expectation maximization (EM) algorithm [15], and deep back projection (DBP) [11]. This last algorithm was initially designed to work with computed tomography (CT). However, it was included here as it uses projections for reconstruction, making it an algorithm very similar to the one proposed in this study. The experiment consisted in reconstructing sinograms from raw data with low counts (small amount of radiopharmaceutical), comparing the reconstruction with the GT image, and analyzing the area of the lesion region. The data used in the experiment was described in Sect. 3.2. Figure 3 shows the reconstruction results for each method. It can be seen that the proposed method has fewer artifacts and is perceptually more similar to the GT. Note also that the DBP algorithm has a lower performance than the proposed one; this could be since, in CT, there is less noise in the images, unlike PET, where the images have many artifacts, which could have affected the algorithm. This can be corroborated quantitatively in Table 1, where it is observed
PET Reconstruction
377
that the proposed method outperforms the other methods in the comparison, the metrics used were the peak signal-to-noise ratio (PSNR) [16], structural similarity (SSIM) [17], and normalized root mean square (NRMS) [18]. For lesion evaluation, Fig. 4 shows the lesion position on the phantom; the rectangles show the lesion and the background areas. Figure 5 shows the lesion surface for each method, the surface is taken from the lesion area of the lesion in the image reconstructed by each method. Once again, the proposed method shows more similarity with the GT in the lesion morphology than the other methods.
Fig. 3. Reconstructed images from the algorithms (a) FBP, (b) EM, (c) DBP, (d) proposed, and (e) GT.
378
J. Mejia et al. Table 1. Reconstruction metrics. PSNR FBP
SIM
NRMSE
10.75519 0.18638 0.80938
EM
27.8709
DBP
24.62008 0.7068
0.84908 0.11281 0.16402
Proposed 37.45186 0.9520
0.03743
Fig. 4. Lesion and background.
For the quantitative evaluation of the lesion contrast, two regions were selected, one within the lesion and the other in its vicinity, see Fig. 4, then the Weber (WC) and Michelson (MC) contrast metrics were calculated. These metrics are given by [19] I − Ib (6) WC = Ib where I is the lesion intensity, and Ib the background intensity. For the Michelson contrast Imax − Imin (7) MC = Imax + Imin where Imax and Imin are the maximum and minimum intensities respectively. Table 2, shows the values of contrast achieved by each algorithm. The proposed method achieves the best result for WC, followed by EM. EM has the best score for MC, even higher than the GT. Note, however, that the lesion area corresponding to EM, Fig. 5b, is less than the area of the GT, and it has a higher peak, which explains a high value of MC since this metric uses the maximum value of the area. Finally, from Fig. 5, it can be observed that the structure of the lesion reconstructed with the proposed method has the closest similarity in structure with the lesion in the GT; this is a good indicator as it could mean more accurate quantification of radiopharmaceutical uptake in the lesion zone.
PET Reconstruction
(a)
(b)
(c)
(d)
379
(e)
Fig. 5. Surface of the lesion reconstructed with (a) FBP, (b) EM, (c) DBP, (d) proposed method, (e) GT.
Table 2. Contrast measures. WC FBP
MC
0.260643 1.081633
EM
0.494905 2.206349
DBP
0.215012 0.674797
Proposed 0.548509 1.340909 GT
0.550393 1.215054
380
5
J. Mejia et al.
Conclusions
Because of the low quality of the PET images when a small dose of radiopharmaceutical is ingested, in this work, a reconstruction algorithm based on GRU is proposed to obtain higher-quality images even when using low doses of the radiopharmaceutical. In a experiment with a mouse phantom, the reconstruction quality was evaluated, and the proposed architecture outperformed the other methods in all the metrics used. In lesion detection, the proposed method obtains high WC results, close to GT; however, the MC value falls below the EM method, which exceeds GT’s real value. Perceptually, it can be observed that the structure of the lesion reconstructed with the proposed method is the closest to the structure of the lesion in the GT; this is a good indicator as it could mean that in tumor quantification studies, the proposed method will be more accurate than the other methods. For future work, it is suggested to train the architecture with real data; note that real data are challenging to obtain due to the use of radiopharmaceuticals that are expensive and usually have a limited duration due to the decay of the radioactive substance. Additionally, most commercial scanners only deliver the reconstructed image, not having access to the raw data (sinogram) before the reconstruction. Finally, it is suggested to use metrics to evaluate the morphology of the lesion to be able to compare with that of the GT.
References 1. Wernick, M.N., Aarsvold, J.N.: Emission Tomography: The Fundamentals of PET and SPECT. Elsevier, Amsterdam (2004) 2. Ahishakiye, E., Bastiaan Van Gijzen, M., Tumwiine, J., Wario, R., Obungoloch, J.: A survey on deep learning in medical image reconstruction. Intell. Med. 1(03), 118–127 (2021) 3. H¨ aggstr¨ om, I., Schmidtlein, C.R., Campanella, G., Fuchs, T.J.: Deeppet: a deep encoder-decoder network for directly solving the pet image reconstruction inverse problem. Med. Image Anal. 54, 253–262 (2019) 4. Yokota, T., Kawai, K., Sakata, M., Kimura, Y., Hontani, H.: Dynamic pet image reconstruction using nonnegative matrix factorization incorporated with deep image prior. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3126–3135 (2019) 5. Cui, J., Liu, X., Wang, Y., Liu, H.: Deep reconstruction model for dynamic pet images. PloS one 12(9), e0184667 (2017) 6. Hu, Z., et al.: Dpir-net: direct pet image reconstruction based on the wasserstein generative adversarial network. IEEE Trans. Radiat. Plasma Med. Sci. 5(1), 35–43 (2020) 7. Zhu, B., Liu, J.Z., Cauley, S.F., Rosen, B.R., Rosen, M.S.: Image reconstruction by domain-transform manifold learning. Nature 555(7697), 487–492 (2018) 8. da Costa-Luis, C.O., Reader, A.J.: Deep learning for suppression of resolutionrecovery artefacts in mlem pet image reconstruction. In: IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), pp. 1–3. IEEE (2017) 9. Dougherty, G.: Digital Image Processing for Medical Applications. Cambridge University Press, Cambridge (2009)
PET Reconstruction
381
10. Cho, K., et al.: Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014) 11. Ye, D.H., Buzzard, G.T., Ruby, M., Bouman, C.A.: Deep back projection for sparseview ct reconstruction. In: 2018 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pp. 1–5. IEEE (2018) 12. Dogdas, B., Stout, D., Chatziioannou, A.F., Leahy, R.M.: Digimouse: a 3D whole body mouse atlas from ct and cryosection data. Phys. Med. Biol. 52(3), 577 (2007) 13. Stout, D.: Creating a whole body digital mouse atlas with pet, ct and cryosection images. Mol. Imaging Biol. 4(4), S27 (2002) 14. Ljungberg, M., Strand, S.-E., King, M.A.: Monte Carlo calculations in nuclear medicine: Applications in diagnostic imaging. CRC Press, Boca Raton (2012) 15. Shepp, L.A., Vardi, Y.: Maximum likelihood reconstruction for emission tomography. IEEE Trans. Med. Imaging 1(2), 113–122 (1982) 16. Chen, L., Chen, Z., Singh, R.K., Vinu, R., Pu, J.: Increasing field of view and signal to noise ratio in the quantitative phase imaging with phase shifting holography based on the hanbury brown-twiss approach. Optics Lasers Eng. 148, 106771 (2022) 17. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process. 13(4), 600–612 (2004) 18. Stephen, K.D., Kazemi, A.: Improved normalization of time-lapse seismic data using normalized root mean square repeatability data to improve automatic production and seismic history matching in the nelson field. Geophys. Prospect. 62(5), 1009–1027 (2014) 19. Remeˇs, V., Haindl, M.: Region of interest contrast measures. Kybernetika 54(5), 978–990 (2018)
Characterization of COVID-19 Diseased Lung Tissue Based on Texture Features Jes´ us Gibr´ an Delgado-Alejandre1 , Diomar Enrique Rodr´ıguez-Obreg´on1 , Alejandro Santos-D´ıaz2,3 , and Aldo Rodrigo Mej´ıa-Rodr´ıguez1(B) 1
Facultad de Ciencias, Universidad Aut´ onoma de San Luis Potos´ı, San Luis Potos´ı, 78295 San Luis Potos´ı, M´exico [email protected] 2 Escuela de Ingenier´ıa y Ciencias, Instituto Tecnol´ ogico y de Estudios Superiores de Monterrey, Tlalpan, 14380 Ciudad de M´exico, M´exico 3 Escuela de Medicina y Ciencias de la Salud, Instituto Tecnol´ ogico y de Estudios Superiores de Monterrey, 64710 Monterrey, Nuevo Le´ on, M´exico
Abstract. Although real time polymerase chain reaction test (RTPCR) is the gold standard method for the diagnosis of COVID-19 patients, the use of Computed Tomography (CT) images for diagnosis, assessment of the severity of this disease and its evolution is widely accepted due to the possibility to observe the lungs damage. This evaluation is mainly made qualitatively, therefore, techniques have been proposed to obtain relevant additional clinical information, such as texture features. In this work, CT scans from 46 patients with COVID-19 were used to characterize the lungs by means of textural features. In the proposed approach, pulmonary parenchyma was delimited using a U-NET previously trained with images from different pulmonary diseases. Texture metrics were calculated using co-occurrence and run-length matrices considering both lungs, right and left lung, as well as apex, middle zone and base lung regions. A boxplot descriptive analysis was performed looking for significant differences between regions of each estimated texture metric. Results show that Gray Level Non-Uniformity (GLNU) and Run-Length Non-Uniformity (RLNU) features have more significant differences between regions, suggesting that these metrics may provide a proper characterization of the pulmonary damage caused by COVID-19.
Keywords: Texture features Lung tissue characterization
1
· COVID-19 · Computed tomography ·
Introduction
More than 6 million deaths have been registered all over the world caused by COVID-19 as of June 2022 [1]. The possibility of observe the damaged caused by COVID-19 in the lungs using CT imaging, has make this tool very c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 382–392, 2023. https://doi.org/10.1007/978-3-031-18256-3_42
Characterization of COVID-19 Diseased Lung Tissue
383
important to assess this disease. Automatic methods have been proposed to differentiate between COVID-19 disease and other pulmonary illnesses [2]. Nevertheless, there is considerably less information in the literature regarding the study of the pulmonary volumetric damage caused by COVID-19 in patients, or the spatial distribution of the disease inside the lungs, due to the challenge of identifying patterns of lesion on the image using only the gray level intensity information. We consider that relevant clinical information can be extracted from a spatial distribution study of COVID-19 in CT images, which may help to assess the severity of the disease and possibly elucidate patient prognosis. In particular, texture features have been used widely, to obtain complementary information from images, regarding the spatial distribution of gray scale values [3]. In this work, we analyzed how textural features behave on different regions of the lung and we proposed two features that could be more useful to differentiate lung regions in images from COVID-19 patients. This information could lead to a better comprehension of the disease and the development of artificial intelligence techniques that exploit these metrics.
2 2.1
Material and Methods Medical Images Dataset
In this work, COVID-19 CT images from 46 patients (53 ± 14 years) were used, provided by the intensive care unit at the Instituto Nacional de Ciencias M´edicas y Nutrici´ on Salvador Zubir´ an (INCMNSZ, Mexico City, Mexico) in NIfTI format. The used CT images were acquired using a lung window, with a matrix size of 512×512 pixels, a range of slices between 70 and 101, a pixel size ranging from 1.25 to 3.75 mm, and 1.25 mm spacing between slices. All medical images were acquired in March 2020, and for all the patients the presence of SARS-CoV-2 was confirmed by positive RT-PCR. The use of these CT images in the present research was approved by the institutional INCMNSZ ethics committee [4]. 2.2
Lung Tissue Segmentation
The public dataset PleThora [5], which contains 402 CT images and ground truth masks of the parenchyma from patients with different lung diseases, was used to train a U-net architecture network. Binary masks from pulmonary parenchyma were obtained from CT images of INCMNSZ’s patients using this previously trained network [4]. Figure 1 shows an example of the obtained segmented lung parenchyma.
384
J. G. Delgado-Alejandre et al.
Fig. 1. Axial CT slice on the left; segmented lung parenchyma image on the right.
2.3
Definition of Pulmonary Regions
Human lungs are heterogeneous organs, considering that right lung is divided into three lobes and is slightly larger than left lung, which has two lobes. In addition to these anatomical characteristics, the predominance of certain pulmonary diseases for apex (upper lobe) or base (lower lobe) have been recognized due to differences in perfusion-ventilation ratio, lymphatic flow, metabolism, and mechanics [6,7]. Therefore, in this work the idea is to know how texture features behave for different lung regions, meaning that it’s necessary to define a lung division by region in the CT images. Considering the anatomical nature of human lungs, the distribution of the CT slices would be unequal and medical validation to indicate the beginning-ending of each range would be needed. Since this isn’t a viable option, for each CT image the total slices were equally distributed in three segments assigned in order of appearance: apex, middle zone and base, respectively. Finally, 9 lung regions were defined as: BL (both lungs); RL (right lung); LL (left lung); RA, RMZ, RB (right apex, right middle zone, right base), and LA, LMZ, LB (left apex, left middle zone, left base). 2.4
GLCM and GLRLM Matrices Definition
Gray level Co-occurrence Matrix (GLCM) identifies the texture of an image by the frequency of pairs of pixels with a specific gray level. The pair of pixels are separated by a distance d specified by a vector d = (dr,dc) where dr and dc are the components of the distance in rows and columns, respectively, in a direction θ [8]. Figure 2 shows an example where two GLCM are obtained from the matrix on the left (Image I), the red ellipses indicate pairs of pixels, notice that one have the relationship d = (0, 1) and the other d = (1, 0). The arrows indicate the corresponding cell on the GLCM for that pair of gray level values.
Characterization of COVID-19 Diseased Lung Tissue
1
3
3
1
1
1
2
2
2
3
2
1
3
3
1
2
1
2
3
1
2
3
1
1
2
1
1
1
3
1
2
1
1
1
2
2
1
1
3
2
1
2
3
1
1
1
d=(0,1)
Image I
385
d=(1,0)
Co-ocurrence matrices
Fig. 2. Example of two GLCM on the right, for two different displacement vectors, obtained from the image matrix on the left (Image I).
On the other hand, Gray-Level Run-Length Matrix (GLRLM) identifies the texture of an image by the number of consecutive collinear pixels (better known as run-lengths) with a specific gray level along a given direction θ [9]. Displacement vector is defined in the same way for GLCM or using an angle of displacement, also a maximum run-length number is defined by the maximum between the height and width of the image. Figure 3 shows an example where two GLRLM are obtained from the matrix on the left (Image I), the red ellipses indicate runs of pixels of the same gray level, notice that one has the direction θ = 0◦ and the other θ = 90◦ . The arrows indicate the corresponding cell on the GLRLM for that run-length and gray level. On the GLRLM, the columns represent the run-lengths, and the rows are the gray levels of the image.
1
3
3
3
2
1
1
1
2
2
2
1
2
3
3
2
Image I
1
2
3
4
1
3
1
0
0
0
2
3
0
1
0
0
3
5
0
0
0
1
2
3
4
1
2
0
1
0
2
3
0
1
3
0
1
1
θ=0°
θ=90° Run length matrices
Fig. 3. Example of two GLRLM on the right, for two different directions θ, obtained from the image matrix on the left (Image I).
Procedure for estimation of GLCM and GLRLM matrices is the same for all 9 lung regions defined in this work. In order to avoid the background presented in every parenchyma slice, the values corresponding to zero were substituted by a NaN (Not a Number) value; therefore, at the moment of searching for pairs of
386
J. G. Delgado-Alejandre et al.
pixels that share a gray level value in common, these values are ignored. Each estimation is made for 4 angles: 0◦ , 45◦ , 90◦ y 135◦ . Initially NIFTI images have around five thousands gray values wich decreases to two thousands for parenchyma images due to parenchyma’s mask. Even with this decrease the number is still large to be processed so, in the order to get a faster processing of the images, the number of gray levels is set to 32. Higher values probably give us more accuracy in results but a lower processing. The maximum run-length value is set to 512 which corresponds to the dimension of parenchyma images. 2.5
Textural Features Estimation
Once GLCM and GLRLM have been obtained, texture features for each lung region are calculated. For GLCM, the most studied features were estimated, defined as follows [8]: Ngl Ngl p(i, j)2 (1) ASM = i=1 j=1 Ngl Ngl
Correlation =
p(i, j)
i=1 j=1
IDM =
Ngl Ngl i=1 j=1
Contrast =
(i − μi )(j − μj ) σi2 σj2
p(i, j) 1 + (i − j)2
Ngl Ngl
p(i, j)(i − j)2
(2)
(3)
(4)
i=1 j=1
where p(i, j) is the co-occurrence matrix, Ngl is the number of gray levels, μi , μj , σi , σj are mean and standard deviation of rows i, and columns j [8]. Angular Second Moment (ASM) measures the number of pixel pairs that are repeated, Contrast is a measure of local variation of pixel pairs, Inverse Difference Moment (IDM) measures the proximity between pixel pairs, and Correlation measures the lineal dependency of pixel pairs. For GLRLM, every existent feature associated with GLRLM will be used, and they are defined as follows [9]: Short Run Emphasis (SRE): Ngl Nrl p(i, j) 1 SRE = nr i=1 j=1 j 2
(5)
Long Run Emphasis (LRE): LRE =
Ngl Nrl 1 p(i, j) · j 2 nr i=1 j=1
(6)
Characterization of COVID-19 Diseased Lung Tissue
387
Gray-Level Non-Uniformity (GLNU): Ngl 1 GLN U = nr i=1
N rl
2 p(i, j)
(7)
j=1
Run-Length Non-Uniformity (RLNU): Nrl 1 RLN U = nr j=1
Ngl
2 p(i, j)
(8)
Ngl Nrl p(i, j) 1 LGRE = nr i=1 j=1 i2
(9)
i=1
Low Gray-Level Run Emphasis (LGRE):
High Gray-Level Run Emphasis (HGRE): Ngl Nrl 1 p(i, j) · i2 HGRE = nr i=1 j=1
(10)
Short Run Low Gray-Level Emphasis (SRLGE): Ngl Nrl p(i, j) 1 SRLGE = nr i=1 j=1 i2 · j 2
(11)
Short Run High Gray-Level Emphasis (SRHGE): Ngl Nrl p(i, j) · i2 1 SRHGE = nr i=1 j=1 j2
(12)
Long Run Low Gray-Level Emphasis (LRLGE): Ngl Nrl p(i, j) · j 2 1 LRLGE = nr i=1 j=1 i2
(13)
Long Run High Gray-Level Emphasis (LRHGE): Ngl Nrl 1 p(i, j) · i2 · j 2 LRLGE = nr i=1 j=1
(14)
where p(i, j) is the run-length matrix, Ngl is the number of gray levels, Nrl is the maximum number of run-lengths, and nr is the sum of all run-lengths (sum of all the cells in the matrix) [9].
388
J. G. Delgado-Alejandre et al.
2.6
Descriptive and Statistic Analysis
A descriptive analysis was performed, looking to compare whole lungs, individual lungs, and lung regions. This comparison permit us to have evidence of the lesion distribution in lungs and allows us to see if COVID-19 affects mainly a region or lobe of lungs. Furthermore, this analysis can expose the texture features that allow the best differentiation of the lung regions, which could be a useful feature for further analysis and the study of new algorithms to assist the diagnosis. For this reason, a boxplot components analysis was made to determine the presence of significant differences between regions for each texture metric. This analysis consists in the study of the boxplot components. These components are interquartile range (IQR) and the length of the whiskers, and they describe the usual range of values in the dataset and how far the values are from the IQR, respectively [10]. Shapiro-Wilk test was used to test normality in distributions, the significance level was set to p=0.05. To corroborate the boxplots results, a one-way ANOVA test and Kruskall-Wallis test were used in values with normal distribution and non-normal distribution, respectively, considering R R2018b (The Mathworks, Natick, MA, a p=0.05 of significance. MATLAB USA) was used to perform image pre-processing, texture features extraction, and descriptive and statistical analysis.
3
Results and Discussion
Table 1 presents the median values of the texture metrics for each one of the 9 lung regions defined. In this case, GLNU and RLNU show a big disparity of values between regions, compared to the rest of texture features, suggesting that these two texture metrics show that the affection of lungs caused by COVID19 is not evenly distributed and that a proper characterization could be done Table 1. Medians of each texture feature for the different lung regions. Bold indicate presence of disparity values between regions. BL (both lungs); RL (right lung); LL (left lung); RA (right apex), RMZ (right middle zone), RB (right base), LA (left apex), LMZ (left middle zone), LB (left base). Texture feature/Region BL
RL
RA
RMZ
RB
LL
LA
LMZ
LB
ASM
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
Contrast
5.57
6.30
6.00
5.01
8.49
6.29
5.97
4.99
7.64
IDM
0.59
0.57
0.58
0.60
0.53
0.57
0.57
0.59
0.55
Correlation
0.93
0.92
0.90
0.95
0.90
0.92
0.88
0.94
0.92
SRE
0.78
0.79
0.79
0.77
0.82
0.79
0.79
0.78
0.81
LRE
3.43
3.10
2.92
3.61
2.68
3.12
3.06
3.32
3.05
GLNU
873.65
514.85
503.38
755.23
247.35
424.45
492.83
555.08
261.19
RLNU
8139.90 5153.10 4545.00 7763.40 2908.20 4166.80 3726.70 5782.90 2942.40
LGRE
0.02
0.02
0.02
0.02
0.02
0.02
0.03
0.02
0.02
HGRE
226.15
234.05
168.38
215.38
279.68
220.67
147.06
222.46
308.38
SRLGE
0.01
0.01
0.02
0.01
0.02
0.02
0.02
0.01
0.01
SRHGE
172.78
182.67
145.77
169.95
221.30
175.66
123.31
173.72
232.00
LRLGE
0.07
0.08
0.10
0.08
0.05
0.07
0.09
0.05
0.04
LRHGE
849.40
814.10
401.10
847.70
1070.70
790.60
318.20
729.00
1204.30
Characterization of COVID-19 Diseased Lung Tissue
389
using these textural features. In addition, Table 2 presents all the significant differences found for each texture metric among the different pulmonary regions, showing complementary information regarding the behavior of the texture metrics throughout the lungs, and it corroborates that GLNU and RLNU are the two texture metrics with more significant differences between lung regions. Table 2. Significant differences between regions of each texture feature. BL (both lungs); RL (right lung); LL (left lung); RA (right apex), RMZ (right middle zone), RB (right base), LA (left apex), LMZ (left middle zone), LB (left base). Texture feature/Region BL ASM
Contrast
♠
IDM
♠
Correlation
♥
SRE
♠
RL
RA
♥
RMZ ♠
RB ♣♥ #
LL
LA
LMZ
♣
♠
♣♥
♣♥ ♠
♦ #
♥
♣♥
♠
♣♥
♥
♣♦ # ♠
♥
♥
♠
♣♥ #
LB ♣♥
♥
GLNU
♦ # ♠
♣♥ ♠
♣♥ ♠
♦ #♠
♣♦ ♥#
♣♥ ♠
♣♥ ♠
♣♠
♣♦ ♥#
RLNU
♦ # ♠
♣♥ ♠
♣♥ ♠
♦ # ♠
♣♦ ♥#
♣♥ ♠
♣♦♥
♣♥ # ♠
♣♦ ♥#
LGRE
♥
♣♦♥ # ♠
#
HGRE
♠
♠
♣♦ #♠
♠
♥
♠
♣♦♥ # ♠
♠
♣♦ ♥#
SRLGE
♥
#
♥
♣♦ ♥# ♠
♦ #
SRHGE
♠
♦♠
♠
♣♥
♠
♣♦♥ # ♠
♠
♣♥ #
LRLGE
♠
♠
♠
♠
♣♦ ♥#
♠
♠
♣♦ ♥#
LRHGE
♣♦♥ # ♠
♠
♣♦♥ # ♠
♠
#
LRE
♣: differences respect to BL ♦: differences respect to RL : differences respect to RA ♥: differences respect to RMZ : differences respect to RB #: differences respect to LL : differences respect to LA : differences respect to LMZ ♠: differences respect to LB
Descriptive analysis by means of boxplots of GLNU and RLNU is shown in Fig. 4, where is possible to see that both features present different IQR sizes between regions which indicates a different concentration of values in each region. Lower whiskers do not present changes of length between regions, while upper
390
J. G. Delgado-Alejandre et al.
whiskers do, meaning that an unequal distribution of data between regions exists. In addition, for both texture features is possible to observe a similar pattern between regions. The highest metric value in both cases corresponds to BL, indicating a considerable variation in gray level values and few run-lengths share the same gray level value, respectively; this may indicate that the disease is distributed in different areas of the lung and with different concentrations. GLNU
Textural feature value
2000
1500
1000
500
0
BL
RL
RA
RMZ
RB
LL
LA
LMZ
LB
LL
LA
LMZ
LB
Lung regions
RLNU 14000
Textural feature value
12000 10000 8000 6000 4000 2000 0
BL
RL
RA
RMZ
RB
Lung regions
Fig. 4. Comparative between GLNU’s values distribution and RLNU’s values distributions for each lung’s region
For the case of individual lungs (RL and LL) the opposite occurs having a considerably less value of GLNU and RLNU, showing few variations in gray level values and that run-lengths share the same gray level value, respectively. This may indicate that it’s necessary to explore the distribution of the regions that make up individual lungs. The comparison between regions of the same lung shows that in both left (LA, LMZ, and LB) and right (RA, RMZ, and RB) lungs, where the higher values for each lung correspond to the middle zone of lungs, follow by the apex, and with bases as the region with the lowest value. In particular, considering that bases have the least variation of gray level values and where the majority of run-lengths share the same gray level value for GNLU and RLNU respectively, they can be considered the most homogeneous
Characterization of COVID-19 Diseased Lung Tissue
391
regions in lungs. This may indicate that these are the regions where pulmonary tissue affected by COVID-19 is most concentrated because of lack of texture, taking into account that ground glass opacities and consolidations areas are more homogeneous (lighter gray values) than healthy areas (mix of dark and light gray values), see Fig. 1. Finally, the results found in this work support the idea that the affection of lung tissue by COVID-19 is not distributed evenly in lungs and provide evidence that relevant clinical information may be extracted from a characterization because it may provide insights to the clinical expert about regions of interest to analyze where the disease might be more concentrated [11]. Nevertheless, it’s worth to mention that, at this point, the results only suggest the presence of a non-uniformity distribution of the lung tissue affected by COVID-19 throughout the lungs, based on the presence of significant differences between regions. However, GLNU and RLNU are indicators of the non-uniformity of gray level values and run-lengths, respectively, but they do not indicate if a low GLNU or RLNU value means the presence or absence of disease in a particular region or vice versa. Therefore, a correlation between texture metric values and the presence of affected lung tissue validated by a clinical expert has to be done.
4
Conclusions
In this work, an analysis of different texture features was made to characterize the lungs in CT images of patients infected with COVID-19, and the results found suggested that the disease is not homogeneously distributed throughout the lungs. In particular, GLNU and RLNU were the texture features with more statistical differences between pulmonary regions. This conclusion may provide relevant clinical complementary information suggesting that COVID-19 tends to be distributed in specific zones of lung and probably is more useful to find disease patterns in pulmonary regions, such as bases or apex, instead of considering the entire lungs. To get more evidence that supports the results found in this work, a comparison with state of the art algorithms and a more extensive CT images database needs to be studied, and a correlation between the texture metric values and validated level of affectation of the disease in order to properly characterize the lungs. In addition, in order to select the best parameters to characterize COVID19 lesion by texture, a large range of values for the parameters related to the estimation of GLCM and GLRLM need to be studied (different values of distance between pixels, gray level values, etc.). Finally, the textural characterization of lungs could be used alongside a volumetric segmentation approach looking for to generate a 3D model of the disease caused by COVID-19 and provide more information regarding the evolution of the disease. Acknowledgments. Diomar E. Rodriguez-Obregon acknowledges the support provided by CONACyT with a doctoral Grant (# 787212). Authors acknowledge the Critical Care and Radiology Departments of the INCMNSZ, for their support providing the used CT images.
392
J. G. Delgado-Alejandre et al.
References 1. WHO. World health organization coronavirus (COVID-19) dashboard (2022) 2. Bakheet, S., Al-Hamadi, A.: Automatic detection of covid-19 using pruned glcmbased texture features and LDCRF classification. Comput. Biol. Med. 137, 104781 (2021) 3. Varela-Santos, S., Melin, P.: A new approach for classifying coronavirus covid19 based on its manifestation on chest x-rays using texture features and neural networks. Inf. Sci. 545, 403–414 (2021) 4. Cendejas-Zaragoza, L., Rodriguez-Obregon, D.E., Mejia-Rodriguez, A.R., ArceSantana, E.R., Santos-Diaz, A.: Covid-19 volumetric pulmonary lesion estimation on ct images using a u-net and probabilistic active contour segmentation. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 3850–3853. IEEE (2021) 5. Kiser, K.J., et al.: Plethora: Pleural effusion and thoracic cavity segmentations in diseased lungs for benchmarking chest ct processing pipelines. Med. Phys. 47(11), 5941–5952 (2020) 6. Nemec, S.F., Bankier, A.A., Eisenberg, R.L.: Lower lobe-predominant diseases of the lung. Am. J. Roentgenol. 200(4), 712–728 (2013) 7. Pr´ o, E.A.: Anatom´ıa cl´ınica. M´edica Panamericana (2014) 8. Haralick, R.M., Shanmugam, K., Dinstein, I.H.: Textural features for image classification. IEEE Trans. Syst. Man Cybern. 6, 610–621 (1973) 9. Tang, X.: Texture information in run-length matrices. IEEE Trans. Image Process. 7(11), 1602–1609 (1998) 10. Walpole, R.E., Myers, R.H., Myers, S.L., Ye, K.: Probability & statistics for engineers & scientists. Pearson Cloth 1, 67–71 (2017) 11. Bao, C., Liu, X., Zhang, H., Li, Y., Liu, J.: Coronavirus disease 2019 (covid-19) CT findings: a systematic review and meta-analysis. J. Am. Coll. Radiol. 17(6), 701–709 (2020)
Glioblastoma Classification in Hyperspectral Images by Reflectance Calibration with Normalization Correction and Nonlinear Unmixing In´es Alejandro Cruz-Guerrero1(B) , Juan Nicolas Mendoza-Chavarr´ıa1 , and Daniel Ulises Campos-Delgado1,2 1
2
Facultad de Ciencias, Universidad Aut´ onoma de San Luis Potos´ı, 78290 San Luis Potos´ı, San Luis Potos´ı, M´exico [email protected] ´ Instituto de Investigaci´ on en Comunicaci´ on Optica, Universidad Aut´ onoma de San Luis Potos´ı, 78210 San Luis Potos´ı, San Luis Potos´ı, M´exico Abstract. In this work, a new normalized reflectance calibration proposal is presented for hyperspectral (HS) images, and evaluated through a nonlinear unmixing classification method. This evaluation was performed on craniotomy HS images to classify regions affected by grade IV glioblastoma tumor. The classification methodology follows a semi-supervised strategy where some pixels in the HS images were manually labeled by a clinical expert. The nonlinear unmixing of the HS images is carried out by using a multilinear model, and the abundances of the estimated endmembers are the distinctive features for classification purposes. The evaluation results show that the proposed calibration decreases the variability of the spectral signatures, increasing the classification accuracy compared to the standard methodology of the state of the art. These results demonstrate that the new formulation allows reflectances calibration without losing characteristic features, which allows better separability among classes than with the standard calibration. Keywords: Hyperspectral imaging calibration
1
· Nonlinear unmixing · Reflectance
Introduction
Hyperspectral imaging (HSI) is an alternative method of non-contact and nonionizing optical imaging, where spatial and spectral information of a scene is captured simultaneously [1,2]. The principle of this technique is based on light incising on a surface, where by the characteristic properties of the materials, it can be absorbed, diffracted, scattered and/or reflected [2]. This information is acquired by a hyperspectral (HS) camera that forms three-dimensional images with the spectral information of the scene, where each spatial location has a response or spectral signature (SS) corresponding to the resulting reflectance [3]. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 393–402, 2023. https://doi.org/10.1007/978-3-031-18256-3_43
394
I. A. Cruz-Guerrero et al.
The SSs allow identification of the compounds present in the scene, since the shape of these signatures is characteristic of each component [2]. Due to this, HSI has been widely used in various scientific fields since its invention by NASA, such as geology, archaeology, vegetation monitoring, water resources, safety control, medical diagnosis, etc. [2]. The high potential of HSI in diagnostics is related to the progressive effects of a pathology, since the absorption, fluorescence, and scattering properties of diseased tissue change with time, which means that HS images can capture quantitative information on the progress of the disease [4]. Despite the advantages of HSI, the SSs can vary depending on the optical equipment, so it is necessary to implement a reflectance calibration step to standardize the SSs between different HS cameras. The standard calibration (SC) in the literature is a linear transformation defined by the following equation: ˆ y, n) = 100 × Iraw (x, y, n) − RD (x, y, n) I(x, RW (x, y, n) − RD (x, y, n)
(1)
ˆ y, n) is the calibrated HS image, RD (x, y, n) and RW (x, y, n) are where I(x, the dark (dark current) and white reference HS images, where x ∈ {1, . . . , X}, y ∈ {1, . . . , Y }, and n ∈ {1, . . . , N }; for the spatial dimensions X × Y and N the number of spectral bands, respectively. The SC limits the amplitudes range for the SSs and compensates for most unwanted effects at the acquisition time [5]. However, there are still problems that affect the captured information, such as the non-ideal response of bandpass optical filters in conjunction with electronics. These phenomena can cause some parasitic effects, such as crosstalk, leakage, and harmonics, resulting in a slight increase in the amplitude of the SSs [6]. In most cases, these effects go unnoticed, since the HS cameras return the discretized spectral response for each pixel; accordingly, there is increased variability and morphology changes in the SSs [6]. The literature has proposed various approaches to compensate for the undesirable effects present during capture, to reduce and offset the impact of these effects. An example is shown in [6], where an algorithm is derived to compensate the spectral response of optical sensors by formulating an inverse regression problem when discretizing each band. However, to solve the inverse interpolation problem, the continuous spectral response of the sensors has to be carefully characterized, which is difficult to access in commercial HS cameras. In addition, different methodologies have been proposed in the state of the art, for example, [7] suggests transfer learning, and [8] describes optimal transformation models; however, these strategies require multiple scenes or calibration palettes, which lead to longer acquisition times and higher computational costs. On the other hand, in [9], a modification of SC is proposed that restricts the allowable values in the reflectances by applying a normalization correction. This proposal has the advantage that the same reference images are used in SC, in addition to preserving the characteristic features of the materials in the scene, and therefore produces better separability in classification problems. Because of this, in this contribution, we present the application of [9] in craniotomy HSI by analyzing its behavior in a classification problem. The classification features were
Glioblastoma Classification in HSI
395
extracted with a nonlinear unmixing algorithm. To validate the results, the proposed calibration (PC) is compared with the SC in (1) by assessing the interand intra-patient classification performance.
2
Method
The methodology followed to contrast the SC and the PC is described in Fig. 1, which consists of five stages: HSI raw input (Fig. 1A), pre-processing (Fig. 1B), semi-supervised data set (Fig. 1C), nonlinear unmixing (Fig. 1D), and classification results (Fig. 1E). The database used in this work consists of four HSIs in-vivo of human brain in the visible and near-infrared wavelengths described in [10]. In these images, some pixels were manually classified by pathologists, selecting four general categories or classes: tumor tissue (TT), normal tissue (NT), hypervascularized tissue (HT), and background (BG) [11]. The preprocessing stage consisted of three steps: image calibration, band reduction, and noise reduction (explained in [11]). In the first stage, (1) and the PC described in Sect. 2.1 are applied. The second step is the reduction of the spectral bands, in which one in five bands is selected, and the extreme ones are eliminated. Finally, in the noise reduction step, a smoothing filter is used. The name of the images are as follows: Op8C1, Op12C1, Op15C1, and Op20C1. Once the pre-processing stage was completed, some pixels were manually labeled by a pathologists and used as training data in a semi-supervised perspective through two different approaches: intra-patient and inter-patient. In the intra-patient scheme, the labeled data from each image were used to generate their analysis, while in the inter-patient case, all labeled data from the HSIs were considered, except for the labeled pixels of the HSI to be classified.
Fig. 1. Overall diagram of the methodology for the evaluation of standard and proposed calibration in the classification of HS images.
396
2.1
I. A. Cruz-Guerrero et al.
Proposed Calibration
The proposed calibration departs from the SC in (1), but now employing differently the raw HS image (Iraw (x, y, n)), the white (RW (x, y, n)) and dark (RD (x, y, n)) references. The new calibration defines two auxiliary HS images: α(x, y, n) = Iraw (x, y, n) − RD (x, y, n),
(2)
β(x, y, n) = RW (x, y, n) − RD (x, y, n) − min α(x, y, n).
(3)
x,y,n
In β(x, y, n), a polarization component was added under the assumption that minx,y,n RW minx,y,n α(x, y, n); therefore, this component establishes a lower bound to guarantee always positive or zero values. Ideally, the global minimum value of α(x, y, n) is greater than or equal to zero, since the reflectances are always positive quantities. However, α(x, y, n) could show negative values caused by undesired effects in the acquisition stage. Therefore, a transformation is applied to guarantee this property. Once the auxiliary variables have been established, we proceed to limit the upper values of β(x, y, n) by: ˆ y, n) = β(x,
β(x, y, n) ∈ [0, 1]. maxx,y (β(x, y, n))
(4)
On the other hand, α(x, y, n) is updated by the same perspective such that α ˆ (x, y, n) =
α(x, y, n) − minx,y,n α(x, y, n) , ˆ y, n) β(x,
(5)
where the offset is applied using the global minimum of α(x, y, n), restricting the minimum allowable value to zero. However, the maximum values are not bounded (ˆ α(x, y, n) ∈ [0, ∞)), so this condition is achieved by the following normalization: ˆ y, n) = 100 × I(x,
α ˆ (x, y, n) ∈ [0, 100]. maxx,y (ˆ α(x, y, n))
(6)
This new approach reduces the spectral variability and non-uniform gain of the HS sensor, while maintaining the separability between classes by ensuring values in the range of [0,100]. 2.2
Nonlinear Unmixing
The nonlinear unmixing stage (Fig. 1D) is used as a features extraction step, in which the HSI data are separated into their pure spectral components or end-members and their fractional contributions in a pixel, generating the abundance maps. The main motivation for performing a spectral decomposition is to identify the characteristic end-members per studied class in the HSI, and to use their abundances as classification features. In this sense, some pixels have
Glioblastoma Classification in HSI
397
been labeled by an expert, whose spectral information is used to establish a semi-supervised classification task. For this goal, the nonlinear extended blind end-members and abundance extraction (NEBEAE, described in [12]) algorithm is used by following the implementation described in [13]. In this way, the classification methodology includes two main steps: a) estimation of the characteristic end-members, and b) computation of abundances. Hence, in the first step, the characteristic end-members are estimated by using the labeled pixels by the clinical expert, and dividing the information into four sets corresponding to the four classes of interest and considering the intraand inter-patient perspectives. Once the sets of labeled pixels were defined, different numbers of end-members per class were manually selected, which are shown in Fig. 2, and are common for both classification perspectives. These values were selected to generate the lowest estimation error in the data reconstruction. In the case of considering just one end-member for the class, this one is obtained by averaging the SSs of those labeled pixels. In all other cases, the end-members were estimated by applying NEBEAE to each set of labeled pixels. In this step, the following hyperparameters in NEBEAE were set: (i) the similarity coefficient ρ = 0.01, and (ii) the entropy weight λ = 0 [14]. In addition, the vertex component analysis was selected as initialization step of the end-members matrix [15]. Once the characteristic end-members were computed, they are used to estimate their respective abundance maps through NEBEAE with ρ = 0 and λ = 0.2 (the end-members matrix on this step was kept fixed) on the HS image to be classified.
Fig. 2. Number of end-members per class in each HS image
398
2.3
I. A. Cruz-Guerrero et al.
Classification Results
The classification map was obtained from the estimated abundances per pixel. First, the abundances of the end-members belonging to the same class were added, then the pixel label was based on the greatest contribution of the class end-member. The classification results were evaluated by considering the pixels labeled by the clinical expert as ground-truth, and using the following metrics: accuracy, sensitivity, specificity, and F1-score.
3 3.1
Results Calibration Results
Figure 3 shows that, except for BG, each calibration provides a different average SS for the studied classes. Looking at these average results, PC provides less variability between end-members within each class specially in TT and HT, which is observed by the greater red shaded area compared to the green one.
Fig. 3. Mean and standard deviation (σ) of the estimated end-members per class.
Glioblastoma Classification in HSI
3.2
399
Classification Results
By comparing the classified maps (see Figs. 4 and 5), a reduction in false positives can be observed in TT for both inter- and intra-patient scenarios using the PC. The PC also provides a better identification of the marking rubber rings. An example of this phenomenon is shown in Fig. 5 A)Op8C1 for the intra-patient strategy, as well as in D)Op20C1. Regardless the calibration method, the most consistent labeled maps were obtained through the intra-patient approach. When analyzing the quantitative results (see Figs. 6 and 7), as in the classified maps, it is observed that the intra-patient strategy generates better performance indices. In this case, both calibrations show similar values. Observing the interpatient results, just in image Op8C1, we obtained better results by using SC, however, in the rest of the images, PC overcame by far the SC performance scores. Especially this remarkable difference was found in the specificity score for the images Op12C1 and Op15C1.
Fig. 4. Results of inter-patient classification performance by using HS images: a) Op8C1, b) Op12C1, c) Op15C1, and d) Op20C1. The first column corresponds to the synthetic RGB image for a reference, the second column illustrates the ground-truth maps, the third column shows the SC labels maps, and the last column presents the PC results. The following colormap was considered: NT (green), TT (red), HT (blue), and BG (black); white pixels correspond to non-labeled data. (Color figure online)
400
I. A. Cruz-Guerrero et al.
Fig. 5. Results of intra-patient classification performance by using HS images: a) Op8C1, b) Op12C1, c) Op15C1, and d) Op20C1. The first column corresponds to the synthetic RGB image for a reference, the second column illustrates the ground-truth maps, the third column shows the SC labels maps, and the last column presents the PC results. The following colormap was considered: NT (green), TT (red), HT (blue), and BG (black); white pixels correspond to non-labeled data. (Color figure online)
Fig. 6. Inter-patient classification performance metrics for each HS image.
Glioblastoma Classification in HSI
401
Fig. 7. Intra-patient classification performance metrics for each HS image.
4
Conclusions
In this work, a new strategy for the calibration of HS images was presented to standardize the SSs of different devices. Once the proposed calibration was established, its impact was analyzed by classifying medical HS images by using nonlinear unmixing. The PC results showed SSs with less variation within the same class. This reduction in variability was reflected in the performance of inter-patient strategy, which reflects a real scenario for a possible application, where the specific information of the patient is not available just previous studied HS images. As future work, we propose to carry out a formal derivation of the proposed calibration departing from the response of the optical filters, as well as, to validate the results on other types of HSIs.
References 1. Lu, G., Fei, B.: Medical hyperspectral imaging: a review. J. Biomed. Opt. 19(1), 010901 (2014) 2. Khan, M.J., Khan, H.S., Yousaf, A., Khurshid, K., Abbas, A.: Modern trends in hyperspectral image analysis: a review. IEEE Access 6, 14118–14129 (2018) 3. Huang, H., Liu, L., Ngadi, M.O.: Recent developments in hyperspectral imaging for assessment of food quality and safety. Sensors 14(4), 7248–7276 (2014) 4. Fei, B.: Hyperspectral imaging in medical applications. In: Data Handling in Science and Technology, vol. 32, pp. 523–565. Elsevier (2020) 5. Geladi, P., Burger, J., Lestander, T.: Hyperspectral imaging: calibration problems and solutions. Chemom. Intell. Lab. Syst. 72(2), 209–217 (2004)
402
I. A. Cruz-Guerrero et al.
6. Pichette, J., Goossens, T., Vunckx, K., Lambrechts, A.: Hyperspectral calibration method For CMOS-based hyperspectral sensors. In: Photonic Instrumentation Engineering IV, vol. 10110, p. 101100H. SPIE (2017) 7. Zhao, S., Qiu, Z., He, Y.: Transfer learning strategy for plastic pollution detection in soil: calibration transfer from high-throughput HSI system to NIR sensor. Chemosphere 272, 129908 (2021) 8. Burger, J., Geladi, P.: Hyperspectral NIR image regression part I: calibration and correction. J. Chemometr. Soc. 19(5–7), 355–363 (2005) 9. Cruz-Guerrero, I.A., et al.: Reflectance calibration with normalization correction in hyperspectral imaging. In: 2022 25th Euromicro Conference on Digital System Design (DSD). IEEE (2022) 10. Fabelo, H., et al.: In-vivo hyperspectral human brain image database for brain cancer detection. In: IEEE Access, vol. 7, pp. 39098–39116 (2019) 11. Cruz-Guerrero, I.A., Leon, R., Campos-Delgado, D.U., Ortega, S., Fabelo, H., Callico, G.M.: Classification of hyperspectral in vivo brain tissue based on linear unmixing. Appl. Sci. 10(16), 5686 (2020) 12. Campos-Delgado, D.U., et al.: Nonlinear extended blind end-member and abundance extraction for hyperspectral images. Sig. Process. 201, 108718 (2022) 13. Chavarr´ıa, J.N.M., Cruz-Guerrero, I.A., Mejia-Rodriguez, A.R., Campos-Delgado, D.U.: Algoritmo de descomposici´ on ciega basado en el modelo de mezcla multilineal. Memorias del Congreso Nacional de Ingenier´ıa Biom´edica 8, 106–109 (2021) 14. Campos-Delgado, D.U., et al.: Extended blind end-member and abundance extraction for biomedical imaging applications. IEEE Access 7, 178539–178552 (2019) 15. Nascimento, J.M., Dias, J.M.: Vertex component analysis: a fast algorithm to unmix hyperspectral data. IEEE Trans. Geosci. Remote Sens. 43, 898–910 (2005)
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein with the Phasor-FLIM Approach Signatures of Laurdan Balam Benítez-Mata1 , Francesco Palomba1 , Zhiqun Tan2 Leslie Thompson3 , and Michelle Digman1(B)
,
1 Laboratory for Fluorescence Dynamics, Department of Biomedical Engineering,
University of California, Irvine, CA 92627, USA [email protected] 2 Institute for Memory Impairments and Neurological Disorders, University of California, Irvine, CA 92627, USA 3 School of Medicine, University of California, Irvine, CA 92627, USA
Abstract. Huntington’s Disease, known for the presence of extended polyQ repeats in the huntingtin protein, has pathological effects on cellular membrane organelles. Here, we describe disturbances in membrane organization caused by the expression of mutant polyQ. We use the environment-sensitive fluorescent probe, LAURDAN, to assess variations in membrane lipid order of SH-SY5Y cells. The cells were induced to express labeled non-pathogenic (Q18mApple) and pathogenic (Q53mApple) proteins. Our approach takes advantage of LAURDAN’s affinity for hydrophobic regions, such as membranes, where it displays a red shift in emission associated with higher membrane fluidity (MF) instigated by changes in dipolar relaxation (DR) from the penetration of water molecules into the lipid bilayer. To assess fluidity in membranes, we use the Phasor analysis where we analyze LAURDAN fluorescence lifetime. In the phasor analysis plot, we identify two axes, one sensitive to MF and another to DR. Here we show that expression of pathogenic polyQ, correlates with increasing membrane fluidity, with no changes in DR processes, that suggests a disturbance in water penetration but not in membrane-lipid composition. Moreover, we show MF and DR processes are not inversely proportional and can be distinguished apart using lifetime measurements. Keywords: Fluorescence microscopy · Membrane fluidity · Huntington’s Disease
1 Introduction Huntington’s disease (HD) is a neurodegenerative disease characterized by the expression of abnormal poly-glutamine (CAG) repeats (polyQ) in the huntingtin (Htt) protein. Expanded polyQ longer than 35 polyQ repeats in the exon 1 of the huntingtin gene, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 403–413, 2023. https://doi.org/10.1007/978-3-031-18256-3_44
404
B. Benítez-Mata et al.
are considered pathogenic (mHtt). These extended polyQ repeats dysregulate the native protein function by assembling into aggregates and forming inclusions [1]–[3]. HD affects cholesterol metabolism by dysregulating the transcription of PGC1α, SREBP and LXR, altering biosynthesis and lowering intracellular levels of cholesterol, leading to membranes deterioration, and compromising cellular organelles now prone to dysfunctionality [4]. In Huntington’s disease there are disturbances in lipids composition that impact membrane fluidity (MF) or water penetration into the membrane bilayer. These types of lipids include but are not limited to cholesterol, GM1 gangliosides, monounsaturated and polyunsaturated fatty acids docosahexaenoic acid (DHA) and arachidonic acid (AA) [5]–[8]. Our work as well as others, have shown that overall, MF increases in mHtt cells [9, 10]. However, membrane fluidity is not only lipid dependent, proteins within the membrane bilayer also have effects in it [11]. Here we use LAURDAN as a membrane order reporter, a probe with relative high quantum yield, photostability and with negligible cytotoxicity. LAURDAN shows a red-spectral shift as the membrane environment polarity increases, providing information about MF and Dipolar Relaxation (DR) (Fig. 1A). Fluorescence lifetime imaging microscopy (FLIM) measures the average time electrons remain in an excited state before decaying back to ground state. The phasor approach enables the graphical representation of lifetimes on a polar coordinate system [12, 13]. Golfetto et al. used the phasor approach to FLIM to describe the sensitivity of LAURDAN lifetime to separately detect MF and DR, which are thought to be inversely proportional, the latter one reports on the mobility of hydrated lipid moieties, either saturated or unsaturated, where the LAURDAN molecule is inserted [14]–[16]. To the best of our knowledge, there are no reports showing how mHtt affects DR in all cell membranes (total membranes). Here we use LAURDAN lifetime to investigate the MF and DR processes in total membranes during the expression of mHtt polyQ53. We quantify MF and DR changes using a three-component analysis and the Phasor approach to lifetime measurements. We hypothesize this approach can distinguish between changes in MF and DR processes in total membranes as a consequence of mHtt disturbing lipids metabolism.
2 Methods 2.1 Materials Chemicals used: LAURDAN in crystal-powder form (6-dodecanoyl-2dimethylaminonaphthalene), Thermo-Fisher Scientific Inc., Ponasterone-A Santa Cruz Biotechnology, Inc., Cell culture reagents from Genesee Scientific Corporation., Antibiotic cocktail made of Zeocin, Invitrogen, and Geneticin, GIBCO. 2.2 Cell Culture We cultured human derived neuroblastoma SH-SY5Y cell, a widely used model in neurodegenerative diseases. Stable cell lines were generated to respond to ponasterone induction such that this inducible system allows for the expression of the expanded
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein
405
Exon 1 (Httex1) containing either 18Q or the pathogenic 53Q repeats followed by the fluorescent tag protein mApple, or just mApple as control. The cells are maintained in DMEM/F12 1:1 medium, supplemented with 120 μg/mL of the antibiotic cocktail and 10% FBS. Cells are passaged every 2–3 days. For experiments, cells are plated 48 h prior to imaging in 8-chambered cover glass dishes (Cellvis). Cells are left to adjust overnight followed by Ponasterone induction, 10 μM, for 24 h to express the fluorescent constructs. Cells are washed with complete medium, stained and incubated with a final concentration of 6 μM LAURDAN dissolved in DMSO (final concentration < 0.2% v/v) for 60 min. Cholesterol depletion experiments, are done with MβCD, diluted in deionized water, final concentration 10 μM. 2.3 Fluorescence Lifetime Imaging Microscopy FLIM acquisition is carried out as previously reported [17]. Briefly, a Zeiss LSM880 coupled to a 2-photon laser (80 fs ultra-short pulse width with an 80 MHz repetition rate) set at 780 nm. A Plan-Apochromat 63x oil immersion objective NA 1.4 (Carl Zeiss). Detection by two high efficiency GaAsP Hybrid Detectors (HPM-100-40, Becker and Hickl GmbH). LAURDAN fluorescence is split with a 495 nm LP dichroic and into two bandpass filters (Semrock) centered at 460/80 nm for Channel 1 and 540/50 nm for Channel 2. Acquisition parameters: frame size is 256 × 256 pixels, pixel dwell time is 16.38 μs/pixel, pixel size is 264 nm. The fluorescence lifetime decays are recorded and calibrated with SimFCS v4 software, available from the Laboratory for Fluorescence Dynamics (www.lfd.uci.edu). For the phasor analysis, a median filter (3 × 3) is applied to all phasor plots using the SimFCS software. 2.4 The Phasor Approach for Lifetime Data The phasor approach relies on the fluorescence time decay at each pixel in an intensity image, this fluorescence decay consists of different photon arrival times to the detectors, emitted upon the fluorophore excitation. The phasor analysis uses a Fourier transformation applied to each pixel of the image to obtain the G(τ) and S(τ) coordinates, corresponding to the x- and y-coordinates in a scatterplot, respectively, using Eqs. 1 and 2. Each pixel in the image is now referred as a phasor, plotted in a polar plot, also denominated as the Phasor plot (Fig. 1B). G(τ ) =
∫∞ 0 Ii,j (t)cos(ωt)dt ∫∞ 0 Ii,j (t)dt
(1)
S(τ ) =
∫∞ 0 Ii,j (t)sin(ωt)dt ∫∞ 0 Ii,j (t)dt
(2)
where Ii,j (t) indicates the recorded intensity in pixel (i,j) at time t, f is the laser repetition frequency, used to obtain the angular modulation frequency, ω = 2πf . The Phasor plot is a semi-circle, where mono-exponential lifetime decays will lie along the line. Multi-exponential decays will be plotted inside the semi-circle.
406
B. Benítez-Mata et al.
Fig. 1. A) Photophysics of LAURDAN. B) The phasor plot and the three-component analysis of phasor distributions of LAURDAN lifetimes for detection channels 1 and 2, blue and red, respectively.
2.5 Phasor Distributions Analyzed by the Three-Cursor Approach The three-component analysis is a way to quantify the contribution of three independent components to a single pixel making use of the addition rules that apply to phasor plots. For this work, three cursors are set based on LAURDAN’s global lifetime changes in both Channel 1 and Channel 2 combined, to quantify MF and DR (Fig. 1B). The line from cursors “A” to “C”, represent the changes referring to MF. A characteristic shift of the phasors position from longer (0,0) to shorter (1,0) lifetimes is related to a shift from nonpolar region (high cholesterol) to polar region (low cholesterol) [16, 17]. To detect DR on LAURDAN’s lifetime, a third cursor, “B”, is placed outside the universal circle, creating the line A-B. The analysis allows the obtention of histograms with the fraction for MF and DR (Fig. 1B). Fluidity fraction =
fc fc + fa
Dipolar relaxation fraction =
b fb + fa
(3) (4)
For every fractional distribution in each axis, the center of mass (CM) is calculated as: i=100 Fi ∗i centre of mass(CM) i=0 (5) i=100 i=0 Fi where Fi represents the fraction corresponding to each axis, either MF or DR. 2.6 Statistics GraphPad Software, Inc. is used to analyze three independent biological replicates with 10 cells each. For MβCD treatments only one replicate with 10 cells each, is used. Fractional histograms show the mean and positive standard deviation (SD). Violin plots are shown as median with minimum and maximum values. For multiple groups comparison an ANOVA with Tukey post-test is performed. For statistical significance, a p-value lower to 0.05 is used.
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein
407
3 Results 3.1 Expression of mHtt with a PolyQ53 Increases Membrane Fluidity in Total Membranes The lifetime phasor plot for LAURDAN’s channel 1 emission in SH-SY5Y cells expressing the different Htt constructs is shown in Fig. 2. The phasor distribution is analyzed with the three-component analysis to assess changes in the fluidity and DR axes (Fig. 2A). For each axis, a fractional histogram ranging from 0 to 100 is used for easier interpretation of data. An increase in fraction, represents an increase in each axis. A color scale map is used to visualize lifetime changes of LAURDAN in each condition, along with an intensity-based image of the fluorescently tagged construct (Fig. 2B).
Fig. 2. Analysis on Channel 1, sensitive to polarity. A) Example of Phasor plot distribution for mApple and polyQ constructs. A three-component analysis is shown for MF and DR axes. B) Exemplary images of LAURDAN‘s intensity emission and pseudo colored lifetime values for Fluidity and DR.
408
B. Benítez-Mata et al.
Expression of polyQ53 increases the fluidity fraction in total membranes when compared against a nonpathogenic polyQ18 and mApple constructs (Fig. 3A). This aligns with an increase in the CM of the fluidity-axis (Fig. 4A). This effect has been previously reported by our group using hyperspectral phasors in a PC12 cell model that expresses a polyQ97-mRuby [9]. When cholesterol is depleted with MβCD for all conditions, the MF increases as previously reported [9, 16]. Worth to mention that the polyQ53 + MβCD condition, although not statistically different from MβCD controls, presents a narrower distribution of CM values, this may be an indicative of a more efficient depletion of lipids in the cells expressing the mutant polyQ53. The DR-axis in channel 1 is virtually not sensitive to DR processes, and variations correspond to the spectral red-shift within channel 1 detection’s bandwidth, which lies within the line of the universal circle in the phasor plot (Fig. 2A).
Fig. 3. Analysis on Channel 1, for each cell a A) Fluidity fraction, top, and B) Dipolar Relaxation Fraction, bottom, histograms are created. We show the average and standard deviations (mean + SD). Negative SD is avoided for simplicity.
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein
409
Fig. 4. Analysis on Channel 1, from each individual histogram in Fig. 3, a Center of Mass for A) Fluidity- and B) DR-axes are calculated. Statistical significance is designated with ****, ***, ** and *. Where p < 0.0001, p < 0.001, p < 0.01 or p < 0.05, respectively.
3.2 Expression of mHtt PolyQ53 Does not Increases DR Processes in Total Membranes Channel 2 is sensitive to MF changes due to the spectral shift, but also has sensitivity to DR processes caused by the types of lipids present in the membrane, which causes a rotation of the phase angle and locates the phasors outside of the universal circle with respect to channel 1 due an apparent time delay of emitted photons with respect to excitation (Fig. 1) [16]. Figure 5 shows the three-component analysis done for channel 2 the same way as done for Channel 1.
410
B. Benítez-Mata et al.
Fig. 5. Analysis on Channel 2 sensitive to DR. A) Example of Phasor plot distribution for mApple and polyQ constructs. A three-component analysis is shown for MF and DR axes. B) Exemplary images of LAURDAN‘s intensity emission and pseudo colored lifetime values for Fluidity and DR.
Changes in the fluidity-axis detected in Channel 2 suggest an amplified effect on MF caused by polyQ53 (Fig. 6A and Fig. 7A). We do not observe a significant difference in DR processes between cells expressing pathogenic and control constructs. Changes in DR are thought to be inversely proportional to MF effects, phenomenon shown in samples treated with MβCD for cholesterol depletion, where MF increases and DR decreases (Fig. 6 and Fig. 7).
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein
411
However, expression of polyQ53 might be affecting the lipids packaging (hydration level) but not their moieties’ dynamics, since DR mainly reports the mobility of hydrated lipid moieties, dependent on the type of lipid, saturated or unsaturated, where the LAURDAN molecule is inserted [14, 15].
Fig. 6. Analysis on Channel 2, for each cell a A) Fluidity fraction, top, and B) Dipolar Relaxation Fraction, bottom, histograms are created. We show the average and standard deviations (mean + SD). Negative SD is avoided for simplicity.
412
B. Benítez-Mata et al.
Fig. 7. Analysis on Channel 2, from each individual histogram in Fig. 3, a Center of Mass for A) Fluidity- and B) DR-axes are calculated. Statistical significance is designated with ****, ***, ** and *. Where p < 0.0001, p < 0.001, p < 0.01 or p < 0.05, respectively.
4 Conclusion Our results indicate that in the absence of inclusions of mHtt polyQ53, there is an increase in membrane fluidity but no changes in DR. This opens a new discussion threat where LAURDAN photophysic-properties report on MF and DR are not inversely proportional processes and can be isolated by measuring the probe’s lifetime. Further studies in lipid composition of membranes are needed to confirm this result. To our knowledge, this is the first time the phasor approach is used for LAURDAN lifetime measurements, which enables the study and isolation of MF and DR processes in an HD cell model. In this work, we show that expressing the expanded mHtt polyQ53 in HD-SH-SY5Y cells, increases the membrane fluidity without further sign of decrease in DR processes, suggesting disturbances in membrane hydration levels but not in lipid packaging and dynamics of their moieties, in the absence of mHtt inclusions. Acknowledgements. Funding agencies, Mexican National Council for Science and Technology (CONACYT, fellowship 739656). The Laboratory for Fluorescence Dynamics (LFD) at the University of California, Irvine (UCI) is supported jointly by the National Institutes of Health through the grant P41GM103540 and UCI.
References 1. Testa, C.M., Jankovic, J.: Huntington disease: a quarter century of progress since the gene discovery. J. Neurol. Sci. 396, 52–68 (2019). https://doi.org/10.1016/j.jns.2018.09.022
Changes in Membrane Fluidity of the Expanded Mutant Huntingtin Protein
413
2. Kim, Y.E., et al.: Soluble oligomers of PolyQ-expanded huntingtin target a multiplicity of key cellular factors. Mol. Cell 63(6), 951–964 (2016). https://doi.org/10.1016/j.molcel.2016. 07.022 3. Kremer, B., et al.: A worldwide study of the Huntington’s disease mutation: the sensitivity and specificity of measuring CAG repeats. N. Engl. J. Med. 330(20), 1401–1406 (1994). https:// doi.org/10.1056/NEJM199405193302001 4. Leoni, V., Caccia, C.: The impairment of cholesterol metabolism in Huntington disease. Biochim. Biophys. Acta - Mol. Cell Biol. Lipids 1851(8), 1095–1105 (2015). https://doi.org/ 10.1016/j.bbalip.2014.12.018 5. Gaus, K., Zech, T., Harder, T.: Visualizing membrane microdomains by Laurdan 2-photon microscopy (Review). Mol. Membr. Biol. 23(1), 41–48 (2006). https://doi.org/10.1080/096 87860500466857 6. Alpaugh, M., et al.: Disease-modifying effects of ganglioside GM1 in Huntington’s disease models. EMBO Mol. Med. 9(11), 1537–1557 (2017). https://doi.org/10.15252/emmm.201 707763 7. Mesa-Herrera, F., Taoro-González, L., Valdés-Baizabal, C., Diaz, M., Marín, R.: Lipid and lipid raft alteration in aging and neurodegenerative diseases: a window for the development of new biomarkers. Int. J. Mol. Sci. 20(1), 5 (2019). https://doi.org/10.3390/ijms20153810 8. Sambra, V., Echeverria, F., Valenzuela, A., Chouinard-Watkins, R., Valenzuela, R.: Docosahexaenoic and arachidonic acids as neuroprotective nutrients throughout the life cycle. Nutrients 13(3), 986 (2021). https://doi.org/10.3390/nu13030986 9. Sameni, S., Malacrida, L., Tan, Z., Digman, M.A.: Alteration in fluidity of cell plasma membrane in huntington disease revealed by spectral phasor analysis. Sci. Rep. 8(1), 1 (2018). https://doi.org/10.1038/s41598-018-19160-0 10. Subczynski, W.K., Pasenkiewicz-Gierula, M., Widomska, J., Mainali, L., Raguz, M.: High cholesterol/low cholesterol: effects in biological membranes: a review. Cell Biochem. Biophys. 75(3–4), 369–385 (2017). https://doi.org/10.1007/s12013-017-0792-7 11. Lenaz, G., Castelli, G.P.: Membrane fluidity: molecular basis and physiological significance. In: Structure and Properties of Cell Membrane: Volume I: A Survey of Molecular Aspects of Membrane Structure and Function, vol. I, pp. 93–136. CRC Press (1985) 12. Digman, M.A., Caiolfa, V.R., Zamai, M., Gratton, E.: The phasor approach to fluorescence lifetime imaging analysis. Biophys. J. 94(2), L14–L16 (2008). https://doi.org/10.1529/bio physj.107.120154 13. Ranjit, S., Malacrida, L., Jameson, D.M., Gratton, E.: Fit-free analysis of fluorescence lifetime imaging data using the phasor approach. Nat. Protoc. 13(9), 1979–2004 (2018). https://doi. org/10.1038/s41596-018-0026-5 14. Jurkiewicz, P., Cwiklik, L., Jungwirth, P., Hof, M.: Lipid hydration and mobility: an interplay between fluorescence solvent relaxation experiments and molecular dynamics simulations. Biochimie 94(1), 26–32 (2012). https://doi.org/10.1016/j.biochi.2011.06.027 15. Bagatolli, L.A.: LAURDAN fluorescence properties in membranes: a journey from the fluorometer to the microscope BT - fluorescent methods to study biological membranes. In: Mély, Y., Duportail, G. (eds), pp. 3–35. Springer, Heidelberg (2013) 16. Golfetto, O., Hinde, E., Gratton, E.: Laurdan fluorescence lifetime discriminates cholesterol content from changes in fluidity in living cell membranes. Biophys. J. 104(6), 1238–1247 (2013). https://doi.org/10.1016/j.bpj.2012.12.057 17. Malacrida, L., Jameson, D.M., Gratton, E.: A multidimensional phasor approach reveals LAURDAN photophysics in NIH-3T3 cell membranes. Sci. Rep. 7(1), 1–11 (2017). https:// doi.org/10.1038/s41598-017-08564-z
A Method for Automatic Monoplane Angiography Segmentation Héctor Emanuel Martín Alcala , Francisco Javier Alvarez Padilla , and Gerardo Mendizabal Ruiz(B) Departamento de Bioingeniería Traslacional, CUCEI, Universidad de Guadalajara, 44430 Guadalajara, Jalisco, Mexico [email protected]
Abstract. The diagnosis of stenosis, characterized by narrowing the lumen on arteries, requires the inspection of medical images acquired using technics such as X-Ray angiography. Recently, convolutional neural networks (CNN) have been successfully applied to automate segmenting arteries on angiographic images. The main challenge of using these models relies on counting many images where the vessels of interest have been manually annotated. Besides being difficult and expensive to produce, obtaining consent for its use in clinical investigations is even more complicated. Thus, this work presents an automatic angiography segmentation method that does not rely on a CNN architecture. Our results indicate the feasibility of using the proposed method for segmenting vascular components on angiographies with accuracies comparable to other CNN-based state-of-the-art methods. Keywords: Angiography · Image segmentation · Hessian based filters
1 Introduction According to the INEGI [1], cardiovascular diseases are the leading cause of death in Mexico and worldwide [2]. Among these diseases, we can find angiopathies, typically provoked by arterial stenosis. Stenosis is narrowing the blood vessels’ lumen, which disturbs the local flow and precludes the adequate irrigation of perfused organs [3]. Stenosis is the foremost cause of strokes, both cardiovascular and cerebrovascular. Typical stenosis symptoms can be fatigue, weakness, chest pain, etc. [4]. Even in the presence of the previously mentioned symptoms, confirming angiopathy requires specific diagnostic studies. In cardiology, the gold standard for stenosis diagnosis is currently X-Ray angiography [5] because it requires low radiation dosages and brings real-time results, which are suitable for real-time treatments. The angiographies are typically made in a procedure room, where the physician looks at a series of images of contrast-dyed arteries obtained using X-Ray imaging. In most cases, the diagnosis is made subjectively by observing these images.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 414–423, 2023. https://doi.org/10.1007/978-3-031-18256-3_45
A Method for Automatic Monoplane Angiography Segmentation
415
Seeking to objectify the diagnosis, Vanninen et al. [6] proposed the quantification of the narrowing of the vascular lumen as a parameter to detect injuries. Recently, it has been shown that computer-aided detection systems can assist practitioners in detecting, diagnosing, and decision-making [7]. Nevertheless, computationally identifying blood vessels becomes a complex task because of the low signal-to-noise ratio, artifacts like bone structures, and contrast variations along arteries [8]. Recently, convolutional neural networks (CNN) have been employed for automatically segmenting angiographic images even in the presence of noise or low quality [9]. However, the main challenge related to developing a CNN model is the large number of images required for training [10] which, in the case of angiography, are challenging to find in open repositories. For instance, it has been reported that the acquisition of angiographies for stenosis classification is practiced only in approximately 0.03% of the USA population [11, 12]. In this work, we present an automatic angiography segmentation method whose results are comparable to another state-of-the-art CNN-based method. The main advantage of our practice is that it does not rely on large databases since it does not employ a supervised approach.
2 Materials and Methods Image segmentation is the process of correctly identifying regions of interest in an image and dividing them. In our problem, the segmentation consists of generating binary masks, which are two-class images that separate interest elements from the background, of angiographic images, where vessels are selected and artifacts and background ignored. The proposed segmentation method consists of four steps. As depicted in Fig. 1, at first, all vessel-like structures are enhanced using Jeman’s vessel enhancement filter. Then, high-gradient boundaries are enhanced to find any focal spot artifacts, and they are removed. Later, image binarization is done using Otsu’s class-separation method. In the end, taking advantage of connected components, artifacts and noise are drawn based on morphological properties.
Fig. 1. Methodology proposal: Diagram depicting the proposed method to segment vessels on angiographies.
416
H. E. Martín Alcala et al.
Vessel Enhancement. To enhance the arteries on images and reduce non-vascular elements and image noise, we use the multiscale vessel enhancement filter proposed by Jerman et al. [14]. This filter, based on the hessian matrix eigenvalues, similarly to its predecessor presented by Frangi et al. [15], gives a close to uniform response to vascular structures and precisely enhances the boundaries between them and the background. An example of the results from the processing with this filter can be seen in Fig. 2.
Fig. 2. From left to right: Original angiography and an example result from applying the Jerman multiscale filter.
Removal of X-Ray Focal Spot Boundaries. In X-Ray devices, the focal spot is the area that receives the electron beam. This is the area that’s later converted into an image. In some X-Ray devices, this focal spot is smaller than the area of the image, as shown in Fig. 3. Thus, its boundaries are shown on images, as in the example shown in Fig. 4.
Fig. 3. X-Ray beam projection: Image showing how the circular X-Ray beam focal spot is projected over a rectangular X-Ray detector, leaving zones without excitation.
The presence of X-Ray focal spot boundaries (FSB) represents a problem when seeing an image as a mathematical function, as Otsu’s method does. Otsu’s process works on histograms of the intensity of images, trying to separate them in a two-class way. Focal spot artifact on images modifies intensity histograms, generating not ideal results on binarization. An example of histograms of an image with focal spot boundaries and without it is shown in Fig. 5.
A Method for Automatic Monoplane Angiography Segmentation
417
Fig. 4. Example of angiography with visible focal spot: The black area on the left is not part of the X-Ray acquired zone.
Fig. 5. Intensity histograms of a vessel-enhanced grayscale image before and after focal spot remotion. From left to right: Intensity histogram before remotion, intensity histogram after remotion.
To remove (FSB), we use the results of a focal-spot-detection oriented Jerman filter, with parameters that allow us to detect its boundaries. After getting the Jerman filter response, binarization is made using a high-pass filter of 99% of image intensity. The circularity of all the components on the binary image is computed. Those components with a circularity greater than 0.9 are selected as candidates for removal, considering their proximity to the image boundaries. An example of the results of this process is shown in Fig. 6. Finally, the components of the binary mask are dilated using a disk structural element of 10 pixels to generate a new mask that is used to remove its content from the vesselenhanced image, as shown in the example in Fig. 7. Mask Binarization. After the Jerman filter and focal spot boundaries movement, we seek to generate a binary mask that entirely separates the arteries from the background. We use the Otsu method, which determines the threshold by maximizing inter-class variance, or separability, from grayscale image histograms [16]. An example of binarizing using Otsu’s threshold on Jerman response images is shown in Fig. 8.
418
H. E. Martín Alcala et al.
Fig. 6. Example of focal spot boundary detection: Focal-spot-detection oriented vessel enhancement response grayscale image, focal spot element mask to be used on focal spot remotion.
Fig. 7. Example of focal spot boundary remotion on vessel-enhanced images. From left to right: Grayscale image before focal spot boundary remotion, grayscale image after focal spot boundary remotion.
Fig. 8. From left to right: Example of a Jerman multiscale filter response, example of binary mask after Otsu’s threshold.
Filtering by Region-Based Features. In the binary mask, it is possible to find some components which do not represent any of the vascular regions of interest. Properties like area, eccentricity, and compacity are analyzed in all the 8-connected components from the binary mask to find and delete these components. An example result is shown in Fig. 9.
A Method for Automatic Monoplane Angiography Segmentation
419
Fig. 9. From left to right: example of an original binary mask, example of binary mask after filtering components with region-based features.
3 Results In this work, we used the DCA1 [13], which is publicly available. This dataset contains 130 coronary angiographies with their corresponding ground truth drawn by an expert cardiologist. The images were acquired at the Cardiology Department of the Mexican Social Security Institute, UMAE T1-Leon, in Mexico. Each angiography is a 300 x 300 pixels grayscale image stored in PGM (Portable Gray Map) format. We used the Dice Score Index [17, 18] to evaluate the performance of the proposed segmentation algorithm. This score is measured using the output image of the method as a testing image and the ground truth found on the dataset as a validator. We compute the Dice Score for each image and then the mean and standard deviation of the score for all 130 images. The result is a mean of 0.6984 with a standard deviation of 0.0021. This result is comparable to the state-of-the-art paper where the dataset is published (Table 1). Table 1. Comparison of results between proposed method and state of the art method. Method
Mean of dice score
Standard deviation
Cervantes-Sanchez et al. [15]
0.6857
Not mentioned
Proposed method
0.6984
0.0021
The proposed method works accurately at the segmentation tasks, where it keeps not only primary vessels but some non-primary vessels, as shown in the example in Fig. 10. Along with the evaluation of the results using the Dice score, a series of images are shown as results examples in Fig. 11.
420
H. E. Martín Alcala et al.
Fig. 10. Example of results of the proposed method: a) original grayscale angiography, b) ground truth mask, c) segmented mask using the proposed method, d) comparison between ground truth and obtained mask, green represents presence in obtained mask but not in ground truth and purple represents presence in ground truth but not in obtained mask.
Fig. 11. Examples of the results obtained with the proposed method. From top to bottom: original angiography, ground truth image, results from the proposed method, results from CervantesSanchez method.
4 Discussion Vessel enhancement filters are widely available, and multiple methods are proposed that use the Hessian matrix to enhance vessel structures on images. Even so, to get a good binary mask, further processing is required. We suggest using Otsu’s threshold selection method to achieve a high-quality binarization, where we keep the information of interest and further analyze the morphology of the components to choose which correspond to interest vessel structures.
A Method for Automatic Monoplane Angiography Segmentation
421
The proposed method removes any focal spot artifacts, as seen in Fig. 11, columns 8 and 9, allowing a better vessel segmentation. The proposed method has been demonstrated to be an effective tool with a performance comparable to the CNN method proposed by Cervantes-Sanchez et al. [13]. Therefore, it meets the requirements with no training needed. Additionally, the segmented vessels include some regions not considered on ground truth masks but could be of interest in some complex analyses, as shown in Fig. 10. Also, we can find a better segmentation when background texture changes more drastically, like in Fig. 11, column 2. We believe it is feasible to generate synthetic stenosis images for training machine learning-based models using architectures known as autoencoders. Autoencoders are a neural network architecture designed to create images with the same statistical properties as the images used to train them [19]. Autoencoders can compress an image, encoding it to a series of descriptors, which are then used by the decoder to try to regenerate the original input [20]. This process allows the autoencoders to produce a synthetic output image from a given input image with the same properties as the ones used to train the network. That way, it is possible to create new synthetic images that look like the original. The segmentation method presented in this work is the first part of a pipeline that employs an autoencoder network architecture to generate synthetic stenosis images. The objective consists of using a binary image representing an artery to generate a synthetic image with properties similar to those generated from an X-Ray angiography. Then, use these synthetic images to complement the data set needed to train a network able to identify stenosis on images. Figure 12 depicts a schematic of the proposed pipeline. The idea is that an autoencoder can generate high-quality synthetic angiographies. The idea is to get accurate vessel segmentation from angiographic images and use them as autoencoder’s input, which gets trained to compress the input mask into a dimensionally reduced latent space and then tries to regenerate the original image from it.
Fig. 12. Synthetic image methodology proposal: segmentation of grayscale angiography, codification of binary segmented image, de-codification from latent space into a synthetic grayscale angiography.
With this method, it will not be necessary to depend on an expert cardiologist to choose vessel regions on angiographies and draw binary masks. This way, we intend to expand the already available dataset with images obtained from a non-public source, automatically generating the corresponding binary masks. Future work will present results of this proposed methodology.
422
H. E. Martín Alcala et al.
References 1. INEGI: Características de las defunciones registradas en México durante 2020. https://www. inegi.org.mx/contenidos/saladeprensa/boletines/2021/EstSociodemo/DefuncionesRegistr adas2020preliminar.pdf. Accessed 04 June 2022 2. WHO: The top 10 causes of death. https://www.who.int/news-room/fact-sheets/detail/thetop-10-causes-of-death. Accessed 04 June 2022 3. Thiriet, M., Delfour, M., Garon, A.: Vascular stenosis: an introduction. In: Lanzer, P. (ed.) PanVascular Medicine, pp. 781–868. Springer, Heidelberg (2015). https://doi.org/10.1007/ 978-3-642-37078-6_32 4. Centers for Disease Control and Prevention: Coronary Artery Disease. https://www.cdc.gov/ heartdisease/coronary_ad.htm. Accessed 04 June 2022 5. Members, T.F., et al.: 2013 ESC guidelines on the management of stable coronary artery disease: the task force on the management of stable coronary artery disease of the European Society of Cardiology. Eur. Heart J. 34(38), 2949–3003 (2013). https://doi.org/10.1093/eur heartj/eht296 6. Vanninen, R.L., Manninen, H.I., Partanen, P.K., Tulla, H., Vainio, P.A.: How should we estimate carotid stenosis using magnetic resonance angiography? Neuroradiology 38(4), 299– 305 (1996). https://doi.org/10.1007/BF00596574 7. Petrick, N., et al.: Evaluation of computer-aided detection and diagnosis systems. Med. Phys. 40(8), 087001 (2013). https://doi.org/10.1118/1.4816310 8. Compas, C.B., Syeda-Mahmood, T., McNeillie P., Beymer, D.: Automatic detection of coronary stenosis in X-ray angiography through spatio-temporal tracking. In: 2014 IEEE 11th International Symposium on Biomedical Imaging (ISBI), pp. 1299–1302 (2014). https://doi. org/10.1109/ISBI.2014.6868115 9. Jang, H., McCormack, D., Tong, F.: Noise-robust recognition of objects by humans and deep neural networks (2020). https://doi.org/10.1101/2020.08.03.234625 10. Thompson, N., Greenewald, K., Lee, K., Manso, G.: The computational limits of deep learning (2020). https://doi.org/10.48550/arXiv.2007.05558 11. Johns Hopkins: Patients Fare Just as Well if Their Nonemergency Angioplasty is Performed at Hospitals without Cardiac Surgery Capability. https://www.hopkinsmedicine.org/news/ media/releases/patients_fare_just_as_well_if_their_nonemergency_angioplasty_is_perfor med_at_hospitals_without_cardiac_surgery_capability. Accessed 04 June 2022 12. United States Census Bureau: Census Bureau Projects U.S. Population of 312.8 Million on New Year’s Day. https://www.census.gov/newsroom/releases/archives/population/cb11-219. html. Accessed 04 June 2022 13. Cervantes-Sanchez, F., Cruz-Aceves, I., Hernandez-Aguirre, A., Hernandez-Gonzalez, M., Solorio-Meza, S.: Automatic segmentation of coronary arteries in X-ray angiograms using multiscale analysis and artificial neural networks. Appl. Sci. 9(24), 5507 (2019). https://doi. org/10.3390/app9245507 14. Jerman, T., Pernuš, F., Likar, B., Spiclin, Ž.: Beyond Frangi: an improved multiscale vesselness filter. In: Proceedings of the SPIE 9413, Medical Imaging 2015: Image Processing (2015). https://doi.org/10.1117/12.2081147 15. Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A.: Multiscale vessel enhancement filtering. In: Wells, W.M., Colchester, A., Delp, S. (eds.) MICCAI 1998. LNCS, vol. 1496, pp. 130–137. Springer, Heidelberg (1998). https://doi.org/10.1007/BFb0056195 16. Otsu, N.: A threshold selection method from gray-level histograms. IEEE Trans. Sys. Man. Cyber. 9(1), 62–66 (1979). https://doi.org/10.1109/TSMC.1979.4310076 17. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297–302 (1945). https://doi.org/10.2307/1932409
A Method for Automatic Monoplane Angiography Segmentation
423
18. Sørensen, T.J.: A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on Danish commons. København: I kommission hos E. Munksgaard (1948) 19. DuMont Schütte, A., et al.: Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation. npj Digit. Med. 4, 141 (2021). https://doi.org/10.1038/s41 746-021-00507-3 20. Bank, D., Koenigstein, M., Giryes, R.: Autoencoders (2021). https://arxiv.org/abs/2003.05991
Lung Segmentation Algorithm and SVM Classification of COVID-19 in CT Images Luis Eduardo Gaeta-Ledesma
and Francisco Javier Alvarez-Padilla(B)
Department of Translational Bioengineering, Division of Cyber-Human Integration Technologies, University of Guadalajara, 44430 Guadalajara, Jalisco, Mexico [email protected]
Abstract. The analysis of COVID-19 by tomographic imaging has been a standard for pandemic management. The application of different types of artificial intelligence algorithms has proven to be an accurate method for disease detection. This study presents a method of lung segmentation and a classification algorithm that allows to discriminate between images that show signs of the disease and those that don’t. In addition, the article seeks to establish what kind of features are relevant when feeding a machine learning algorithm. Texture features extracted from Gray Label Concurrence Matrix (GLCM) and a Gabor filter are used for this purpose. Then, we trained and evaluated a SVM algorithm using different combinations of features. It is found that the features extracted from the Gabor filter work better than those extracted from the GLCM, finding that those features focused exclusively on intensity description work better than those focused on spatial description, at least in early stages. Keywords: Image processing · Computed tomography · Lung segmentation · Machine learning · SVM · GLCM · Gabor filter
1
Introduction
Since the beginning of the COVID-19 pandemic, the use of computed tomography (CT) for the differential diagnosis of the disease has been an essential tool for medical personnel [1]. Because of this, the characterization of early radiological signs and methods of analysis focused on COVID-19 have proliferated throughout the pandemic. One of the most frequent signs of this disease and which in turn appears in very early stages is the Ground Glass Opacity (GGO) as seen in Fig. 1, which consists of a slight increase in pulmonary attenuation, also present in other types of pneumonia [2]. There have been several approaches to the detection of COVID-19 through digital image processing and the use of artificial intelligence algorithms. The use of algorithms based on deep learning stands out among them. For example, Xu et al. [3], describes an application of deep learning focused on classifying and segmenting the GGO regions of COVID-19 in c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 424–433, 2023. https://doi.org/10.1007/978-3-031-18256-3_46
Lung Segmentation and COVID-19 Classification
425
contrast to those caused by early-stage of influenza A; or the study by Zang et al. [4], which describes disease progression and status using convolutional neural networks combined with V-NET bottleneck structures. Despite the widespread use of deep learning and that its results support its application not only in the medical area, but in data science in general, these algorithms have some limitations. One of them is the difficulty in interpreting their results; it is challenging to understand exactly what led to a certain prediction commonly called “black box”. Another one is the need of a large amount of data at the training stage in order to achieve significant results [5]. To reduce the volume of required data, approaches such as Sun et al. (2020) [6] present the selection of features for the classification of COVID-19 using machine learning methods which help to generate a robust basis for generating more efficient classification models. This kind of approaches do not focus on the features at the individual level OR on how these may vary their performance according to the stage of the disease in which they are applied. Therefore, this study focuses on analyzing a small set of features applied in a SVM model in order to find which features are relevant and classify COVID-19 at different stages of disease progression. For this purpose, the following structure is followed: in Sect. 2 a lung segmentation method is presented, which prepares our region of interest for further characterization and classification using a SVM algorithm. A SVM is chosen among other machine learning algorithms because it works well even without a large training dataset and gives a clear idea of the weight of individual characteristics in its performance. In Sect. 3 the results of the automatic segmentation of lungs compared with a manual segmentation by experts, and the precision and accuracy of the machine learning algorithm according to the type of features used for training are presented. Finally, Sect. 4 analyzes obtained results and which features are relevant when training an artificial intelligence algorithm to classify CTs of patients with COVID-19.
(a) Healthy lung.
(b) GGO.
Fig. 1. Lung with no sign of pneumonia (a) vs lung with COVID-19 signs (b). GGOs and consolidations can be found (in red bounding boxes). (Color figure online)
426
2
L. E. Gaeta-Ledesma and F. J. Alvarez-Padilla
Materials and Methods
This section explains the development of the algorithm, described in the following general steps: 1. Lung segmentation: because the COVID-19 mainly signs are found in the lungs, no pertinent data (i.e. noise) is filtered from external lung regions. 2. Classification: the SVM algorithm is used for automatic image classification. This process is divided into the following two steps: (a) Characterization: first order features are extracted using a Gabor filter and second order features using GLCM, which are prepared to feed our machine learning algorithm. (b) Training and evaluation: SVM is trained using both groups of features, then each of them separately and finally a subgroup with the first and second order features that have presented better results of precision and accuracy. 2.1
Lung Segmentation
The present segmentation algorithm is based on the approach presented by Tseng et al. 2009 [7] and Brown et al. (1997) [8]. Our contribution is to add both thorax and full body CT scans to the training process. The lung segmentation algorithm is started by generating a mask using a thresholding method. Lung parenchyma has a range from −1024 HU to −500 HU [8], therefore, being this the region of interest, this range is taken in a fixed thresholding for the extraction of this first mask. 0 if CT (x, y) > −500, M ask(x, y) = (1) 1 If CT (x, y) ≤ −500; This mask retains mainly parenchyma and air. To remove the external air to the body, a series of morphological operations are performed.An erosion is applied using a structuring element with 26-neighborhood to remove the smallest connected components (here called objects), as well as to separate the lungs from the connection they have with the external air through the nostrils or mouth in the case of full body CT scans Fig. 2. Subsequently, the largest object, which corresponds to the external air is deleted, obtaining a mask that corresponds to the body. By calculating the intersection of the first mask, with the body mask is obtained a preliminary segmentation of the lungs Fig. 3 (e,f). The largest object is extracted, which in this case corresponds to the lungs, thus eliminating noise of the image. Due to threshold segmentation, some areas within our region of interest were deleted. So, in order to obtain our final segmentation, the internal contours of the lung are detected and filled. Thus obtaining the lung segmentation mask Fig. 4.
Lung Segmentation and COVID-19 Classification
427
(a) Pre-erosion (b) Eroded imimage. age.
Fig. 2. Erosion operation on a full body CT scan. It can be seen in (a) the mouth and nostrils (in red bounding box), which are eliminated when the erosion is applied (b). (Color figure online)
Fig. 3. Preliminary lung segmentation: (a, d) First mask obtained from the application of fixed thresholding; (b, e) body mask, resulting from the elimination of the external air from the image; (c, f) result of the intersection of the two masks it can be observed in the sagittal section some objects that do not belong to the lung (in red bounding box). (Color figure online)
428
L. E. Gaeta-Ledesma and F. J. Alvarez-Padilla
(a) Sagittal section.
(b) Axial section.
Fig. 4. Lung segmentation. It can be seen that its taken all the space delimited by the pleura, separated from the rib cage (i.e. internal lung mask information).
2.2
Classification
Once we have obtained the segmentation masks of our images, we are ready for the classification stage, which is divided into two parts: in the first one we will obtain the characteristics that will feed the SVM, while the second one is the training of them and their consequent evaluation. Characterization Once the lung segmentation is performed, it is time to characterize the region of interest. Two types of descriptors are applied for this purpose: the application of a Gabor filter and GLCM features. The open-source package pyradiomics [9], is used for this purpose, which extracts several features focused on medical images. The application of the Gabor filter is justified in its wide use as a method of describing spectral features. Since GGOs mainly modify the intensity of certain lung regions, it is expected that the application of this filter are able to generate features that highlight the presence of this sign. The second type of descriptors are obtained from the GLCM, which gives us second order characteristics, relating the frequency of our image intensities to their spatial location. These characteristics are obtained by calculating how often a spatial relationship occurs between two voxels with a specific intensity, thus giving information about texture patterns in the image.Once both sets of features are obtained for each image, the results are stored in two dataframes, ready for the training stage.
Lung Segmentation and COVID-19 Classification
429
Training and Evaluation With the obtained descriptors, the classification algorithm is fed. For this purpose, a SVM is used and configured as follows: – Parameter regularization:: 1. – Kernel function: Linear . – Class weight: Balanced, inversely proportional to class frequencies in the input data . – Iteration limit: No limit. – Gamma: 1 Accuracy and precision metrics are calculated using 5-fold cross validation. This metrics are defined as: Accuracy =
TP + TN TP + TN + FP + FN
Precision =
TP TP + FP
where TP, TN, FP and FN represent True Positive, True Negative, False Positive, and False Negative, respectively.
3
Experimentation and Results
30 positive and 15 negative COVID-19 CT images. The positive ones were taken from the National Genomics Data Center [10]. These are classified as follows according to the stages proposed by Pan et al. (2020) [11]: 1. Early stage: 0 - 4 days after the first symptoms, distributed GGO are present in the lower lobes unilaterally or bilaterally. 2. Progressive stage: 5 - 8 days after the first symptoms, the infection presents more aggressively in different lobules, with GGO, crazy-paving pattern, and consolidation. 3. Peak stage: 9 -13 days ter the first symptoms, the infected lung area presents dense consolidations in a prevalent manner, GGO are found diffusely and residual parenchymal bands. 4. Absorption stage: after 14 days of the first symptoms, the consolidations are absorbed and an extensive GGO is observed as a result of this absorption. The negative ones, taking advantage of a local database, from the Institut Godinot, Reims, France of patients with other pathologies that do not affect internal lung data. The images were performed in auto mA mode with adaptive statistical iterative reconstruction, native collimation of 16 × 1.25 mm, resolution matrix of 512 × 512 × 351 in chest CTs and 512 × 512 × 680 in full body, with a voxel size of 0.97 × 0.97 × 0.5 mm3 .
430
L. E. Gaeta-Ledesma and F. J. Alvarez-Padilla
Adding as stage 0 the negative images to COVID-19, the distribution of our database is as follows: 0. 1. 2. 3. 4.
23% in negative stage. 14% in early stage. 30% in progressive stage. 30% in peak stage. 3% in absorption stage.
The classification is made taking stages 0 to 3, stage 4 was left aside due to the few images available in our dataset corresponding to it. Automated segmentations are evaluated using manually performed segmentations as ground truth. The Sorensen-Dice and Jaccard coefficients were used to evaluate the similarity between the automatic segmentation and the ground truth, the higher the coefficient, the more similarity there is between the segmentations. The results are presented in Table 1 divided into COVID-19 positive and negative images. The SVM was fed using these groups of features: 1. 2. 3. 4.
All features. Gabor Features. GLCM Features. Top 4 Gabor features according to their performance in precision and accuracy on an individual basis (Gabor4 ). 5. Top 4 GLCM features according to their performance in precision and accuracy on an individual basis (GLCM4 ). 6. Best Gabor and GLCM features at the same time (Gabor4 + GLCM4 ).
The precision and accuracy of these groups of characteristics are presented in Table 2, dividing the accuracy (A) and the precision (P) for each of the stages. Finally, Tables 3 and 4 presents the individual results for each type of feature. Table 1. Segmentation ratios. (a) COVID-19. Image 1 2 3 4 5 6 7 8 9 10
(b) NO COVID-19.
Sorensen
Jaccard
Image
0.9792 0.9834 0.9765 0.9502 0.9681 0.9448 0.9417 0.9526 0.9457 0.9702
0.9593 0.9637 0.9534 0.9312 0.9421 0.9234 0.9203 0.9321 0.9232 0.9567
1 2 3 4 5 6 7 8 9 10
Sorensen
Jaccard
0.9553 0.9745 0.9434 0.9681 0.9478 0.9513 0.9607 0.9691 0.9355 0.9319
0.9145 0.9456 0.9134 0.9231 0.8986 0.9145 0.9234 0.9201 0.8934 0.8867
Lung Segmentation and COVID-19 Classification
431
Table 2. Results of feature grouping from SVM classification. Group of features Stage 0 A P
Stage 1 A P
Stage 2 A P
Stage 3 A P
All features
0.77 0.84 0.84 0.84 0.77
0.91 0.75 0.25
All gabor
0.85 0.85 0.87 0.82 0.88
0.86 0.85 0.5
All GLCM
0.77 0.57 0.82 0.76 0. 77 0.91 0.73 0.3
Gabor4
0.82 0.91 0.81 0.84 0.88
0.86 0.76 0.3
GLCM4
0.83 0.91 0.83 0.64 0.88
0.86 0.77 0.25
Gabor4 +GLCM4 0.84 0.91 0.83 0.76 0.86
0.86 0.84 0.3
Table 3. Results of gabor filter features from SVM classification. Features
Stage 0 A P
10 Percentile 90 Percentile Energy
Stage 1 A P
Stage 2 A P
Stage 3 A P
0.68 0.88 0.81 0.88
0.9
0.78
0.77 0.21
0.68 0.88 0.83 0.78
0.9
0.78
0.74 0.21
0.74 0.58 0.66 0.59
0.67 0.56
Entropy
0.77 0.9
Interquartile range
0.87 0.88 0.9
0.48 0.78 0.78
0.86 0.88 0.78
0.75 0
0.56 0.78
0.78 0.21
0.93 0.78
0.77 0.21
Kurtosis
0.8
Maximum
0.77 0.59 0.88 0.52
0.92 0.52
0.76 0
Mean absolute deviation
0.9
0.9
0.77 0.21
Mean
0.68 0.81 0.6
0.9
0.88 0.78 0.73
0.93 0. 78 0.77 0.8 0.78
0.66 0.56
0.77 0
Median
0.68 0.43 0.88 0.42
0.93 0.56
0.77 0
Minimum
0.77 0.55 0.70 0.56
0.77 0.56
0.77 0
Range
0.77 0.59 0.56 0.56
0.57 0.56
0.77 0
0.78
0.56 0.78
0.77 0.21
0.77 0.88 0.66 0.88
0.71 0.74
0.71 0.18
Robust mean absolute deviation 0.77 0.88 0.5 Root mean squared Skewness
0.66 0.81 0.88 0.45
0.91 0.74
0.78 0.75
Total energy
0.68 0.62 0.88 0.74
0.9
0.68
0.77 0.15
Uniformity
0.79 0.9
0.88 86.33 0.89 0.78
0.77 0.21
Variance
0.80 0.66 0.88 0.88
0.93 0.74
0.77 0.16
432
L. E. Gaeta-Ledesma and F. J. Alvarez-Padilla Table 4. Results of GLCM Features from SVM classification.
Features
Stage 0 A P
Stage 1 A P
Stage 2 A P
Stage 3 A P
Autocorrelation
0.64 0.69 0.77 0.58 0.88 0.68 0.77 0.1
Cluster prominence
0.74 0.65 0.63 0.56 0.66 0.56 0.74 0
Cluster shade
0.74 0.64 0.75 0.64 0.78 0.56 0.77 0.25
Cluster tendency
0.92 0.80 0.75 0.67 0.76 0.74 0.77 0.87
Contrast
0.7
0.64 0.88 0.69 0.9
Correlation
0.8
0.90 0.9
Difference average
0.77 0.64 0.8
0.65 0.82 0.63 0.77 0
Difference entropy
0.77 0.64 0.8
0.72 0.8
Difference variance
0.70 0.64 0.88 0.63 0.9
0.7
0.7
0.77 21.42
0.93 0.78 0.77 21.42 0.63 0.77 0 0.78 0.77 21.42
Inverse difference
0.77 0.55 0.46 0.72 0.48 0.56 0.77 0
Inverse difference moment
0.71 0.55 0.46 0.56 0.48 0.56 0.77 0
Inverse difference moment normalized 0.77 1
0.87 0.6
0.9
0.64 0.77 0.21
Inverse difference normalized
0.77 0.92 0.82 0.61 0.88 0.6
Informative measure of correlation 1
0.80 0.9
0.91 0.62 0.93 0.74 0.77 0.21
Informative measure of correlation 2
0.8
0.83 0.69 0.86 0.71 0,78 0.21
Inverse variance
0.77 0.56 0.50 0.56 0.52 0.56 0.77 0
Joint average
0.64 0.69 0.76 0.63 0.77 0.67 0.79 0.1
Joint energy
0.68 0.70 0.74 0.56 0.76 0.56 0.77 0
0.9
0.77 0
Joint entropy
0.68 0.57 0.66 0.62 0.68 0.56 0.74 0
Maximal correlation coefficient
0.71 0.88 0.86 0.71 0.88 0.74 0.78 0.18
Maximum probability
0.71 0.76 0.78 0.62 0.8
0.56 0.74 0
Sum average
0.64 0.69 0.77 0.63 0.8
0.67 0.77 0.1
Sum entropy
0.72 0.8
0.64 0.77 0.2
Sum Squares
0.89 0.83 0.76 0.69 0.78 0.70 0.77 0.75
4
0.89 0.62 0.9
Conclusions
The proposed lung segmentation method, on both chest and full body CT, had robust results in both Sorensen’s and Jaccard’s indices (0.9681 and 0.942) mainly in early stages of the disease. However, due to the presence of GGOs consolidations (soft tissue density), which are mostly discarded by the segmentation, the effectiveness of the classification reduces dramatically in samples of stage 3 of COVID-19, as can be seen in the Table 2. It is also noted that for this last reason, we excluded from the SVM modeling samples from stage 4 of COVID-19. The features that have more weight when classifying the images are those coming from the Gabor filter giving better results in all stages those that give us information about the distribution of intensities within the image, such as kurtosis or
Lung Segmentation and COVID-19 Classification
433
variance, and being less useful for the classification by stages those that provide information about the intensities as a whole, such as mean, median, maximum or range. While most of the features obtained from the GLCM have an precision less than 60, where the features with a higher precision are those that leave aside the spatial description and focus more on intensity distribution, such as contrast, correlation or difference variance. However, selecting for training those images with prevalent GGO (stages 1 and 2), the accuracy using the GLCM features increases; so that texture descriptors become a discriminative feature when characteristic disease (GGO) are scattered throughout the region of interest. Indeed, based on this experimentation, GLCM features do not improvesignificantly the results of the Gabor features, even taking into account only their best performers being negligible when using the two types at the same time.
References 1. Ding, X., et al.: Wearable sensing and telehealth technology with potential applications in the coronavirus pandemic. IEEE Rev. Biomed. Eng. 14(1), 48–70 (2020) 2. Bao, C., Liu, X., Zhang, H., et al.: Coronavirus disease 2019 (COVID-19) CT findings: a systematic review and meta-analysis. J. Am. Coll. Radiol. 17(6), 701– 709 (2020) 3. Xu, X., Jiang, X., et al.: A deep learning system to screen novel coronavirus disease 2019 pneumonia. Engineering 6(10), 1122–1129 (2020) 4. Zhang, H., et al.: Automated detection and quantification of COVID-19 pneumonia: CT imaging analysis by a deep learning-based software. Eur. J. Nucl. Med. Mol. Imaging 47(11), 2525–2532 (2020). https://doi.org/10.1007/s00259-020-04953-1 5. Shorten, C., Khoshgoftaar, T.M., Furht, B.: Deep Learning applications for COVID-19. J. Big Data 8(1), 1–54 (2021). https://doi.org/10.1186/s40537-02000392-9 6. Sun, L., Mo, Z., et al.: Adaptive feature selection guided deep forest for COVID-19 classification with chest CT. IEEE J. Biomed. Health Inform. 24(10), 2798–2805 (2020) 7. Tseng, L.-Y., Huang, L.-C.: An adaptive thresholding method for automatic lung segmentation in CT images. In: AFRICON 2009, pp. 1-5. IEEE (2009) 8. Brown, M.S., et al.: Method for segmenting chest CT image data using an anatomical model: preliminary results. IEEE Trans. Med. Imaging 16(6), 828–839 (1997) 9. Van Griethuysen, J.J., Fedorov, A., et al.: Computational radiomics system to decode the radiographic phenotype. Can. Res. 77(21), e104–e107 (2017) 10. Ning, W., Lei, S., Yang, J., et al.: Open resource of clinical data from patients with pneumonia for the prediction of COVID-19 outcomes via deep learning. Nat. Biomed. Eng. 4(12), 1197–1207 (2020) 11. Pan, F., et al.: Time course of lung changes at chest CT during recovery from coronavirus disease 2019 (COVID-19). Radiology 295(3), 715–721 (2020)
IOT in Health and Bioinstrumentation
Disinfection Method Based on UV-C Light Using the Internet of Things for Cleaning Hospital Areas (COVID-19) Stephanie Carolina Ju´ arez-Garc´ıa1 , Misael S´ anchez-Magos1,2 , 2(B) , Christi Torres-Vargas1 , Iv´ an Matehuala-Mor´ an 1 ´ , Ricardo Bautista Mercado1 , Francisco Mu˜ noz del Angel Juan Jes´ us Mej´ıa Fern´ andez1 , and Fanny Alvarado1 1
Biomedical Engineering Department, Instituto Nacional de Ciencias M´edicas y Nutrici´ on Salvador Zubir´ an, Av. Vasco de Quiroga No. 15, Col. Belisario Dom´ınguez Secci´ on XVI, 14080 Ciudad de M´exico, M´exico 2 Tecnologico de Monterrey, School of Engineering and Science, Calle del Puente 222, Col. Ejidos de Huipulco Tlalpan, 14380 Ciudad de M´exico, M´exico [email protected]
Abstract. Since the SARS-CoV-2 transmission can occur by contact with surfaces contaminated with respiratory secretions and other fluids like faeces or saliva, the superficial disinfection has been one of the main problems during the COVID-19 pandemic. Cross-contagion has been observed between health personnel and cleaning staff from hospitals attending COVID-19 patients. The problem was solved through the implementation of a contact-less disinfection system that reduces the COVID-19 exposition of sanitation workers from healthcare facilities. This work presents the results observed from the implementation of an Ultraviolet-C (UV-C) disinfection method controlled and monitored using an Internet of Things (IoT) scheme. Also, implementation experiences obtained from the application of the proposed solution at the Instituto Nacional de Ciencias M´ edicas y Nutrici´ on Salvador Zubir´ an (INCMNSZ) are discussed in this article. The main contribution of this work relies in the fulfillment of a disinfection proceeding that helps reducing the cross-contagion between the cleaning staff of hospitals attending the COVID-19 pandemic. Keywords: SARS-CoV-2 · COVID-19 of things · Disinfection system
1
· Ultraviolet-C light · Internet
Introduction
The emergent pathogen SARS-CoV-2 has provoked an acute health crisis due to the COVID-19 disease that this virus causes [1]. Healthcare systems have collapsed in many countries because the disease is hard to diagnose. But to c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 C. J. Trujillo-Romero et al. (Eds.): CNIB 2022, IFMBE Proceedings 86, pp. 437–447, 2023. https://doi.org/10.1007/978-3-031-18256-3_47
438
S. C. Ju´ arez-Garc´ıa et al.
reduce the number of infections, some protective measures have been taken [2]. These include sneezing into a handkerchief, keeping a safe distance between people and frequent hand washing [3,4]. The two most usual ways by which the virus is spread throughout the body are inhaling droplets or aerosols suspended in the air that was produced by the respiration and speech of asymptomatic individuals [4] and being in contact with surfaces that are contaminated by respiratory secretions or other bodily fluids like mucus, saliva, feces, etc. One factor to take into account is that although the most common symptoms of COVID-19 are fever, dry cough and tiredness [5,6], in most of the cases the subject starts gradually presenting mild symptoms [7]. Other factors are at-risk groups, such as elder people [8] or patients with chronic conditions like high blood pressure, cancer, diabetes or heart or lung conditions [9], because they are more predisposed to present severe symptoms which include pneumonia, organ failure, heart problems, blood clots, among others which may complicate the patient’s situation [10,11] or lead to death [12]. One of the challenges that the countries are facing is that the population thinks that given the measures taken (social distancing [13], isolation [14], travel restrictions [15], obligatory quarantine [16], etc.), the COVID-19 has been controlled. Additionally, the World Health Organization (WHO) has reported that in spite of the safety protocols implemented, around 10% of the worldwide infection rate correspond to medical staff [17]. Along with the increasing demand for hospital’s beds due to COVID-19; these situations caused several shutdowns, from healthcare facilities to laboratories all over the world [18,19]. Therefore, an alternative solution, alongside vaccines, must be implemented to contain the contagion. And given that most cases occur because the person is in contact with surfaces that are infected [20] or with droplets that contain infected mucous secretions [21], then the solution must be directly related to disinfect environmental surfaces. An effective method to do so is UV-C radiation, which consists of a high-energy type of UV with a short wavelength that allows to disinfecting water, air and non-porous surfaces [22]. Moreover, UV-C can kill diverse types of bacteria, destroying the outer protein membrane of several viruses and changing their genetic material to induce their inactivation. The wavelength of this type of radiation is in the interval between 200 and 280 nm [23], but the highest efficiency against COVID-19 occurs when the value is 254 nm [24]. In addition, the vast majority of germicidal lamps also work at this value of wavelength because of its efficiency [25]. But this does not necessarily imply that longer or shorter wavelengths cannot be used to disinfect surfaces [26]. Table I shows the doses of UV-C needed to reach the Log inactivation of different bacteria, spores and viruses. 1.1
Paper Distribution
This work presents an IoT-based solution for the surface disinfection inside hospitalization rooms of healthcare facilities attending COVID-19 patients. Disinfection is carried out by a UV-C mobile platform allocated in the highly transited areas without the necessity of costly or complex installations. The proposed
Disinfection Method for Cleaning Hospital Areas
439
Table 1. UV-C dose to accomplish a logarithmic reduction without photo-reactivation on spores, bacteria and virus. UV-C dose for Log reduction (mJ/cm2 ) 1* 2* 3* 4* 5* Spores Aspergillius niger
132
−
−
−
−
Bacillus subtilis
20
39
60
81
−
Bacteria Escherichia coli
4.4
6.2
7.3
8.1
9.2
Pseudomonas stutzeri
100
150
195
230
−
Staphylococcus aureus
3.9
5.4
6.5
10.4
−
Compylobacter jejuni
1.6
3.4
4
4.6
5.9
Salmonella spp.