Functional Brain Mapping: Methods and Aims [1st ed.] 9789811568824, 9789811568831

Citation preview

Brain Informatics and Health

Vassiliy Tsytsarev Vicky Yamamoto Ning Zhong   Editors

Functional Brain Mapping: Methods and Aims

Brain Informatics and Health Editors-in-Chief Ning Zhong, Department of Life Science & Informatics, Maebashi Institute of Technology, Maebashi-City, Japan Ron Kikinis, Department of Radiology, Harvard Medical School, Boston, MA, USA Series Editors Weidong Cai, School of Computer Science, The University of Sydney, Sydney, NSW, Australia Henning Müller Switzerland

, University of Applied Sciences Western Switzerland, Sierre,

Hirotaka Onoe, Graduate School of Medicine, Kyoto University, Kobe, Japan Sonia Pujol, Department of Radiology, Harvard Medical School, Boston, MA, USA Philip S. Yu, Department of Computer Science, University of Illinois at Chicago, Chicago, IL, USA

Informatics-enabled studies are transforming brain science. New methodologies enhance human interpretive powers when dealing with big data sets increasingly derived from advanced neuro-imaging technologies, including fMRI, PET, MEG, EEG and fNIRS, as well as from other sources like eye-tracking and from wearable, portable, micro and nano devices. New experimental methods, such as in toto imaging, deep tissue imaging, opto-genetics and dense-electrode recording are generating massive amounts of brain data at very fine spatial and temporal resolutions. These technologies allow measuring, modeling, managing and mining of multiple forms of big brain data. Brain informatics & health related techniques for analyzing all the data will help achieve a better understanding of human thought, memory, learning, decision-making, emotion, consciousness and social behaviors. These methods also assist in building brain-inspired, human-level wisdom-computing paradigms and technologies, improving the treatment efficacy of mental health and brain disorders. The Brain Informatics & Health (BIH) book series addresses the computational, cognitive, physiological, biological, physical, ecological and social perspectives of brain informatics as well as topics relating to brain health, mental health and well-being. It also welcomes emerging information technologies, including but not limited to Internet of Things (IoT), cloud computing, big data analytics and interactive knowledge discovery related to brain research. The BIH book series also encourages submissions that explore how advanced computing technologies are applied to and make a difference in various large-scale brain studies and their applications. The series serves as a central source of reference for brain informatics and computational brain studies. The series aims to publish thorough and cohesive overviews on specific topics in brain informatics and health, as well as works that are larger in scope than survey articles and that will contain more detailed background information. The series also provides a single point of coverage of advanced and timely topics and a forum for topics that may not have reached a level of maturity to warrant a comprehensive textbook.

More information about this series at http://www.springer.com/series/15148

Vassiliy Tsytsarev Vicky Yamamoto Ning Zhong •

•

Editors

Functional Brain Mapping: Methods and Aims

123

Editors Vassiliy Tsytsarev University of Maryland Baltimore, MD, USA

Vicky Yamamoto University of Southern California West Hollywood, USA

Ning Zhong Department of Information Engineering Maebashi Institute of Technology Maebashi, Japan

ISSN 2367-1742 ISSN 2367-1750 (electronic) Brain Informatics and Health ISBN 978-981-15-6882-4 ISBN 978-981-15-6883-1 (eBook) https://doi.org/10.1007/978-981-15-6883-1 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 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 Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Preface

Neurobiology is developing rapidly: new, more effective methods are being developed, increasingly less traumatic methods of neurovisualization are being offered, and new molecular probes are used to study brain functioning. Brain imaging technology has a long history. Usually it is customary to take autumn 1895 as the starting point, when the German physicist Wilhelm Roentgen discovered mysterious X-rays that can enlighten human body, including brain, through and through. X-ray experiments gave a huge impetus to the development of medicine— for the first time, doctors were able to look inside a person without opening it. After many years, MRI and fMRI appeared in the scientific arsenal—magnetic resonance imaging, as well as functional magnetic resonance imaging invented in 1973 by chemists Paul Lauterbur and Peter Mansfield. This event marked the beginning of the era of functional brain imaging in the modern sense of the word. It should be noted that in clinical practice, including neurological and mental diseases the success of treatment is still largely determined by correct diagnosis and monitoring of the effectiveness of therapy. These tasks are solved using various brain imaging methods. Accepted to associate them with the representation of the particular functions in the brain, but in biological reality they are much more complicated. Currently brain diseases comprised of neurodegenerative, behavioral, neurodevelopmental and mental disorders together represent leading causes of morbidity and mortality in over the world but in spite of amazing scientific efforts to elucidate the disease and to develop adequate therapies, our understanding remains incomplete and our therapeutic options are still limited. Further development of the therapy of brain diseases is impossible without the consistent improvement of animal models of diseases of the neural system. Work with animal models, in turn, requires new methods of brain imaging, which have higher spatial and temporal resolution as well as chemical selectivity. The book is a brief overview of the current state of neuroscience, primarily in terms of clinical, preclinical and computational methods of brain imaging. The book includes newly updated information about the imaging structure and functions of the nervous system of both humans and the main used as animal model for studying the principles of the brain development and functioning. Particular v

vi

Preface

attention is given to those areas of the brain science that have been developing most intensively in recent years: molecular imaging, functional brain imaging, neurophotonics, optogenetics and neurocomputing. The variety of imaging technologies is great, and it is not possible to describe all of them in one volume so the most common and interesting ones were chosen for this book. We aimed to provide the reader with the opportunity to receive the information about the relationship of the structure of the brain and physiological functions in a clear and accessible form. The book is intended for students of biological, psychological and medical schools, scientists, all interested in neurobiology and working in this field. Baltimore, USA West Hollywood, USA Maebashi, Japan

Vassiliy Tsytsarev Vicky Yamamoto Ning Zhong

Contents

Principles of Functional Brain Imaging General Descriptions on MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hidenao Fukuyama, Tomohiro Ueno, and Hector Sanchez Lopez Neurons and Plasticity: What Do Glial Cells Have to Do with This? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nicolangelo Iannella and Michel Condemine

3

13

Molecular Brain Imaging Transcranial Dynamic Fluorescence Imaging for the Study of the Epileptic Seizures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vyacheslav Kalchenko, Alon Harmelin, David Israeli, Babak Kateb, Igor Meglinski, Qinggong Tang, Nitish V. Thakor, Alla Ignashchenkova, Anna Volnova, and Vassiliy Tsytsarev Critical Elements for Connectivity Analysis of Brain Networks . . . . . . . Jean Faber, Priscila C. Antoneli, Noemi S. Araújo, Daniel J. L. L. Pinheiro, and Esper Cavalheiro

49

67

Brain Optical Imaging Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art with Emphasis on Pre-clinical and Clinical Studies . . . . . . . . . . . . . . . . 111 Ron D. Frostig Implantable CMOS Fluorescent Imaging Devices . . . . . . . . . . . . . . . . . 129 Kiyotaka Sasagawa, Makito Haruta, Yasumi Ohta, Hironari Takehara, Takashi Tokuda, and Jun Ohta

vii

viii

Contents

Neural Computation and Data Analysis Data Analysis Method for Neuroimaging Data: Task-Related Component Analysis and Its Applications to fNIRS Data . . . . . . . . . . . 149 Hirokazu Tanaka, Takusige Katura, and Hiroki Sato Towards Automated Processing and Analysis of Neuronal Big Data Acquired Using High-Resolution Brain-Chip Interfaces . . . . . . . . . . . . . 175 Mufti Mahmud, Claudia Cecchetto, Marta Maschietto, Roland Thewes, and Stefano Vassanelli Conclusion and Future Work Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 Vassiliy Tsytsarev

Editors and Contributors

About the Editors Vassiliy Tsytsarev holds a Ph.D. in Neuroscience from Saint Petersburg State University, Russia. Soon after graduation he moved to Japan and began working at the Brain Science Institute of RIKEN, and the Human Brain Research Center, Kyoto. Functional brain mapping, neural circuits and different types of brain optical imaging are his main scientific interests. In Japan, Vassiliy worked in the field of auditory neuroscience using intrinsic optical imaging (IOS) and voltage-sensitive dye imaging. After seven years in Japan he moved to the United States, where he has worked at several universities; for the past six years, at the University of Maryland School. His current focus is on functional brain mapping, epileptic studies and neural network function in the rodent somatosensory system, which offers a perfect specimen for many types of neuroscience research, including models of neural diseases. Vassiliy is the author and co-author of more than 40 publications in peer-reviewed magazines, and several book chapters. He is a senior editor for the Journal of Neuroscience and Neuroengineering, serves on the board of directors of the Society for Brain Mapping and Therapeutics (SBMT), and on the editorial boards of other scientific journals. Dr. Vicky Yamamoto is a cancer scientist at USC-Keck School of Medicine in the Department of Otolaryngology/Head and Neck Surgery with more than 15 years of research experience ranging from developmental neurobiology and stem cell to molecular targeted therapy. Prior to joining USC, Dr. Yamamoto worked at Mount St. Mary’s College, The Scripps Research Institute, Cedars-Sinai Medical Center, and California Institute of Technology. She has significant teaching experience and mentored numerous students. She graduated from Mount St. Mary’s College with a BS in biological sciences and a BA in chemistry. She received a Ph.D. in biochemistry and molecular biology from Keck School of Medicine of USC. Dr. Yamamoto’s pioneering work on a Wnt co-receptor was conducted at the laboratory of David Baltimore, a Nobel laureate-1975, at California Institute of

ix

x

Editors and Contributors

Technology. She received a prestigious fellowship from California Institute for Regenerative Medicine (CIRM) to investigate key mechanism that regulates the development of stem cells into neurons. Dr. Yamamoto has been a founding member of the Society for Brain Mapping and Therapeutics (SBMT) and currently serving as an executive director. She is also an active board member of the Brain Mapping Foundation. Prof. Ning Zhong is currently head of Knowledge Information Systems Laboratory, and a professor in Department of Life Science and Informatics at Maebashi Institute of Technology, Japan. He is also director and an adjunct professor in the International Web Intelligence Consortium (WIC) Institute, Beijing University of Technology, China

Contributors Priscila C. Antoneli Escola Paulista de Medicina (EPM), Department of Neurology and Neurosurgery, Discipline of Neuroscience, Lab. of Neuroengineering and Neurocognition, Universidade Federal de São Paulo— UNIFESP, São Paulo, Brazil Noemi S. Araújo Escola Paulista de Medicina (EPM), Department of Neurology and Neurosurgery, Discipline of Neuroscience, Lab. of Neuroengineering and Neurocognition, Universidade Federal de São Paulo—UNIFESP, São Paulo, Brazil Esper Cavalheiro Escola Paulista de Medicina (EPM), Department of Neurology and Neurosurgery, Discipline of Neuroscience, Lab. of Neuroengineering and Neurocognition, Universidade Federal de São Paulo—UNIFESP, São Paulo, Brazil; Centro Nacional de Pesquisa em Energia e Materiais CNPEM, Campinas, Brazil Claudia Cecchetto Okinawa Institute of Science and Technology, Okinawa, Japan Michel Condemine Condemine Consulting, Paris, France Jean Faber Escola Paulista de Medicina (EPM), Department of Neurology and Neurosurgery, Discipline of Neuroscience, Lab. of Neuroengineering and Neurocognition, Universidade Federal de São Paulo—UNIFESP, São Paulo, Brazil Ron D. Frostig Departments of Neurobiology and Behavior and Biomedical Engineering, The Center for the Neurobiology of Learning and Memory, University of California, Irvine, CA, USA Hidenao Fukuyama Graduate School of Medicine, Kyoto University, Kyoto, Japan

Editors and Contributors

xi

Alon Harmelin Vice President for Administration and Finance, Weizmann Institute of Science, Rehovot, Israel Makito Haruta Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan Nicolangelo Iannella Department of Biosciences, University of Oslo, Oslo, Norway Alla Ignashchenkova Saint Petersburg State University, Translational Research Institute, Saint Petersburg, Russia David Israeli Psychiatric Array, Kaplan Hospital, The Hebrew University of Jerusalem, Jerusalem, Israel Vyacheslav Kalchenko Department of Veterinary Resources, Weizmann Institute of Science, Rehovot, Israel Babak Kateb Founding Chairman of the Board of Directors, CEO and Scientific Director, President & Scientific Director, Society for Brain Mapping & Therapeutics (SBMT), Brain Mapping Foundation, Pacific Palisades, CA, USA; Society of Brain Mapping and Therapeutics, Santa Monica, CA, USA; Brain Mapping Foundation, West Hollywood, CA, USA; National Center for NanoBioElectronics, Los Angeles, CA, USA; Brain technology and Innovation Park, Los Angeles, CA, USA Takusige Katura NeU Corporation, Creation Work Unit, Japan Hector Sanchez Lopez School of Information Technology and Electrical Engineering, The University of Queensland, Brisbane, QLD, Australia Mufti Mahmud Department of Computing and Technology, Nottingham Trent University, Nottingham, UK Marta Maschietto Department of Biomedical Sciences, University of Padova, Padova, Italy Igor Meglinski School of Engineering and Applied Science, Aston University, Birmingham, United Kingdom Jun Ohta Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan Yasumi Ohta Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan Daniel J. L. L. Pinheiro Escola Paulista de Medicina (EPM), Department of Neurology and Neurosurgery, Discipline of Neuroscience, Lab. of Neuroengineering and Neurocognition, Universidade Federal de São Paulo— UNIFESP, São Paulo, Brazil

xii

Editors and Contributors

Kiyotaka Sasagawa Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan Hiroki Sato Department of Bioscience and Engineering, College of Systems Engineering and Science, Shibaura Institute of Technology, Koto City, Tokyo, Japan Hironari Takehara Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan Hirokazu Tanaka Faculty of Information Technology, Department of Intelligent Systems, Tokyo City University, Setagaya, Tokyo, Japan Qinggong Tang Stephenson School of Biomedical Engineering, University of Oklahoma, Norman, OK, USA Nitish V. Thakor Department of Biomedical Engineering, John Hopkins University, Baltimore, USA Roland Thewes Sensor and Actuator Systems, Technische Universität Berlin, Berlin, Germany Takashi Tokuda Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan; Future Interdisciplinary Research of Science and Technology (FIRST), Tokyo Institute of Technology, Tokyo, Japan Vassiliy Tsytsarev Department of Anatomy and Neurobiology, University of Maryland, Baltimore, MD, USA Tomohiro Ueno Graduate School of Medicine, Kyoto University, Kyoto, Japan Stefano Vassanelli Department of Biomedical Sciences, University of Padova, Padova, Italy Anna Volnova Saint Petersburg State University, Translational Research Institute, Saint Petersburg, Russia

Principles of Functional Brain Imaging

General Descriptions on MRI Hidenao Fukuyama, Tomohiro Ueno, and Hector Sanchez Lopez

1 What is an MRI? MRI is an abbreviation for Magnetic Resonance Imaging. The magnetic resonance imaging represents measurement equipment and a method to obtain an image of magnetic resonance signals. The magnetic resonance signals indicate signals of Nuclear Magnetic Resonance (NMR). Since the word image means a visual representation, the image to be obtained is a spatial distribution of NMR signals. In this sense, signal acquisition procedure in MRI consists of two parts: generation and observation of NMR signals, addition of spatial information to NMR signals. Following the two parts of the acquisition procedure, MRI measurement equipment can be classified into corresponding two components, NMR signal acquisition system and spatial encoding system. The NMR signal acquisition system is composed of a magnet, radio frequency (RF) coils and an RF spectrometer. The magnet is a superconducting magnet in most of cases and produces a static magnetic field, which generates a signal source of NMR. The RF coils and spectrometer are for generating, observing and manipulating NMR signals. On the other hand, the spatial encoding system is composed of magnetic field gradient coils and their divers and controllers, which add spatial information to generated NMR signals before their observation. H. Fukuyama (B) · T. Ueno Graduate School of Medicine, Kyoto University, 53 Shogoin-Kawahara-cho, Sakyo-ku, Kyoto 606-8507, Japan e-mail: [email protected] T. Ueno e-mail: [email protected] H. S. Lopez School of Information Technology and Electrical Engineering, The University of Queensland, Brisbane, QLD 4072, Australia e-mail: [email protected] © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_1

3

4

H. Fukuyama et al.

As we see in the previous paragraph, MRI is formed from NMR system and spatial encoding system. Thus, MRI is a modality to obtain a spatial distribution of NMR signals. Although the two system are mutually dependent, MRI image contrasts can be explained in the NMR system and imaging time in the spatial encoding system. MRI image quality is, however, affected by both systems. When we extract biological information from MRI images, we can focus on spatio-temporal behaviors of local NMR signals. In the next section, we see how local biological environment affects NMR signals.

2 What an NMR Signal Tells About Under a magnetic field, biological tissue becomes magnetized. When we apply a certain radio frequency pulse (short duration oscillating magnetic field), the magnetized tissue reacts to the pulse and send back a radio frequency pulse. This generated radio frequency pulse corresponds to an NMR signal. This phenomenon, however, occurs only on a predetermined frequency, resonant frequency. The resonant frequency is determined by magnetic field strength and a nucleus to be observed. Therefore, the NMR signal tell us what kinds of nuclei and how much those nuclei exist in a tissue, since the signal intensity is proportional to quantity of nuclei. In a usual MRI, quantity of protons is measured in a minimum unit volume (pixel or voxel) of an image. The information on quantity of nuclei is very similar to what other imaging modalities can offer, such as X-ray absorption in CT, reflection coefficients in Ultra Sound. What makes MRI stand out among other modalities is its ability to differentiate soft tissues with various contrasts. Moreover, we can relate changes in contrasts with biological functions. This ability comes from characteristics of NMR signals. Surrounding magnetic micro environments affect the NMR signals. These micro environment effects appear in temporal behaviors of the NMR signals and correspond to tissue characteristics. This means that a nucleus feels its surrounding environment, behaves accordingly and produces affected NMR signals. The micro environments are local magnetic fields generated by surrounding nuclei and interactions with the nuclei. For example, deoxyhemoglobin in a blood vessel affect magnetic field nearby. Therefore, when proportion of the deoxyhemoglobin changes in blood vessels, the nearby NMR signal changes accordingly. Since the proportion of the deoxyhemoglobin relates with tissue activities, we can extract functions of tissues from the local NMR signal changes. When the change in the deoxyhemoglobin comes from neuronal activities (BOLD effects), information on brain function can be obtained (fMRI). The case of deoxyhemoglobin corresponds to a time dependent change. Similar but static examples are quantities of iron and myelin in a brain tissue. An iron deposit and myelin have different magnetic susceptibilities from surrounding tissue. These differences modify local magnetic environments and local NMR signals. Thus, iron deposits due to aging and/or neurodegenerative disease and demyelination can be detected from the local NMR signal changes. In these

General Descriptions on MRI

5

examples, T 2 * , one of properties of the NMR signals, changes. The local magnetic susceptibility, χ, can be calculated from the NMR signal changes. The other example of magnetic micro environment change is that nucleus itself move to different environments. The nucleus movements can be diffusion or flow. The diffusion is a random movement and its effects appear as phase coherence of the NMR signal. The phase coherence means how much nuclei behave in the same way. During NMR signal acquisition, larger diffusion causes larger loss of the phase coherence then larger decrease of local NMR signal. Usually, the nuclei are in restricted geometry, such as inside and outside of cell. Since the restricted geometry restricts diffusion to a cell size during the signal acquisition, microstructure of tissue can be inferred from diffusion effects in the local NMR signals. Moreover, due to special one directional shape of axon, connectivity of neurons can be obtained from direction of diffusion, which is referred to as Diffusion Tensor Imaging, DTI. When the restricted geometry breaks, diffusion can not be limited. Thus, cerebral hemorrhage can be detected with diffusion. The other movement of nuclei is a flow. In a flow, nuclei move as a group to a certain direction such as direction of a blood vessel. Therefore, local nuclei move out of the local region and fresh nuclei come in. Local NMR signals decrease with regards to moving out nuclei and the signals increase in the case of coming in nuclei. Using this property, blood vessels can be detected. The above examples are that a micro magnetic environment is modified by other kinds of nuclei or by transportation of nuclei to other environments. Most commonly encountered micro magnetic environment changes are interactions between same kind of nuclei. These micro magnetic environments represent tissue properties and can be used to differentiate soft tissues. These interactions change two NMR signal properties, longitudinal relaxation time, T 1 , and transverse relaxation time, T 2 . In the transverse relaxation time, there is slightly different time constant, T 2 * . T 2 * includes other effects on T 2 , such as local magnetic field differences. T 1 , T 2 and T 2 * depend on tissue structures and magnetic field, and their dependences are different from each other. Therefore, combination of T 1 , T 2 and T 2 * make it possible to differentiate tissues.

3 Magnetic Fields in MRI In the previous section, we see microscopic magnetic environments. Since MRI has magnetic eyes, various macroscopic magnetic field components are used. For the NMR part, a static, usually large, homogeneous magnetic field component and a high frequency (radio frequency) magnetic field components are used. For the MRI imaging part, magnetic field gradients are used. The static homogeneous magnetic field is referred to as B0 field and generated by a large magnet, which has enough space for a human body or head inside. The homogeneous region corresponds to a field of view of MRI and is usually situated at the center of the inside space. A body or head to be imaged is placed at the center of the homogeneous region. The large magnet can be a permanent magnet or

6

H. Fukuyama et al.

a resistive magnet or a superconducting magnet. At higher fields above 0.3 T, the superconducting magnet is a usual choice. This static homogeneous magnetic field is utilized for producing an NMR signal source. The high frequency magnetic field is referred to as B1 field and classified into two fields: transmitting field (B1 + ) and receiving field (B1 −). The high frequency magnetic field uses oscillating magnetic field, which frequency linearly depends on the strength of the static magnetic field. In usual proton cases, 1.5 T static magnetic field needs 60 MHz, 3 T field 120 MHz. Its strength is, however, order of 10 μT and much weaker than that of the B0 field. The transmitting field is for manipulating NMR signals and the receiving field is the NMR signal itself. The transmitting field is generated by a saddle or bird cage coil and the receiving field is detected by the transmitter coil or specialized receiver coils. The specialized coils are a surface coil, a phased array coil (combination of surface coils) and a volume coil such as a solenoid coil. To increase signal to noise ratio, the receiver coil is placed closest to a body or head. The transmitter coil is placed outside of the receiver coil to ensure homogeneous transmitting field. The magnetic field gradients are additional magnetic fields, which strengths linearly depend on spatial positions, to make NMR signals as a function of position. There are three magnetic field gradients (Gx , Gy , Gz ), which correspond to three orthogonal directions of field changes, although the magnetic field direction is same as the B0 field. The field gradient null point is matched with the center of the imaging region. The magnetic field gradient coils are placed between the B0 magnet and RF coils (transmitter and receiver coils).

4 Principals of NMR There are two types of NMR methods, cw-NMR and pulsed NMR. In the cw-NMR, the B1 field is continuously applied. In the pulsed NMR, the B1 field is applied for just short duration of time but a few times depending on NMR signal properties to be obtained. In MRI, pulsed NMR is the standard method and explained in this section. The B0 field induces magnetization, which is an NMR signal source, in tissues of a head or body. The magnitude of the magnetization is proportional to the strength of local magnetic field. The main contributor to the local magnetic field is the B0 field so that 3 T MRI has twice larger signal source than 1.5 T MRI has. Strictly speaking, the magnetization is a property of a group of atoms or molecules in a tissue but it won’t cause any problem to understanding of MRI that each atom or molecule is treated as having own magnetization. The magnetization is a vector so that it has a direction and a magnitude. The induced magnetization aligns along the direction of the B0 field. The magnetization reacts to the B1 + field with predetermined frequency, which is called resonant frequency. The resonant frequency is determined by the B0 magnetic field and types of nucleus. When the B1 + field is applied with a certain duration and strength, the magnetization changes its direction from the original B0 field direction.

General Descriptions on MRI

7

This angle is referred to as a flip angle (FA). The flip angle is proportional to the product of the B1 field duration and strength. The flipped magnetization shows precession to the B0 field direction. The orthogonal component to the B0 direction corresponds to the B1 − field and induces oscillating voltage in the receiver coil. This induced voltage is the NMR signal to be measured. The oscillating frequency is proportional to the strength of local magnetic field. Due to local magnetic field differences, the NMR signal decays with time. This signal decay is referred to as free induction decay (FID). When 90° flip angle is used, the FID signal becomes largest since the orthogonal component becomes maximum. The flipped magnetization rotates according to the local magnetic field strength. Therefore, the strongest local field rotates the magnetization fastest and the weakest, slowest. If 180° flip angle is applied after the 90° pulse, the rotation difference is reversed. The fastest rotating magnetization is placed to the end of line, the slowest the top of line. After the time between 90 and 180° pulses passes, the fastest rotating magnetization catches up with the slowest one. During the catching up period, all the magnetization aligns little by little and an NMR signal appears accordingly. At the catching up moment, all the magnetization align along the same direction and the NMR signal reaches at maximum. Then directions of the magnetization lose focus and the NMR signal decays same as the FID signal. This NMR signal is referred to as an echo signal, especially spin echo signal. Instead of using 180° pulse, the rotation difference can be reversed by applying a magnetic field gradient. Since the magnetic field gradient dominates the local magnetic field difference, reversing the field gradient direction reverses the rotation speed difference. In the case of using the field gradient, the echo signal is referred to as the gradient echo signal. In the gradient echo, existing local magnetic field differences are not reversed and affect the signal. In addition, the first pulse can have any flip angle, not limited to the 90° pulse. The time between the first pulse and maximum echo signal is referred to as TE (time of echo) in MRI. Effects of T 2 and T 2 * on NMR signals can be controlled by changing TE in MRI. Since the 180° pulse reverses rotation differences of magnetization, effects of local magnetic fluctuations after the 180° pulse can be canceled out if the fluctuation are same as before the pulse. On the other hand, effects of the fluctuations that are different from those before the pulse remain. The ratio of these asymmetrical fluctuations increases in longer TE. Thus, a spin echo signal decrease more in longer TE than in shorter TE. Since T 2 correlates with inherent local magnetic field fluctuations, in a tissue with shorter T 2 , a larger signal decrease is observed. Using longer TE, a T 2 weighted signal can be obtained. The same situation exists in a gradient echo signal. In this case, however, T 2 * effects instead of T 2 effects appear since existing local magnetic field differences cannot be canceled out. Thus, a longer TE gradient echo signal will be a T 2 * weighted signal. The inherent local magnetic field fluctuations bring back the flipped magnetization to the equilibrium state in which all magnetization align along the direction of the B0 field. This process takes much longer time than T 2 or T 2 * since different components from those in the T 2 process contribute. This time constant is a longitudinal relaxation time, T 1 . The T 1 effects on the NMR signal can be controlled by the repetition time,

8

H. Fukuyama et al.

TR. In the case that NMR signals are obtained in a repeated manner, the repetition time is defined as a time between successive NMR pulse sequences. In shorter TR, the equilibrium state cannot be reached, but in longer TR, the equilibrium state can be accomplished. Using shorter TR in the NMR signal acquisition, the NMR signal intensity become stronger in a tissue with shorter T 1 than in that with longer T 1 . This shorter TR signal will be a T 1 weighted signal. Considering the recovering process, we should go back to the group property of the magnetization. The equilibrium state is recovered not by slowly rotating all the magnetization to the direction of the B0 field but by slowly placing the magnetization in order. As a group of atoms or molecules in a minimum unit of an MRI image, in which there are large number of atoms or molecules, the mean magnetization loses the flipped component, becomes zero and then gains the B0 field component during the recovering process. Although an NMR signal reflects microscopical atomic properties of a tissue, we cannot detect an NMR signal from a single atom and should treat as a group of atoms or molecules. Another way to obtain the T 1 weighted signal is applying 180° pulse before the first detection pulse. After 180° pulse, all the magnetization is flipped to the opposite direction and starts to recover to the equilibrium state with each T 1 in tissues. The T 1 weighted signal can be detected by applying a usual NMR sequence after some recovering time, which is much shorter than T 1 . This recovering time is referred as to the inversion time, TI in MRI. The first 180° pulse is referred to as the inversion pulse. When a magnetic field gradient is applied for short duration twice before and after the 180° pulse in the same direction, stationary magnetization does not accumulate rotation differences during the NMR signal acquisition. As for diffusing magnetization, the rotation differences cannot be canceled out since the diffused magnetization move to a different place from the original place and is affected by different strength of the magnetic field gradient. Thus, its NMR signal intensity decreases. Larger diffusion decreases the signal intensity more. Note that effects of the diffusion are originated only from displacement along the direction of the applied magnetic field gradient. This gradient applied signal is referred to as a diffusion weighted signal. The signal intensity decrease can be controlled by changing strength of the gradient field, its duration and separation time between two gradient pulses. These parameters are combined to one quantity, the b value. As we see, various NMR signal contrasts can be introduced by controlling acquisition parameters. However, it is impossible to single out one parameter in single acquisition. All signal contrasts are mutually dependent. That is the reason why the word, weighted, is used. There are considerable efforts to obtain quantitative parameters in MRI, which is referred to as quantitative mapping or imaging. With the quantitative imaging, comparisons among various institutes or MRI machines will be more accurate. There is another way to use NMR signals to identify tissue properties. Different molecular structures induce different local magnetic fields. Thus, different molecules have slightly different resonant frequencies. These frequency differences are referred to as chemical shifts, which are expressed in ppm. Note that the chemical shift comes

General Descriptions on MRI

9

from the same nuclear species such as proton and is much smaller than difference in resonant frequencies between different nuclear species. When an NMR signal is transformed to an NMR spectrum, a large water peak and various small peaks can be observed. These peaks correspond to each molecule and their intensities indicate their quantities. In MRI, this method is referred to as NMR spectroscopy and in the case of imaging as MR spectroscopy.

5 How Are Spatial Distribution of NMR Signals Obtained? In MRI, NMR signals with local tissue properties are placed in proper positions of an image space. The magnetic field gradients add spatial information to NMR signals. The spatial encoding of the magnetic field gradients is, however, different from that of an optical CCD camera. In the optical case, a spatially localized signal, which contains local information only, is obtained with a local CCD sensor and becomes a part of an optical image at the position of the sensor just like a piece of a jigsaw puzzle. On the other hand, the spatially encoded NMR signal, MRI signal, consists of many local NMR signals in which spatial weighting is different from each other. Thus, the single MRI signal contains the whole information of the field of view. By changing the strength of the magnetic field gradients, the spatial weighting in each local NMR signal changes and its changing rate depends on the position of the local NMR signal. After collecting all signals, MRI signals can be differentiated to local NMR signals using the changing rate and then transformed into an MRI image. In this transformation, the collected MRI signals are placed according to the strength and direction of the magnetic field gradients and Fourier transformed. The collected MRI signals are referred to as the k space data. The MRI sequence corresponds to the way how the k space data is collected and how the signal contrasts are added. There are three types of spatial encoding in MRI, in which applied magnetic field gradient are referred to as slice, read and phase gradients. The slice gradient is for spatially localizing MRI signals same as in the optical signals. The read gradient is for encoding spatial information into an MRI signal as the resonant frequency. The phase gradient is for encoding spatial information as a phase of a complex MRI signal. Combination of the three spatial encoding methods make it possible to obtain a 2D image and a 3D volume image. Basically, in a 2D image, all three encoding methods are used, and in a 3D image, a read gradient and two phase gradients are used. The slice gradient is applied at the same timing as an NMR detection pulse (the first pulse in a gradient echo case, and 90 and 180° pulses in a spin echo case). The slice gradient makes a resonant frequency as a function of position along the direction of the gradient. The envelop of the applied NMR detection pulse is modified to a Sinc function or gaussian to have a narrower frequency window. Only in the frequency window, the NMR pulse can work properly. Combining the slice gradient and the modified NMR pulse, only magnetization in the predetermined frequency window reacts and thus obtained spatial information of MRI signals is limited to the window

10

H. Fukuyama et al.

along the slice direction. The spatial width of the window is referred to as the slice thickness and determined by the strength of the slice gradient and the length of the modified NMR pulse. The read gradient is applied during echo signal acquisition. Since the gradient makes a resonant frequency as a function of position, the echo signal consists of local NMR signals which resonant frequencies are determined by positions along the direction of the read gradient. Therefore, with the read gradient, spatial information of MRI signals can be distinguished along the direction of the read gradient. The phase gradient is applied between the first NMR detection pulse and the echo signal. Under the phase gradient, the MRI echo signal accumulates phase differences which magnitude depends on the position along the direction of the gradient and the applied time duration. To extract changing rates of the phase differences, the phase gradient should be applied several times to different echo signals with different magnitude. The extracted changing rates are used to differentiate spatial information along the direction of the phase gradient. This is the reason why MRI imaging time is longer than in other imaging systems. The MRI imaging time is a product of the repetition time, TR and number of acquired MRI signals that is the same as the phase gradient steps. To shorten the imaging time, several multiple echo sequences, in which multiple echo signals are obtained in a single TR, have been developed, such as turbo spin echo, GRASE (gradient echo and spin echo) and EPI (echo planer imaging).

6 How Characteristics of the Main Magnet Affect MRI Image Quality? Basically, in HTS magnet MRI, the magnet for the B0 field is replaced from a low temperature superconducting magnet to a high temperature superconducting magnet. In this section, the effects on an MRI image due to usage of different magnets are discussed. As seen in the previous sections, the B0 field induces magnetization in a body or head. Then, the B0 field is treated as a homogeneous background for the image acquisition process. If spatial inhomogeneity exists in the B0 field, the local magnetic field differences become sum of local tissue properties and the B0 field differences. Thus, obtained tissue properties will be smeared. In addition, an MRI image resolution will be degraded since the linearity of magnetic field gradient becomes inaccurate. The same situation will happen in the case that temporal fluctuations exist in the B0 field. Therefore, the spatial homogeneity and temporal stability are important characteristics of the B0 field magnet. To compensate these effects, the magnetic field gradient coils are specially designed in the HTS MRI. The other aspect of using a different B0 field magnet is a structure difference, especially in metallic parts. The switching of the magnetic field gradients induces eddy currents in metallic parts nearby. The eddy current itself reduces the strength of

General Descriptions on MRI

11

the magnetic field gradient and induces extra magnetic fields in the imaging region. Therefore, the eddy current causes image blurs. Moreover, superconducting currents may be disturbed. This degrades the spatial homogeneity and temporal stability of the B0 field magnet and may make magnet operation unstable. To mitigate the eddy current problems, the magnetic field gradient coils are adjusted to the structure change in the HTS magnet.

Hidenao Fukuyama M.D., Ph.D. is a medical doctor, specialty of Neurology. Graduated Kyoto University in 1975, and got the Ph.D. degree in 1981 from Graduate school of Medicine of Kyoto University, and following serving as the assistant professor, associate professor, of department of Neurology, Kyoto University Hospital, and appointed as the professor of Human Brain Research Center, Kyoto University in 2001 and retired 2015. Since then, working as the program specific professor of Leading program for medico-engineering. Over then 30 years, working as the medical imaging researcher and managing the graduate students, in particular MRI hardware and analyzing technique as well as diagnosis. Since 2016, appointed as the professor of Beijing Institute of Technology, Beijing, China, for giving lectures to the students of engineering with regard to Neuroimage technique and Neuroscience.

Neurons and Plasticity: What Do Glial Cells Have to Do with This? Nicolangelo Iannella and Michel Condemine

1 Introduction Synapses, neurons, and neural circuits have long been associated as the entities where plastic synaptic changes can be observed to take place. These changes rely on the activation of biochemical processes within subcellular compartments of the synapses between neurons as a result of stimulation and associated neural activity. They can occur on either the pre-synaptic or post-synaptic side of the synapse, or both. This association between stimulus inducing neuronal activity and synaptic change has been the subject of many studies that have shown the involvement of both neural activity and calcium [1–3]. Such studies have primarily been driven by the “assumed” microscopic structure of the synapse. In its basic form, this structure views the synapse as being composed of two pieces in close proximity to each other; namely the presynaptic terminal and the post-synaptic spine. This complex is commonly referred to as the synapse. When the synapse is activated by a strong electrical signal called an action potential or spike, this triggers neurotransmitters, such as glutamate, to be released from internal stores within the pre-synaptic terminal and expelled into the space between the pre and post-sides called the synaptic cleft. Released neurotransmitter then quickly drifts, binds, and unbinds with receptors located on the of the post-synaptic spine-head. These binding and unbinding triggers the activation of a set of biochemical signaling cascades that also generates small electrical signals observable in the cell body (soma) of the recipient neuron. When this process of activity and synaptic transmission occurs regularly it can lead to

N. Iannella (B) Department of Biosciences, University of Oslo, Blindern, P.O. Box 1066, 0316 Oslo, Norway e-mail: [email protected] M. Condemine Condemine Consulting, 5 Rue Poliveau, 75005 Paris, France © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_2

13

14

N. Iannella and M. Condemine

observable structural and functional changes in electrical and biochemical responses. This is the outcome of synaptic plasticity. Despite the strong focus on neurons and neural communications, the brain is not simply composed of neurons alone. There are non-neuronal cells in the brain that are known to be present, such as Astrocytes, Oligodendrocytes, and Ependymal cells. These non-neuronal cells are collectively called glial cells and have been estimated to make up about half of the human brain. Tomography studies have shown that the human brain is composed of 86 billion neurons and an equal number of glial cells [4, 5]. Historical experiments on glial cells set out to establish whether these cells produced electrical signals similar to those seen in neurons but observed and established that these cells are essentially electrically inert. This led many to believe that glial cells were not major players in the processing of information in the brain and probably played the role of neuronal maintenance. Consequently, in most neurology and neuroscience textbooks, glial cells are often stated as cells that make sure that both neurons and the synaptic connections between neurons in cortical circuits are maintained in a physiologically healthy state but play no role in the processing of information in the brain [6–8]. Glial cells are typically known as the glue within the brain that keeps neurons in place. To this end, these cells have largely been ignored when investigating the formation and refinement of cellular functional properties. Since the turn of the century, however, evidence has begun to emerge indicating that glial cells may not behave as the brain’s passive elements but play more subtle yet active roles. Here, we review glial cell physiology, followed by a discussion of neural-glial signaling and current efforts in modeling neural-glial communications and potential roles that glia may play in synaptic plasticity and brain function.

2 Glial Cell Types Sometimes called “the other half of the brain”, Glia are non-neuronal cells that are commonly known to be responsible for maintaining the biochemical balance in the brain and keeping neurons in a physiologically healthy state. There are four main types of glial cell in the central nervous system (brain) with another three different types in the peripheral nervous system. The first type of glial cell is called an Astrocyte that is typically found in the central nervous system (CNS) [9–11]. This sub-type has a characteristic star-shaped morphology with branching processes that extend and wrap around the synapses of connecting neurons [11]. Astrocytes can be further divided into two types, fibrous and protoplasmic. Fibrous astrocytes are typically found in the brain’s white matter and have long thin branches while protoplasmic astrocytes are found in grey matter and have highly branched processes that are thicker and shorter than their fibrous counterparts [12–14]. The second type of glial cell found in the CNS is called Oligodendrocytes. These are cells that coat the axons of neurons to form the myelin sheath and provides electrical insulation to axon resulting in faster propagation of action potentials (spikes) from one pre-synaptic neuron to a post-synaptic recipient [15, 16]. The third type is known as the Ependymal

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

15

cell and are primarily found forming a lining in the brain’s ventricular system and the spinal cord’s central canal [17]. They have a unique morphology where their apical surface are covered by two types of protrusions called microvilli and layer of organelles forming eyelash like extensions from the apical surface called cilia [17]. These protrusion allow the cell the absorb and circulate cerebrospinal fluid (CSF) throughout the brain [18, 19]. The final type of glial cell in the brain are called Radial glia, which are derived from neuroepithelial cells during neurogenesis [20]. During nervous system development, radial glia act in dual roles by providing a scaffold for newly generated neurons to migrate into and further act as neuronal progenitor cells [21]. In the developed brain, radial glia can be found in the retina (Müller glia) [22] and the Cerebellum (Bergmann glia) [23], since they have migrated (after losing their branching attachments) to the surface of the cortex where most will transform themselves into astrocytes via the process of gliogenesis [24].

3 Glial Cell Function Glia are known to play many roles in the CNS where some simply provide physical support to neurons while other are actively involved in supporting and providing nutrition to the biochemical environment around neurons and synapses on order to play a role in maintaining brain homeostasis [11]. They are also known to be involved in creating the micro-architectures in the brain, provide the brain with the necessary infrastructure to store and distribute energy to neurons, control the growth and development of neuronal axons and dendrites [25]. Furthermore, they take part in both synaptogenesis and synaptic maintenance and provide the brain, as a whole, with a biochemical defense capability against chemical and immunological perturbations [25]. Historically thought of as just “gap fillers”, Astrocytes are now known to play a multitude of roles [25]. The first of these roles is their involvement in establishing the brain’s micro-architecture where through a process called “tiling”, which leads to the formation of many structural subunits, that are nearly independent of each other, through the subdivision of protoplasmic astrocytes in the grey matter [26]. These cells occupy their own region and create sub-anatomical domains within the spatial extent of their branching processes [12, 27]. Within this region the astrocyte’s membrane covers both synaptic connections and neuronal membranes [25, 28, 29]. Furthermore, some of their branching processing also connect to the walls of nearby blood vessel, thus establishing a structural neuron-astrocyte-blood vessel complex commonly called the neurovascular unit [29, 30]. These individual domains also integrate themselves through gap junctions (electrical synapses) located on peripheral branch locations to form superstructures called astroglial syncytia [31, 32]. This does not result in the formation of one superstructure, but in turn, these syncytia are structurally segregated being formed within defined anatomical structures, like an individual somatosensory cortical barrel [31, 32].

16

N. Iannella and M. Condemine

Due to their strategic location enveloping synapses, astrocytes are well placed to control the biochemical environment around neurons [15, 28]. Astrocytes proactively control the concentration of potassium K+ in the extracellular space surrounding the synapse and neuron, and are therefore responsible for extracellular K+ homeostasis in the brain [28, 33]. Due to astrocyte’s internal molecular machinery, they can also simultaneously control the flow and concentration of water, various ions, neurotransmitters, and metabolites within this neurovascular complex [31, 32]. Experiments have shown that K+ concentration is under the direct control of astrocytes. Under normal physiological conditions, neural activity increases K+ concentration from 3 mM at rest to a maximum of 12 mM, while pathological conditions K+ concentration can surpass higher values [34]. From a biophysical viewpoint using the Goldman-Hodgkin-Katz relation, the increased extracellular K+ concentration can change the reversal potential of potassium channels and modulate the overall resting potential of the cell [35]. Astrocytes, however, directly control the K+ concentration levels since they possess several mechanisms to remove K+ . The first is a passive mechanism where K+ is taken up at location of high concentration through inward rectifying K+ channels and spatially redistributed at site with lower concentration values through the same astrocyte or the coupled astrocytic network [31, 32]. The second is an active mechanism where removal of excess K+ takes place through the Na+ –K+ pump that is powered by adenosine triphosphatase (ATP) activity resulting in an increase in intracellular K+ concentration [32]. An important physiological function performed by astrocytes is the regulation of glutamate in the extracellular space. When activity dependent release of glutamate results in an excess, glutamate, despite being critical for neuronal signaling, then adversely acts as a neurotoxin leading to neuronal cell death. Fortunately, astrocytes possess powerful machinery that can remove up to 80% glutamate from the extracellular space [36]. They achieve this through excitatory amino acid transporters (EAAT); specifically EAAT1 and EAAT2 are exclusively expressed in astrocytes where these glutamate transporter mechanisms utilize the energy stored via the transmembrane Na+ gradient so that the efflux of one K+ ion and the influx of 3 Na+ ions and one H+ ion provides the energy to transport a single glutamate molecule. This in turn leads to an increase in intracellular Na+ but is counterbalanced by efflux through Na+ –Ca2+ exchanger. Note that both of these molecular mechanisms are co-localized in the astrocyte’s membrane [26, 36]. Astrocytes further assist in the homeostasis of glutamatergic transmission via a biochemical process that acts as a glutamateglutamine transportation mechanism. Astrocytes enzymatically convert glutamate into glutamine via astrocytically derived glutamine synthetase. Fortunately, neurotransmitter receptors are not responsive to glutamine, which is not toxic to neurons and can be safely transported through extracellular space to pre-synaptic terminals. Glutamine is then absorbed back into the terminal where, once internalized, it is converted into glutamate to be re-used by the pre-synaptic neuron to signal other neurons [26, 36]. Another import role played by astrocytes is that they play a central role in the neurovascular complex where local blood flow and metabolic support for neural circuits are integrated together. Blood vessels in the brain are covered by the

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

17

branching end-feet processes while the remaining branching processes of the astrocyte target the neuronal membrane, synapse, or axon. In this situation the astrocyte can be viewed as a gateway between the vascular and neural systems [36]. Sensory inputs to neurons increase their activity and consequently triggers an increase in the concentration of intracellular calcium Ca2+ in astrocytes. This Ca2+ response can be interpreted as an integrating signal for the neurovascular complex and ultimately leads to the release of biochemical compounds that regulate blood flow [36]. Moreover, the readout signal that is captured during magnetic resonance tomography originates from astrocyte activity [37, 38]. Furthermore, astrocytes are the only cells in the CNS that act as a battery or energy storage buffer by synthesizing glycogen, thus providing neurons with localized metabolic support where neurons can uptake glycogen, release its internalized glucose, and converts this glucose to lactate which consequently provides neurons with energy [36–38]. Additionally, experiments have shown that astrocytes have a role to play in the development, maintenance, and stability of synapses and to some degree, can control the resulting connectivity pattern of neuronal circuits [25, 28]. Astrocytes achieve this through the production and secretion of biochemical factors that facilitate the formation of synapses (synaptogenesis). Without these secreted factors synapse formation is greatly depressed since synapse formation critically depends on cholesterol production and secretion of specific proteins from astrocytes, potentially providing the raw materials to build new membranes [28]. Astrocytes can also produce factors that affect specific proteins and signaling cascades critical for the formation and development of synapses by regulating the density of post-synaptic glutamatergic α-amino3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) and N-Methyl-D-aspartic acid (NMDA) receptors in spines via the actions of tumor necrosis factor alpha (TNFα) and the activity-dependent release of brain-derived neurotrophic factor (BDNF) produced by astrocytes [28]. The astrocyte’s ability to control the formation and maturation synapses places this cell type a unique position where it can also influence the density of synapses across neurons by ensheathing areas of neuronal membrane, thus spatially restricting where synapses can form [39]. Astrocytes also secrete factors like proteolytic enzymes that react and disassemble the extracellular matrix leading to synaptic pruning where the astrocyte’s membrane may enter and essentially replace the synapses by isolating the post-synaptic spine from the presynaptic terminal [15, 16, 28, 39]. Subsequently, when the brain is in a pathological (diseased) state, the occurrence of this process can be prominent [16, 25]. Due to their strategic loci, astrocytes are also well placed to modulate the flow of sensory electrical signals cortical circuits; in the hippocampus they have been shown to modulate synaptic transmission by suppressing synaptic transmission through the astrocytic release of ATP and its conversion to adenosine (by hydrolysis with ectonucleotidase) can then bind with neuronal adenosine receptors, resulting in the inhibition of synaptic transmission [25, 39]. Furthermore, astrocytes can also receive synaptic inputs, since they have receptors that are activated by neurotransmitters. Studies have recorded spontaneous “miniature” excitatory potentials suggesting that the sites of neurotransmitter release are in close proximity [15]. Finally, astrocytes

18

N. Iannella and M. Condemine

can also trigger the myelinating activity of oligodendrocytes since the activity from neurons leads to the release of ATP and causes astrocytes to secrete a regulatory protein called cytokine leukemia inhibitory factor (LIF) that promotes oligodendrocyte’s myelination activity. From this point of view, astrocytes can be seen to play a coordinated executive role [40]. Oligodendrocytes, on the other hand, are the glial cells responsible for myelination and are found in the white matter of the brain [25, 40]. This glial type also plays several supporting roles such as aiding the production of trophic factors like glial cell derived neurotrophic factor (GDNF) and brain derived neurotrophic factor (BDNF) [41]. These factors are secreted proteins that are involved in the survival of existing neurons, as well as encouraging the differentiation and growth of new neurons and synapses. Furthermore, oligodendrocytes are primarily responsible for the creation of myelination across the neuronal membrane. This leads to a reduction of ion leakage and consequently decreases the capacitance of the neuron’s bilipid membrane [41]. Myelin also leads to a significant increase in transmission speed and changes the nature of signal propagation from a travelling wave to saltatory (jumping) propagation of action potential spikes occurring at the nodes of Ranvier located in between the oligodendrocyte derived myelin [35, 42]. This specialization leads to transmission speeds to be linearly related to the diameter of the axon, whereas transmission speeds in unmyelinated axons is proportional to the square root of the diameter [35, 42]. Conversely, satellite oligodendrocytes are found in the grey matter regions of the brain and regulate the extracellular environment where they remain close to neurons but do not create myelin sheaths on neuronal membranes [43]. The third type of glial cell present in the CNS are ependymal cells whose sole purpose is to maintain the homeostasis of Cerebrospinal fluid (CSF) throughout the brain. Located in the ventricles and the central canal of the spinal cord, their apical surfaces form a choroid system that regulates CSF so that a relatively constant amount is maintained. Their basal membranes are characterized by branch like extensions that attach themselves to astrocytes thereby enabling an glial cell network to be established and provides a substrate by which glial-glial signaling could, in principle, provide an alternate pathway for neural signals to propagate [44]. The final glial cell type is the Radial Glial cell. The morphology of these cells is best described as a dipole or bipolar with branches emanating from opposite ends of the cell body. This particular type of glial cell are known to span wide areas of cortex during the development period [11, 45, 46]. These cells are interesting as they act as progenitor cells and go on the produce neurons, astrocytes and oligodendrocytes [45–48]. Radial cells have the ability to divide asymmetrically into a new glial cell and either a new neuron or an intermediate progenitor daughter cell (IPC) [46–48]. Within the subventricular zone, IPCs can go on to further divide symmetrically in subventricular zone to generate new neurons [48, 49]. Local bio—chemical cues and their underlying macromolecules have been shown to control how the radial glial cell subdivides and the type of daughter cells that are made [50]. Biochemical signaling by macromolecules such as Notch and Fibroblast Growth Factor (FGF) have been shown to play roles in these processes, in particular they regulated the rate of neurogenesis and consequently the production of both new neurons and radial

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

19

glia, which in turn affects the growth, surface area and volumetric expansion of the cerebral cortex, including the emergence of folds and ridges in the cerebral cortex [48, 50, 51, 52]. Being confined within the skull, the gyri belonging to this system of folds and ridges, provides a means to increase the surface area of the brain without increasing the overall volume, thus limiting brain size [51, 52]. Another function that radial glial cells possess occurs near the end of cortical development, where radial glia detach themselves from the ventricles and migrate toward the cortical surface to undergo gliogenesis, where most will transform themselves into astrocytes [49, 51, 52]. In vitro studies suggest that radial glia can also transform into oligodendrocytes via generating oligodendrocyte progenitor cells. This suggests that oligodendrocytes are derivatives of radial glia however whether this process actually occurs in the developing mammalian brain requires more evidence [50, 53].

4 Signaling in Glial Cells Like neurons, astrocytes and other glial cells possess second messenger metabotropic receptors that are (intrinsically or extrinsically) coupled to specific intracellular signaling cascades since they are complex biophysical entities that employ many different proteins and macromolecules [10, 27, 54]. These complex biochemical interactions give rise to cellular mechanisms that essentially allows the glial cell to respond to some external or sensory input. Glial cell responses result from exciting the Endoplasmic Reticulum (ER), leading to an increase in intracellular Ca2+ through the activation of IP3 and Ryanodine receptors [10, 54]. Like the ER of the neuron, stimulating metabotropic receptor in glia leads to the production of IP3, resulting in the internalized release of Ca2+ from the ER into the cytosol of the glial cell, producing a measurable calcium signal and hence can be considered as a substrate for glial excitability [27]. These Ca2+ signals are not restricted to a single glial cell but can propagate through a network of glial cells, connected by connexon channels or gap junctions, called a glial syncytium [11]. These gap junctions aids communication between cells to form a Ca2+ wave that travels through the astrocytic network, thus potentially providing an alternative communication pathway for both neurons and glial cells alike [11, 27, 54, 55]. The properties of gap junctions between neurons has been well studied and provides networks with both a fast means of neural communication and several important functional properties such as synchronization [42]. By extrapolating these finding, the presence of gap junctions between glia can potentially provide a potent mix of novel functional attributes and fast cellular communications within glial networks [54, 56]. This form of communication can be viewed as a parallel communications channel; however, the precise physiological and functional properties are still to be determined.

20

N. Iannella and M. Condemine

5 Signal Transmission from Glial Cells Signal transmission in the brain is typically associated with voltage dependent release of neurotransmitter from the axon terminals of pre-synaptic neurons that can quickly diffuse and bind with AMPA and NMDA receptors located in the PSD of the postsynaptic spine [42]. Experiments have now established that astrocytes and other glial cell types can also release the same types of transmitters as neurons, such as glutamate and γ -aminobutyric acid (GABA), into the extracellular space [11, 28, 56]. Consequently, with the knowledge that glial cells are activated via glutamate from neurons and the expectation that neurotransmitters released from glia into the extracellular space of the synaptic cleft can depolarize a post-synaptic neuron, provides an additional line or channel of communication between neurons. In fact, this leads naturally to the intriguing concept of gliotransmission as an additional line of communication between neurons [9, 36]. Most of the neurotransmitters released from glial cell are no different to those released by neurons, such as glutamate, GABA, and ATP, however there are several that are uniquely from glial cell origin, namely taurine [36] and kynurenic acid [1]. Mechanisms underlying neurotransmitter release from neuron have been extensively studied, which leads one ask what types of mechanisms can lead to the exocytosis or release of transmitters from glial cells? To date there have been several different confirmed mechanisms that give rise to gliotransmitter release [36]. The first of these mechanisms is gliotransmitter release via transporters, namely through the exchange and reuptake involving the cystine-glutamate antiporter (organic anion transporters); the second is via diffusion processes namely through high permeable channels such as Cl− channels or P2X7 purinoceptors; and finally via the more familiar mechanism of Ca2+ -dependent exocytosis [26, 56, 57]. Recent experiments have confirmed that astrocytes express an important family of proteins including synaptobrevin 2, syntaxin 1, and 23 kDa, which are known to be involved in exocytotic release [7, 26, 36, 58, 59]. Furthermore, the presence of V-type proton ATPase (V-ATPase) is known to create a proton concentration gradient that drives ATP-based transport of glutamate into internalized vesicles, and consequently glutamate is ready to be released by ongoing activity at some later point in time [57, 58, 60]. The presence of the vesicular glutamate transporters (VGLUT1, VGLUT2 and VGLUT3) has also been confirmed through immunoelectron microscopy [36]. Functionally, glutamate release from glial cells, such as astrocytes, result in detectable responses in neurons that includes the generation of an inward mediated NMDA current leading to the elevation of Ca2+ concentration in the recipient neuron and furthermore, glial-originating glutamate can alter neuronal excitability through the modulation of synaptic transmission leading to the synchronization of neural activity [57, 61]. One must keep in mind that on the subject of information processing in the brain, the role and relevance of glial cells is still controversial and is clearly the next big frontier for neuroscience that requires further investigation, in particular how direct and indirect communication between neurons via glial cells alters the function or neuronal circuits [62].

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

21

6 A Synaptic Ménage à Trois: The Tripartite Synapse Concept Given that glial cells, including astrocytes, typically wrap themselves around both the pre-and post-synaptic terminals, are known to release glutamate and respond to glutamatergic inputs. This sets up a unique biological structure where cell-to-cell communication can potentially be both functionally richer and allow neural communication to proceed on different timescales simultaneously [9, 63, 64]. The close proximity of the glia to both pre-and post-synaptic membranes has been observed for much of the brain, including both gray and white matter [28]. For example, in hippocampus 57% of axonal-to-dendritic synapses are covered by glial membranes [28] (in this case the membrane of astrocytes) where there has been a clearly noticed tendency that between 60–100% of large synapses to be covered by such membranes while for small (and probably less functionally active) synapses this percentage range drops to approximately 40–60% [65]. In the olfactory bulb of the rat the number of synapses that are covered by glia membrane is over 60% [34]. Interestingly, in the Glomerular layer, single excitatory synapses that were covered by glial membrane stood at 27%, while that ratio more than doubled (to 72%) for reciprocal synapses [34]. Similar numbers of glial covered reciprocal synapses in the external piriform layer also stood 76% [34]. Furthermore, such glial-synaptic contacts can also illustrate unique patterning as observed in the cerebellum where virtually all the synaptic contacts onto Purkinje cell dendrites that originate from parallel fibers are surrounded by the membranes of Bergmann glial cells [27]. It has been found that an individual Bergmann glial cell can contact and enwrap 2000–6000 parallel fiber-Purkinje cell synapses [27]. The physical organization of tripartite synapses is such that the individual membranes contributed by both glial cell and neurons are in close proximity to each other [28]. Their close apposition allows the glial cell membrane to be exposed to the same set of neurotransmitters that post-synaptic neuronal membranes are subject to [26, 28, 36, 66]. In this context the glial cell could be viewed as eavesdropping on the messages transmitted between neurons. From the physiological point of view, the membranes of glial cells, e.g. astrocytes, possess a family of receptors similar to their neuronal neighbor that respond to various neurotransmitters and can elicit an calcium based response to presynaptic activity [11, 66]. This similarity in the types of receptors expressed in both glial cells and their neuronal neighbors in a given brain region is observed throughout the brain including the basal ganglia, cerebral cortex, and the cerebellum [28]. Despite this similarity within a given region of the brain, these region-based set of receptors are not homogeneous or invariant across the entire brain, but change across different brain regions [57]. This raises an interesting question as to why these closely positioned membranes share near identical sets of neurotransmitter sensitive receptors and what are the functional consequences of such features. The addition of a glial membrane into the traditional picture of synapse being a dual membrane complex composed of a pre-synaptic terminal ending and the post-synaptic membrane (spine head) naturally leads to the notion of the tripartite

22

N. Iannella and M. Condemine

synapse, forming a tri-membrane complex that contains the three equally important aforementioned constituents. This construct permits synapses to literally communicate to its receiving membranes in a parallel manner where neurotransmitter released from the terminal activates the appropriates receptors in both the glial and postsynaptic neuronal membranes. This leads to the generation of an electrical signal in the post-synaptic neuron and a (slower) calcium signal in the glial cell [28, 57]. Ultimately this calcium signal can lead to the release of neurotransmitter which will modulate the activity of (pre-synaptic and post-synaptic) neuronal membranes [9, 26, 36, 57]. The significant question the neuroscience community is beginning to ask is whether glia (astrocytes included) play critical active roles in information processing. Experiments have established that signaling in glia operates on a slower timescale, typically of the order of seconds to minutes, than fast electrical signaling in (and between) neurons [11]. The consequences of this dichotomy of signaling on information processing and the dynamics and function of neural circuits in the brain is currently not known, but some attempts have begun to investigate these issues both experimentally and computationally [28, 57]. From the experimental side, we know that glia are involved in synaptic plasticity however it is not clear whether they play a direct or more of a modulatory role in forming functional neural circuits [28, 65, 67]. Furthermore, their involvement in synaptic transmission has now started to cast questions with regards to the role of glia in brain function specifically the operation of functional networks, despite originally being thought of playing no role in neurotransmission [7, 67]. Since various receptors in glia can be directly activated by neurotransmitter released into the extracellular space of the synaptic cleft in response to some stimuli, coupled with their ability to release neurotransmitters, thus providing an additional signal to the post-synaptic membrane, raises many questions of their role and functional consequences of reciprocal signaling between neurons and glia at the level of neural populations [36, 60, 63]. This issue has largely been unexplored. From the computational side, glia and their slower calcium-based response could play one of two different roles. The first is that the glial cell simply behaves as a temporal integrator (a low pass filter in time) where the resulting calcium signal is simply integrating its own inputs and feeds this signal back into the neuron [64, 68]. The second role is that the glial cell may behave more as a modulator, by keeping the synapse (and consequently the neuron) in a preferred operational range of activity [61]. Another property that they provides is spatial specificity of synaptic transmission in the sense that glial membranes can physically minimize or stop synaptic spillover of neurotransmitter onto neighboring synapses on dendrites, thus increasing the degree of spatial precision or conversely isolate a synapse from receiving interference from the extracellular environment [34, 65, 69]. Despite these roles at the synaptic level, the consequence of glial cell actions on neurons at the level of functional neural populations remains largely unexplored, however some theoretical studies have begun to investigate the actions and consequences of neural-glial signaling on neuron and network responses [70–73].

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

23

7 Investigating the Impact of Neural-Glial Signaling on Network Dynamics Discovering the currently unknown roles played by glia in the operation of functional neuronal populations will benefit from the synergy of experimental and theoretical techniques, typically known as data-driven science [74]. In such a situation where only a few facts are known, theoretical and computational techniques can be used to provide predictions that can be tested experimentally [74]. These methodologies will play important roles in clarifying the roles played by glia in brain. To this end, there have been some modeling efforts to help understand how glia, and in particular astrocytes, influence the dynamics of networks of spiking neurons [72, 73, 75, 76, 77, 78, 79, 80, 81, 82]. Such models are invariably more complex than typical studies of network dynamics of spiking neurons as the communication between neuron and glia requires a minimum of three variables, namely neuronal depolarization (voltage), release of neurotransmitters, and calcium. Note that in order to generate meaningful predictions requires a description of how glial cells respond to neuronal input and visa-versa. Here, the efforts from two independent groups of Sosnovtseva [72, 73, 75, 76] and Jung [83–85], have made it possible to begin investigating neural-glial signaling in neural populations computationally. These studies have inspired others to investigate the potential roles that glial cells could play in information processing, synaptic transmission and synaptic plasticity [78–82]. We now present a relatively simple modeling framework that can be used as a starting point for studying the influence of glial cells on the neuronal populations. This model is based upon the known biophysics but is more complex due to the need to capture the correct nature of communication between glia and neurons. We start by presenting a standard biophysically inspired model of the neuron, then the dynamics of the glial cell, and formulate their reciprocal interaction. There are a number of neuron models that one can choose from ranging the Izhikevich model [86] through to detailed biophysical models that contain a variety of ion channels including calcium channels and their own calcium based subsystem. As a starting point, one can model a network of neurons using the Izhikevich model [86] described by dν = 0.04v 2 + 5v + 140 − u + Iinput + Iglial dt du = a(bv − u) dt if v > vthres then v → vreset u →u+d

24

N. Iannella and M. Condemine

where a = 0.02, b = 0.2, d = 8, vthres = 30, vreset = −65, I input represents the synaptic current generated by post-synaptic receptors, such as AMPA or Gammaaminobutyric acid (GABAA ) receptors, that are triggered by inputs originating from other neurons in the neural population, and I glial is the current generated from glia cell inputs. Both AMPA and GABAA are functionally described by Iinput (t) =

g (t) = i

 j

 exp





  Wi g i (t) EAMPA,GABAA − v

i

−t − tji τd



 − exp

−t − tji



τo

H t − tji ,

where gi (t) is the total conductance to the synapse, t j i is the arrival time of the jth incoming spike from the ith presynaptic neuron to the post-synaptic neuron, W i the corresponding synaptic weight, EAMPA, GABAA represents the reversal potential for AMPA and GABAA , H(·) is the Heaviside step function, τ o and τ d are the onset and decay time constants, respectively. Although the corresponding synaptic conductances for these inputs sum linearly, there are other synaptic currents where their underlying conductances do not, thus summing in a nonlinear manner. These types of currents are the NMDA, which is dependent on the post-synaptic voltage in a nonlinear fashion and the G-protein activated metabotropic gamma-aminobutyric acid GABAB can also be used in simulations. In order to be computationally efficient, the glial cell calcium signal response was adopted from [87], and behaves as a simple integrator of its inputs, given by the following set of equations [61, 88]:     d Ca2+ = −ϕ + σj δ t − tj dt j    dϕ = α β Ca2+ − ϕ dt where ϕ is a recovery variable and σ j acts as a weight or efficacy of the neuronal spike input to the glial cell. This captures the glial cell’s ability to respond to the release of glutamate from pre-synaptic neurons, δ() represent the arrival time of the spike, and α = 0.001 and β = 0.01 are constants so that they reproduce the slow time course of calcium change observed in experiments by Kawabata et al. [89]. Glutamate released from glial cells is triggered by the elevation of calcium that is described by a simple first order process where glial cell glutamate release occurs when calcium concentration exceeds a threshold as described below [61, 87], μ

       d glu − glu + Ca2+ − Ca2+ thres − κλ if Ca2+ > Ca2+ thres  = − glu − κλ otherwise dt

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

η

25

 dλ = −λ + glu dt

where [Ca2+ ]thres = 0.0018 mM, κ = 200 is the coupling time constant between glutamate and the recovery variable λ, μ and η are time constants with values 0.5 and 10 s, respectively. The glial current I glial generated in the neuron is related to the calcium concentration of the glial cell, which is given by the following experimentally determined relationship first derived and quantified by Nadkarni and Jung [83, 85]      Iglial = 2.11θ log Ca2+ − 196.69 log Ca2+ − 196.69 , where θ(·) is the Heaviside step function. These aforementioned equations provide a computationally efficient starting point to investigate how glial cells may influence the response properties of neurons and neural populations. As a toy example, one can design a simple feed forward network based upon a regular spiking Izhikevich neuron model and the presented simplified model of an astrocyte. The output (Izhikevich) neuron was stimulated by a layer of 100 excitatory and 20 inhibitory neurons that fired stochastically. Presynaptic inhibitory neurons fired with an average rate of 10 Hz via homogeneous Poisson processes, while excitatory cells stochastic firing activity was encoded a set distinct input patterns were used whose representation was such that they spanned an abstract feature space. These stimuli were used to generate spike activity in the excitatory neurons. The set of patterns are individually described by a Gaussian profile of finite width σ, was constructed to cover an abstract feature space were given by:   (x − xc )2 , G(x) = Amin + (Amax − Amin ) exp − 2σ 2 where x represents a location of the feature space, xc is the centre of the input profile, and σ = 10, and Amin and Amax are the minimal and maximal amplitudes, respectively. The glial cell was also stimulated the presynaptic excitatory neurons, but it received its strongest input from the output neuron. The glial cell, in turn, also provided input to the output neuron, forming a recurrently connected pair. The presynaptic glutamatergic inputs to the glial cell provided a modulatory drive. Typically, studying the behaviour of spiking neural networks normally excludes the contributions of glial cells to neural firing, despite mounting evidence that different classes of glial cells influence the spike responses of neuron. The additional complexity introduced by glial cell influences will also impact the response properties of neural populations and thus, our current understanding of brain-based information processing. The impact of different classes of glial cells on network properties is largely unknown but it is clear that the addition of an additional cell classes, whose dynamics and interactions with neurons functions on different timescales, are well placed to influence network properties on multiple timescales. Given that glial cell numbers in the mammalian

26

N. Iannella and M. Condemine

brain equals the numbers of neurons, there is a clear imperative that the majority of published network-based studies need to be revisited and re-examined, especially in terms of network dynamics. From the engineering viewpoint, the inclusion of glial cells to neural network architectures provides additional pathways and processes that influence neural networks on different spatial and temporal scales. These influences can be localised between a small number of nearby neurons on slower timescales or widespread between various neural populations. What is clear is that one can investigate not only glial cell influences on neural dynamics but also glial network influences on neural populations at various timescales. Put simply, the study of multiscale network-to-network dynamics. Such investigations are clearly in their infancy but provide an opportunity to impact other research areas including machine learning and AI. Below, some examples are presented where the inclusion of astrocytes can influence the spiking patterns of neurons. Figure 1 represents the case with no astrocyte and presents the membrane potential of the output neuron. Figure 2 corresponds to the case when a glial cell is present and is recurrently connected to the output neuron, while receiving modulatory glutamatergic inputs from presynaptic neurons. In this configuration, the glial cell can strongly influence the firing pattern of the output neuron. Figure 2 presents the corresponding membrane voltage of the output neuron, the glial cell calcium profile and the current generated by the glial cell in the output neuron. Comparing the two cases, glia are capable of shaping and modulating the spike trains of neurons and consequently influence the response properties of neural populations. The inclusion of glial cells allows novel network derived influences of network dynamics. Figure 3d–e presents how the various variables [Ca]astro , λ, ϕ, [glu] and Iastro vary and how they contribute towards the model glial cell’s influence. Neuronal firing in the absence of glial cell input Membrane Potential (mV)

50

0

-50

-100

0

1

2

3

msec

4

5 4

10

Fig. 1 Simulated membrane potential trace of the model neuron stochastically driven by populations of excitatory and inhibitory neurons in the absence of glial cell input

Neurons and Plasticity: What Do Glial Cells Have to Do with This? Neuronal firing in the presence of glial cell input

50

Membrane Potential (mV)

27

0

-50

-100 0

1

2

3

msec

4

5 4

10

Fig. 2 Simulated membrane potential trace of the model neuron stochastically driven by populations of excitatory and inhibitory neurons in the presence of glial cell input

8 Glial Cells and Synaptic Plasticity Having experimentally confirmed that at tripartite synapses, signaling between the neuron and the glial cell is reciprocal in nature where the glial cell can respond to neuronal inputs while concomitantly modulating synaptic transmission, through the action of active transporters (complex transport systems) of neurotransmitter molecules and by releasing molecules known to alter synaptic transmission such as ATP and D-serine [57]. It is the glial cell’s ability to release molecules that play a role in altering synaptic transmission and/or the connection itself has prompted questions about their functional role during synaptic plasticity. One such molecule that is synthesized and released by glial astrocytes is D-serine, a macromolecule known to serve as a co-agonist of NMDA receptors, a post-synaptic receptor ubiquitously for its involvement in synaptic plasticity [57]. A recent experiment has clarified how D-serine impacts NMDA receptor induced long term potentiation (LTP) formed by the Schaffer collateral synapses in the CA1 region of the hippocampus [90]. Using Ca2+ chelators to block increases in the intracellular Ca2+ concentration of glial astrocytes abolishes the induction of LTP and leads to decreased NMDA receptor currents at nearby synapses. Henneberger et al. (2010) illustrated that under these conditions, the addition of D-serine would rescue LTP induction [90]. Furthermore, by blocking D-serine synthesis in astrocytes through the inhibition of its enzymatic serine racemase in conjunction with high frequency stimulation that leads to the depletion of available D-serine also suppresses LTP [90]. D-serine is not the only substance to influence synaptic plasticity. ATP and its hydrolyzed version adenosine have also been shown to affect synaptic plasticity outcomes. Two photon Ca2+ imaging studies of hypothalamic neurons have revealed that glia-derived ATP leads to synaptic scaling of glutamate based currents in a multiplicative fashion by activating purinergic P2X receptors [91]. The underlying process is dependent on group 1 mGluR1 activation in astrocytes, which leads to an

28

N. Iannella and M. Condemine Glial induced neural current I

Astro

Glial cell current (

A)

20 15 10 5 0 1

0

2

3

4

msec

a) 10

-6

5 4

10

Astrocyte

(mM)

2 1.5 1 0.5 0 1

0

2

b) 10

Calcium Concentration (mM)

2.5

4

-3

5 4

10

Astrocyte Ca

2 1.5 1 0.5 0 0

c)

3

msec

1

2

3 msec

4

5 4

10

Fig. 3 Glial specific variables that underlies the glial cell’s input current to the model output neuron. Note how the largest deviations in each variable occurs in synchrony

Neurons and Plasticity: What Do Glial Cells Have to Do with This? 10

[glu] (mM)

1

29

Astrocyte [glu]

-4

0.5

0 1

0

2

3

4

msec

d) 10

-5

5 4

10

Astrocyte

1 0.8 0.6 0.4 0.2 0

e)

0

1

2

3

msec

4

5 4

10

Fig. 3 (continued)

increase in intracellular Ca2+ concentration from internal stores via IP3 production and leads to the release of ATP and adenosine into the synaptic cleft [57, 91]. For Schaffer collaterals and their synapses in CA1, the accumulated adenosine can act on adenosine 1 receptors (A1 R), which suppresses glutamate release from the presynaptic terminal resulting in an increase in dynamic range for LTP induction [57]. Both ATP and adenosine are capable of diffusing to other synapses thus providing a crosstalk mechanism by depressing glutamate release and synaptic transmission [92]. Another important protein is glial derived cytokine tumor-necrosis factor-α (TNF-α) which has been shown to increase the expression in AMPA receptors in hippocampal cultures and suggests that TNF-α may have a role to play in NMDA receptor mediated LTP and LTD [93]. In the hippocampus, this protein is known to directly contribute to the process of synaptic scaling, which allows a neuron to globally adjust the synaptic efficacy of all its excitatory and inhibitory synapses in response to alterations of prolonged activity [93, 94]. Experiments have shown that the mechanism behind these adjustments is the result of increasing the surface expression of AMPA receptors in excitatory synapses, while simultaneously decreasing the efficacy of inhibitory synapses by decreasing GABAA receptor expression [93–95].

30

N. Iannella and M. Condemine

Moreover, it has been shown that bidirectional control of synaptic strength is regulated by at least two opposing signals, with TNF-α increasing the level of excitation while reducing inhibition and other signals like brain-derived neurotrophic factor (BDNF) decreasing the amount of excitation while increasing inhibition [93]. Glia, through the poorly understood actions of glial-derived GABA and glycine, have been shown play a role in shaping synaptic transmission and plasticity at inhibitory synapses. The GABA transporters GAT1, expressed in both neurons and glia, and GAT3 (only expressed in glial cells) have illustrated that they play an important role in glial GABA uptake in vivo thus controlling the extracellular levels of GABA in the synapse [96]. Work with GAT3 knockout mice has illustrated that this knockout are more resistant to induced seizures since they require higher doses of GABAA antagonist, when compared to normal wild type mice, before experiencing a seizure [97]. Furthermore, modifying GABA uptake by genetic removal of the glycine transporter GlyT1 (mainly expressed in brain stem and spinal cord glia) led to severe motor and respiratory defects [98, 99]. Together these observations highlight that glial cells play critical roles in synaptic transmission and plasticity. Glutamate uptake in glial cells is also involved in the modulation of synaptic plasticity through the activity of two (high affinity) excitatory amino acid transporter subtypes (GLAST/EAAT1) and (GLT1/EAAT2) found on the glial membrane close to excitatory synapses [57]. These glial transporters prevent both over accumulation of glutamate and over-stimulation of glutamate receptors thus preventing toxic levels of extracellular glutamate to accumulate, hence avoiding excitotoxic cell death of neurons to occur [100, 101]. These transporters are regulated by EphA, a member of the Eph receptor tyrosine kinase family and their associated ephrin ligands [102]. This signaling system is bidirectional in nature and hence can reverse the roles of Eph and ephrin, since a Eph receptor can act as a ligand and similarly a ephrin can act as a receptor [102]. Reports have illustrated that Eph receptor signaling plays a role in development of dendritic spines, where Ephrin-B ligands activate dendritic EphB receptors led to spine morphogenesis and maturation while the activation of EphA receptors are important for regulating spine morphology of pyramidal cells in adult hippocampus [101]. Experiments have also shown that activation of EphA4 reduces both the density and length of dendritic spines whereas loss of EphA4 activity has the opposite effect [101]. A recent study has illustrated that the ephrin-A3 ligand found in astrocytes is critical in maintaining EphA4 activation and normal spine morphology in vivo [100–102]. Notably, LTP at CA3-CA1 synapses is modulated by EphA4 in the postsynaptic CA1 neuron and its corresponding ligand, ephrin-A3, found in astrocytes [100]. Additionally, increased GLT1 expression in astrocytes and regulation of GLT1 transcription at mossy fiber (MF)-CA3 synapses led to marked deficits in mGluR-dependent LTD, suggesting that GLT1 participated in preventing activation of pre-synaptic receptor [103]. Taken together these findings suggest that neuronal EphA4 and glial ephrin-A3 signaling can control synapse morphology and the expression of glial glutamate transporters, thus providing an important mechanism that allows astrocytes to bidirectionally regulate both neuronal function and synaptic plasticity [57, 101, 104].

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

31

9 Glial Cells and Structural Plasticity During early development, synapses are excessively established but are later eliminated in an activity dependent manner to form functional circuits. Synapses are dynamic objects that undergo structural remodeling via constant formation and elimination of dendritic spines [105, 106]. Despite implicating neural activity as the precursor to change, the underlying cellular and molecular processes that lead to such structural changes in the brain is poorly understood. Recent studies have shown that glial cells can play an important role in eliminating synapses including neural debris, such as clearing away fragmented axons and lipid membranes left over from synaptic elimination [107]. Significantly, in vitro studies using purified retinal ganglion cells isolated rodent retina have shown that specific biochemical signals derived from glia play a role in synapse formation (synaptogenesis) [107]. Barres [108] showed that few synapses are formed when retinal ganglion cells are cultivated in serum-free media, whereas in the presence of astrocytes or a astrocyte conditioned medium, their numbers dramatically increased by tenfold and five to sevenfold for these respective preparations, where newly formed synapses are free from defects and possess functional AMPA receptors. Using this preparation, previous studies have uncovered that astrocytes secrete three different biochemical factors. The first factor induces the formation of silent synapses (normal synapse which lack AMPA receptors on the postsynaptic side of the synapse). The second, those that increase pre-synaptic activity and increases neurotransmitter release probability. Finally, those that lead to the insertion of glutamate receptors (such as AMPA) and thus an increase in synaptic weight [109]. Using conditioned cultures, thrombospondins (TSPs) have been identified as one of these secreted proteins and belongs to the family of large extracellular matrix proteins, which gives rise to an increase in the number of silent synapses [109]. Removing TSPs from astrocyte conditioned medium reduces synaptogenic activity leading to fewer synapses being formed. This illustrates that TSPs are both necessary and sufficient for the formation of new synapses in vitro. At an early postnatal age in vivo, TSP1 and TSP2 are expressed in developing astrocytes at the time of synaptogenesis, but their expression is downregulated in adults [28]. Significantly, experiments have shown that mice which lack these proteins had fewer excitatory synapses, highlighting that TSPs are important signaling proteins for in-vivo excitatory synapse formation and structural plasticity. TSP binds to the α 2 δ−1 subunit of voltage dependent calcium channels (its corresponding neuronal receptor) [107]. Eroglu and Barres [110] have illustrated that, for all five types of TSP in mammalian brain, synapse formation can be induced when TSP binds to the type 2 epidermal growth-factor-like repeats of the von Willebrand Factor type A (vWFA) domain in the α 2 δ−1 subunit [110]. These authors further postulated that the interaction between TSP and the α 2 δ−1 subunit activates a synaptogenic signaling complex, that may involve calcium channels, and gives rise to new synapses to be formed. Furthermore, TSPs have been shown to bind with another family of macromolecules important for regulating both the size and function of synapses, called Integrins [110, 111]. Recent

32

N. Iannella and M. Condemine

studies have further implicated TSP1 as a ligand for the synaptic adhesion molecule, neuroligin [111] and highlights that TSPs may have the ability to form and modulate synaptic function, both pre-and post-synaptically. In an additional twist, in vitro experiments have also shown that the neurontin drug gabapentin can bind to the α 2 δ 1 subunit and block the formation of TSP-induced excitatory synapses without affecting established contacts [110]. Furthermore, gabapentin can also disrupt excitatory synapse formation between neurons throughout brain by blocking the ability of TSP to bind with the α 2 δ 1 subunit, which triggers synapse formation [110] and provides additional evidence that glial cells, such as astrocytes, can effectively promote synapse formation in the in vivo brain. As TSP secreted from astrocytes only instructs the synapse formation, there must naturally be other secreted macromolecular signals that facilitate glutamate receptor insertion into the PSD of the post-synaptic spine membrane, hence allowing astrocytes to control synaptic efficacy and function. The precise factors responsible for AMPA receptor insertion into the post-synaptic membrane in the in vivo mammalian brain are not well known, however in vitro studies have shown that adding apolipoprotein E bound to cholesterol to cultured RGC led to two observable changes. These cholesterol-induced changes were firstly, a 69% increase in the number of excitatory synapses and finally a 12-fold increase in (excitatory) presynaptic transmitter vesicle release [112, 113]. Both cholesterol and apolipoprotein E increases synaptic responses in autaptic synapses in cultured RGC by increases in pre-synaptic function and dendritic differentiation [112, 113]. This suggests as potential role for cholesterol during brain development of the mammalian brain in vivo.

10 Glia and Brain Dysfunction Brain dysfunction, such as stress related disorders, dementia, or Alzheimer’s, and the development of effective treatments has been identified as an international priority area that, if left untouched, will cost the world economy trillions of dollars in care and treatment. Understanding the root causes and development of brain dysfunction or disease will be critical in avoiding or developing effective long-term treatments for such medical conditions. At the core of understanding the pathology of brain disease, lies the fact that brain activity will not be the same when comparing normal to pathological conditions. Thus, placing the glial cell in a unique position since, as stated earlier, they can influence network activity as well as alter the excitability of single neurons, making them targets for developing new treatments. Such gliocentric approaches to brain dysfunction has the potential to transform our understanding of the development and treatment of these disorders. Current literature has illustrated that glia, including astrocytes, are important for the development, maintenance, and refinement of neural circuits [16, 25, 28, 57]. Hence it stands to reason that any impact on their pathology is likely to have a dramatic effect on normal brain function, potentially leading to some disorder in the brain. For this reason glia, in particular, astrocytes are becoming the focus for

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

33

understanding brain dysfunction such as autism, anxiety, and depression, as these seem to be closely linked to the physiological state of glia [25]. Recent research has started to investigate how astrocyte dysfunction contributes to the emergence of synapse-based disorders like autism and schizophrenia. Significantly, developmental diseases like Down’s syndrome and Fragile X syndrome have illustrated that astrocyte dysfunction contributes to such disorders.

11 Down’s Syndrome In the case of Down’s syndrome, cognitive deficits have been linked to altered number of spines, specifically experiments have demonstrated that Hippocampal cells grown in the presence of astrocytes from Down’s syndrome patients resulted in reductions in spine activity and density [114]. Garcia et al. [114] has identified that the level of TSP1, an astrocyte secreted factor known to modulate spine numbers, was markedly lower in astrocytes from Down’s syndrome patients. The same study further showed that restoration of TSP1 levels rescued both spine number, function, and densities. Thus, demonstrating that astrocyte secreted factors contribute to synaptic defects and suggests that TSP1 may be used as a specific treatment.

12 Fragile-X Syndrome Another disease that is associated with cognitive deficits is fragile-X syndrome. Similar to Down’s syndrome, fragile-X is caused by the mutation of the Fragile X mental retardation 1 (FMPR1) gene which encodes the RNA-binding macromolecule fragile-X mental retardation protein (FMRP) and is only found in synaptic apparatus where ribosomes transcribe mRNA into FMRP [39, 115]. Polyglutamine repeats in the FMR1 gene leads to a loss of FMRP expression in synapses. This has been shown in FMR1 knockout mice whose phenotype captures many of the physiological and clinical traits displayed by Fragile [115, 116]. Furthermore, Hippocampal neurons grown in the presence of FMR1 deficient astrocytes show reduced spine density and abnormal dendritic morphology, however such effects are significantly reversed when neurons from FMR1 deficient mice are growth on a monolayer of normal wildtype astrocytes [115]. These experiments indicate that astrocyte dysfunction plays a role leading to abnormal synaptic and dendritic development, and further highlights the possibility that in vivo defects in spine development are related to abnormal neuralglial signaling.

34

N. Iannella and M. Condemine

13 Rett Syndrome Rett syndrome is another developmental disorder that is known to be caused by the loss of transcriptional repressor methyl-CpG-binding protein 2 (MECP2) [117]. Experiments with mice that lacked the MECP2 gene illustrated many of the same features of the human variant of the disease including cognitive impairment, respiratory abnormalities, and a decrease in brain mass and volume. This suggests that there is both an important role (along with a critical window) for MECP2 function during postnatal development [117, 118]. A recent study which cultured normal or wildtype hippocampal neurons with astrocytes from MECP2 deficient astrocytes lead to dendrites with stunted dendritic morphologies, suggesting that astrocytic deficits in MECP2 impair normal dendritic growth [117, 118]. This is further supported by the in vivo observation that reactivation of MECP2 in astrocytes can partially rescue the development of neuronal dendrites in MECP2 deficient mice [116]. These findings raise an interesting question of whether MECP2 controls gene transcription of astrocyte secreted proteins involved in normal brain development and thus MECP2 loss could impair the production and secretion of synaptogenic proteins from astrocytes and lead to structural deficits in dendritic morphologies akin to those observed in Rett syndrome.

14 Epilepsy Epilepsy is another common neurological disease that affects millions of people worldwide with 80% of cases occurring in developing countries where the rates of new cases being detected are up to 140 per 100,000 people [119]. This brain disorder is typically characterized by the onset of involuntary seizures that vary from brief and near undetectable episodes to long periods of shaking. The precise cause of epilepsy is not known but there are those that have developed the condition through brain injury, stroke, brain tumors, brain infections, and birth defects. In most cases, epileptic seizures result from excessive neural activity in the cerebral cortex [120– 123]. In about 70% of cases epilepsy can be controlled through medication or other relatively inexpensive means. For the remaining 30% of cases that do not respond to medication then other options include dietary changes and surgery are also options to help manage the condition [119]. The relationship between seizure states and glial cell activity has recently been studied with increased scrutiny where immune system responses and activation of glia are known take part in both the human form of the disorder and in animal models [124–128]. It is the link between astrocyte activity to immune system responses that has caught the interest of many. Astrocytes like other glial cells are known to be involved with this regulation and they also regulate the permeability of blood brain barrier [25, 129, 130, 131, 132]. In epilepsy patients, however, the regulation of permeable compounds through the blood brain barrier is impaired and may have been caused via a number of different avenues

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

35

like trauma and auto-immune disorders [55]. Specifically, in the case following brain injury either through stroke or head injuries results in the upregulation of adhesion molecules leading to an increase in proinflammatory cytokines like IL-1β [55]. Furthermore, in patients that have a genetic susceptibility to brain inflammation are also linked to an increased risk of developing epilepsy [55]. Significantly, astrocytes, like other glial cells, are potentially a source of stored Ca that can influence synaptic transmission, thus by biochemical reduction of calcium signaling in astrocytes may provide a novel therapeutic target for future treatments and medication [25].

15 Parkinson’s Disease Parkinson’s disease (PD) is a progressive yet chronic neurodegenerative disorder that targets the brain’s motor functions and is characterized by involuntary tremors, posture instability, reduced mobility and control of limbs. To date, the standard treatment this disorder has been to administer the dopamine precursor L-DOPA. In patients, this compound is known to reduce the majority of symptoms but can have some severe side effects [133–135]. For these reasons, research in PD is focusing to obtain a detailed understanding of biochemistry underlying this complex brain disorder so that new treatments aimed at either slowing or stopping the advancement of PD can be developed while minimizing side effects. PD research has demonstrated that there is a glial component to the disease which has been observed in both PD patients and animal models of the condition. Recent studies have shown that glial cells can play either a protective role or a destructive one on dopaminergic neurons depending on a glial cells activated state [136–138]. Specifically, astrocytes produce neurotrophic factors like GDNF which supports the survival and growth of dopaminergic neurons located in the striatum [139]. During neurodegeneration, oxidative stress leads to increases in reactive oxygen species (ROS) that ultimately leads to degeneration of dopaminergic neurons and detectable macromolecular changes in the substantia nigra. The effects of ROS can be reduced by an enzyme called glutathione peroxidase. This compound is normally present in astrocytes located in the mesencephalon and can essentially stop the reaction of transforming hydrogen peroxide to hydroxyl based radicals [140–143]. The precise roles and mechanisms of glutathione peroxide in PD are currently unknown. Another source of ROS production in the substantia nigra of PD patients originates from autoxidation processes that leads to the loss of dopamine [143]. A novel in vitro study illustrated that astrocytes have the necessary molecular machinery to deal with ROS internally and thus effectively protecting neurons from the effects of ROS. This study showed that astrocytes that contain enzymes that destroy dopamine, like monoamine oxidase type B, also possess glutathione peroxidase. This allows the astrocyte to deal with ROS agents intracellularly without significant impact on neurons [144]. However, this seemingly protective nature of glia is not clear cut but could be state dependent.

36

N. Iannella and M. Condemine

This is supported by studies illustrating that glial cells respond to inflammation by producing known chemical compounds: cytokines, prooxidant reactive compounds, and prostaglandin, all of which are known to damage neurons [145]. In this regard, inflammatory responses of glial cells assist the progression of the disorder. Other chemical imbalances of glial origin are present in PD patients, such as higher expression levels of TNF-α, INF-γ and IL-1β and elevated expression levels of IL-2, IL-4, and IL-6 (proinflammatory cytokines) in the basal ganglia, thus suggesting increased cytokine production associated with the progression of PD [146–148]. These are not the only chemical imbalances in PD, the upregulation in the expression of major histocompatibility complex (MHC) molecules like HLA-DR positive microglia and β2-microglobulin has also been observed [149]. To complicate matters further, the involvement cytokines such as TNF-α, IL-1β, and IFN-γ are strong activators of nitric oxide synthase production in glia, ultimately leading to increased expression of CD23, the low affinity immunoglobulin E receptor [150]. Moreover, CD23 is strongly expressed in the substantia nigra of PD patients but not in healthy control subjects [139]. Furthermore, nitric oxide can also lead to iron released from the ferritin protein, thus adding a toxic component as well as increased formation of hydroxyl radicals [25, 141]. Overall, it seems that glial cell activity and cytokine production are delicately balanced where overproduction of cytokines leads to oxidative stress, which ultimately leads to the activation of proapoptotic signaling pathways and cell death in dopaminergic neurons, thus aiding the progression of PD.

16 Alzheimer’s Disease Another brain disorder that is gaining more attention is Alzheimer’s disease as its commonly associated with the elderly resulting in age-related cognitive decline characterized by progressive deficits in short term memory, cognitive impairment and changes in personality. Alzheimer patients are known to have greatly reduced brain mass (more than 35% in weight), where affected areas include the limbic system and the temporal lobe of the cerebral cortex. The molecular pathways responsible for this condition are currently unknown, however histological studies have suggested that an over production of β-amyloids, a macromolecule known to adversely affect specific biochemical signaling pathways that leads to the phosphorylation of the τ protein and the acceleration of the condition, in addition to other cellular changes including mitochondrial dysfunction. This has become known as the “amyloid cascade hypothesis” [151–153]. Furthermore, there seems to be a link between the production of βamyloids and the activity of microglia and astrocytes. Changes to amyloid precursor proteins through genetic mutation leads to increased production of β-amyloid-42 (Aβ-42) peptide; and this in turn gives rise to amyloid induced neurotoxicity that includes oxidative stress, free radical formation, and inflammation in the brain [154]. The complex nature underlying the onset and progression of Alzheimer’s disease lies in how environmental and genetic factors combine and interact over time, resulting in multiple alteration to the underlying biochemistry of the brain that is not attributable

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

37

to any single mechanism. Many genes have been connected to Alzheimer’s disease but the apolipoprotein E c4 (APOE-c4) gene has attracted attention since patients bearing this gene have been connected to a 50% increase in risk of developing Alzheimer’s [154]. Notably, APOE is acknowledged to be a protein that is involved in the development and repair of the brain primarily made and secreted by astrocytes. This raised the question whether astrocyte dysfunction contributed to the advance of Alzheimer’s disease. In particular, it was believed that the APOE-c4 isoform reduced the astrocyte’s ability to capture and internalize (phagocytosis) deposits of β-amyloids, hence allowing Alzheimer’s disease to progress [155]. However, recent experimental data has presented a controversy. Some studies have demonstrated that astrocytes take part in the protection of neurons by degrading and clearing β-amyloids deposits [156]. Whereas other data has shown that astrocytes responses to chronic stress increases βamyloids deposits through the overexpression of β-secretase and provides a scaffold for combining APP with γ -secretase, thus allowing the progression of the disease [155]. Other studies have additionally linked the genesis of Alzheimer’s disease to inflammatory processes. Notably the response to increased levels of β-amyloids leads to both astrocytes and microglia to protect neurons, but in turn, also leads to the release cytokines which can alter and even impair the function of neurons and contribute to their degeneration [157, 158]. Furthermore, astrocyte dysfunction could degrade the operation of the blood brain barrier. This adversely influences β-amyloid production in glia via clearance through transport processes involving the receptor for advanced glycation end products (RAGE) and the LDL-receptor-related protein 1 (LRP1), as they both play important roles in balancing the clearance and production of β-amyloids [159, 160]. Alzheimer patients are known to have altered relative distributions of these receptors. Finally, Alzheimer’s patients have altered calcium signaling in both neurons and glia. Due to the importance of calcium in regulating biochemical signaling in both neurons and glia, including neuronal excitability, changes to calcium will ultimately impact neural-glial signaling, especially the generation of calcium waves in glia which forms the basis of communication between neurons and glial cells permitting glia, including astrocytes, to modulate neural activity. Invitro studies involving cultured astrocytes have shown that the addition of β-amyloid increases both the amplitude and velocity of evoked calcium waves [161, 162]. This consequently leads to changes in the messages being communicated between neurons and glia, which may eventually give rise to a pathological state.

17 Conclusions We have seen that glial cells, once thought to be the substance that keeps neurons in place and the brain together, are now being viewed as complex biochemical components that not only help protect neurons from various forms of biochemical perturbations, but they also involved in neuronal regulation and synaptic plasticity. Glial

38

N. Iannella and M. Condemine

cell involvement in immune system responses adds an additional level of complex dynamics, however it is their constant communication with neurons and their ability to regulate the activity of synapses on different timescales that neural-glial signaling is starting to be viewed as an important contributor toward information processing in the brain. The precise actions of glia in both synaptic plasticity and information processing in the brain still needs to be fully elucidated, however it is their immune system responses in both normal and dysfunctional states that is starting to attract increased scientific attention, as it seems that biochemical signaling provided by glia may in fact provide unique pathways and treatments for a range of conditions and disorders that are targeted, and at the same time minimize any negative impact on neurons and functional circuits in general. In order to better understand the glial cell actions on neurons and the nervous system in general, especially with respect to information processing, the development of new mathematical and computational models/techniques will play a pivotal role in understanding the significance of neural-glial signaling for synaptic plasticity, cortical development, and information processing in the brain. Such future theoretical/computational studies will provide novel insights that experimentalists can check to verify or debunk model predictions. Significantly, at this instant in time, theoretical models of cortical network dynamics and development that consider neural-glial signaling are few and far between, however future theoretical and computational studies are expected to provide novel predictions especially in the area of brain dysfunction and aid data-driven development of new yet novel treatments. Acknowledgements N. Iannella would like to thank Prof Stephen Coombes and the staff at the School of Mathematical Sciences at the University of Nottingham, UK for hosting and making it possible to write this chapter. N. Iannella’s contribution was supported by the People Programme (Marie Curie Actions) of the European Union’s Seventh Framework Programme (FP7/2007-2013) under REA grant agreement No PCOFUND-GA-2012-600181.

References 1. Rose CR, Konnerth A (2001) Stores not just for storage: intracellular calcium release and synaptic plasticity. Neuron 31:519–522 2. Zucker RS (1999) Calcium-and activity-dependent synaptic plasticity. Curr Opin Neurobiol 9:305–313 3. Shouval HZ, Bear MF, Cooper LN (2002) A unified model of NMDA receptor-dependent bidirectional synaptic plasticity. Proc Natl Acad Sci 99:10831–10836 4. Azevedo FAC, Carvalho LRB, Grinberg LT, Farfel JM, Ferretti REL, Leite REP, Lent R, Herculano-Houzel S et al (2009) Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. J Comp Neurol 513:532–541 5. Snell RS (2010) Clinical neuroanatomy. Lippincott Williams & Wilkins 6. Kang J, Jiang L, Goldman SA, Nedergaard M (1998) Astrocyte-mediated potentiation of inhibitory synaptic transmission. Nat Neurosci 1:683–692 7. Bacci A, Verderio C, Pravettoni E, Matteoli M (1999) The role of glial cells in synaptic function. Philos Trans R Soc B: Biol Sci 354:403–409

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

39

8. Kurosinski P, Götz J (2002) Glial cells under physiologic and pathologic conditions. Arch Neurol 59:1524–1528 9. Verkhratsky A, Butt A (2013) Neuroglia: definition, classification, evolution, numbers, development. Glial Physiol Pathophysiol 73–104 10. Verkhratsky A, Butt AM (2013) Glial physiology and pathophysiology, Wiley 11. Laming PR, Syková E (1998) Glial cells: their role in behaviour, Cambridge University Press 12. Bushong EA, Martone ME, Jones YZ, Ellisman MH (2002) Protoplasmic astrocytes in CA1 stratum radiatum occupy separate anatomical domains. J Neurosci 22:183–192 13. Miller RH, Raff MC (1984) Fibrous and protoplasmic astrocytes are biochemically and developmentally distinct. J Neurosci 4:585–592 14. Oberheim NA, Wang X, Goldman S, Nedergaard M (2006) Astrocytic complexity distinguishes the human brain. Trends Neurosci 29:547–553 15. Allen NJ, Barres BA (2005) Signaling between glia and neurons: focus on synaptic plasticity. Curr Opin Neurobiol 15:542–548 16. Allen NJ, Barres BA (2009) Neuroscience: glia—more than just brain glue. Nature 457:675– 677 17. Spassky N, Merkle FT, Flames N, Tramontin AD, Garc´ıa-Verdugo JM, Alvarez-Buylla A, (2005) Adult ependymal cells are postmitotic and are derived from radial glial cells during embryogenesis. J Neurosci 25:10–18 18. Abney ER, Bartlett PP, Raff MC (1981) Astrocytes, ependymal cells, and oligodendrocytes develop on schedule in dissociated cell cultures of embryonic rat brain. Dev Biol 83:301–310 19. Schnitzer J, Franke WW, Schachner M (1981) Immunocytochemical demonstration of vimentin in astrocytes and ependymal cells of developing and adult mouse nervous system. J Cell Biol 90:435–447 20. Hartfuss E, Galli R, Heins N, Götz M (2001) Characterization of CNS precursor subtypes and radial glia. Dev Biol 229:15–30 21. Anthony TE, Klein C, Fishell G, Heintz N (2004) Radial glia serve as neuronal progenitors in all regions of the central nervous system. Neuron 41:881–890 22. Bringmann A, Pannicke T, Grosche J, Francke M, Wiedemann P, Skatchkov SN, Osborne NN, Reichenbach A (2006) Müller cells in the healthy and diseased retina. Prog Retinal Eye Res 25:397–424 23. Levitt P, Rakic P (1980) Immunoperoxidase localization of glial fibrillary acidic protein in radial glial cells and astrocytes of the developing rhesus monkey brain. J Comp Neurol 193:815–840 24. Bonni A, Sun Y, Nadal-Vicens M, Bhatt A, Frank DA, Rozovsky I, Stahl N, Yancopoulos GD, Greenberg ME (1997) Regulation of gliogenesis in the central nervous system by the JAK-STAT signaling pathway. Science 278:477–483 25. Barres BA (2008) The mystery and magic of glia: a perspective on their roles in health and disease. Neuron 60:430–440 26. Kettenmann H, Verkhratsky A (2011) Neuroglia, der lebende Nervenkitt. Fortschritte der Neurologie\textperiodcentered Psychiatrie 79: 588–597 27. Grosche J, Matyash V, Möller T, Verkhratsky A, Reichenbach A, Kettenmann H (1999) Microdomains for neuron–glia interaction: parallel fiber signaling to Bergmann glial cells. Nat Neurosci 2:139–143 28. Eroglu C, Barres BA (2010) Regulation of synaptic connectivity by glia. Nature 468:223–231 29. Eroglu C, Barres BA, Stevens B (2008) Glia as active participants in the development and function of synapses. In: Structural and functional organization of the synapse. Springer, pp 683–714 30. Metea MR, Newman EA (2006) Glial cells dilate and constrict blood vessels: a mechanism of neurovascular coupling. J Neurosci 26:2862–2870 31. Verkhratsky A, Toescu EC (2006) Neuronal-glial networks as substrate for CNS integration. J Cell Mol Med 10:869–879 32. Giaume C, Koulakoff A, Roux L, Holcman D, Rouach N (2010) Astroglial networks: a step further in neuroglial and gliovascular interactions. Nat Rev Neurosci 11:87–99

40

N. Iannella and M. Condemine

33. Dani JW, Chernjavsky A, Smith SJ (1992) Neuronal activity triggers calcium waves in hippocampal astrocyte networks. Neuron 8:429–440 34. Higashi K, Fujita A, Inanobe A, Tanemoto M, Doi K, Kubo T, Kurachi Y (2001) An inwardly rectifying K+ channel, Kir4. 1, expressed in astrocytes surrounds synapses and blood vessels in brain. Am J Physiol-Cell Physiol 281:C922–C931 35. Hille B et al (2001) Ion channels of excitable membranes, vol 507. Sinauer Sunderland, MA 36. Verkhratsky A, Kirchhoff F (2007) Glutamate-mediated neuronal–glial transmission. J Anat 210:651–660 37. Figley CR, Stroman PW (2011) The role (s) of astrocytes and astrocyte activity in neurometabolism, neurovascular coupling, and the production of functional neuroimaging signals. Eur J Neurosci 33:577–588 38. Gordon GRJ, Choi HB, Rungta RL, Ellis-Davies GCR, MacVicar BA (2008) Brain metabolism dictates the polarity of astrocyte control over arterioles. Nature 456:745–749 39. Clarke LE, Barres BA (2013) Emerging roles of astrocytes in neural circuit development. Nat Rev Neurosci 14:311–321 40. Ishibashi T, Dakin KA, Stevens B, Lee PR, Kozlov SV, Stewart CL, Fields RD (2006) Astrocytes promote myelination in response to electrical impulses. Neuron 49:823–832 41. Bradl M, Lassmann H (2010) Oligodendrocytes: biology and pathology. Acta Neuropathol 119:37–53 42. Kandel ER, Schwartz JH, Jessell TM, Siegelbaum SA, Hudspeth AJ (2000) Principles of neural science, vol 4. McGraw-hill, New York 43. Baumann N, Pham-Dinh D (2001) Biology of oligodendrocyte and myelin in the mammalian central nervous system. Physiol Rev 81:871–927 44. Johansson CB, Momma S, Clarke DL, Risling M, Lendahl U, Frisén J (1999) Identification of a neural stem cell in the adult mammalian central nervous system. Cell 96:25–34 45. Rakic P (1971) Neuron-glia relationship during granule cell migration in developing cerebellar cortex. A Golgi and electonmicroscopic study in Macacus rhesus. J Comp Neurol 141:283– 312 46. Rakic P (1972) Mode of cell migration to the superficial layers of fetal monkey neocortex. J Comp Neurol 145:61–83 47. Rakic P (1988) Specification of cerebral cortical areas. Sci 241:170–176 48. Rakic P (2009) Evolution of the neocortex: a perspective from developmental biology. Nat Rev Neurosci 10:724–735 49. Doetsch F, Caille I, Lim DA, Garc´ıa-Verdugo JM, Alvarez-Buylla A (1999) Subventricular zone astrocytes are neural stem cells in the adult mammalian brain. Cell 97:703–716 50. Kriegstein A, Alvarez-Buylla A (2009) The glial nature of embryonic and adult neural stem cells. Annu Rev Neurosci 32:149 51. Rash BG, Lim HD, Breunig JJ, Vaccarino FM (2011) FGF signaling expands embryonic cortical surface area by regulating Notch-dependent neurogenesis. J Neurosci 31:15604– 15617 52. Rash BG, Tomasi S, Lim HD, Suh CY, Vaccarino FM (2013) Cortical gyrification induced by fibroblast growth factor 2 in the mouse brain. J Neurosci 33:10802–10814 53. Mo Z, Zecevic N (2009) Human fetal radial glia cells generate oligodendrocytes in vitro. Glia 57:490–498 54. Reuss B, Unsicker K (2005) Neuroglia. In: Kettenmann H, Ransom BR (eds). Oxford University Press, New York 55. Ricci G, Volpi L, Pasquali L, Petrozzi L, Siciliano G (2009) Astrocyte–neuron interactions in neurological disorders. J Biol Phys 35:317–336 56. Perea G, Araque A (2005a) Glial calcium signaling and neuron–glia communication. Cell Calcium 38:375–382 57. Paixão S, Klein R (2010) Neuron–astrocyte communication and synaptic plasticity. Curr Opin Neurobiol 20:466–473 58. Auld DS, Robitaille R (2003) Glial cells and neurotransmission: an inclusive view of synaptic function. Neuron 40:389–400

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

41

59. Kettenmann H, Verkhratsky A et al (2008) Neuroglia: the 150 years after. Trends Neurosci 31:653 60. Achour SB, Pascual O (2012) Astrocyte–neuron communication: functional consequences. Neurochem Res 37:2464–2473 61. Pasti L, Volterra A, Pozzan T, Carmignoto G (1997) Intracellular calcium oscillations in astrocytes: a highly plastic, bidirectional form of communication between neurons and astrocytes in situ. J Neurosci 17:7817–7830 62. Pfrieger FW (2010) Role of glial cells in the formation and maintenance of synapses. Brain Res Rev 63:39–46 63. Araque A, Parpura V, Sanzgiri RP, Haydon PG (1999) Tripartite synapses: glia, the unacknowledged partner. Trends Neurosci 22:208–215 64. Perea G, Araque A (2005b) Properties of synaptically evoked astrocyte calcium signal reveal synaptic information processing by astrocytes. J Neurosci 25:2192–2203 65. Witcher MR, Kirov SA, Harris KM (2007) Plasticity of perisynaptic astroglia during synaptogenesis in the mature rat hippocampus. Glia 55:13–23 66. Verkhratsky A, Orkand RK, Kettenmann H (1998) Glial calcium: homeostasis and signaling function. Physiol Rev 78:99–141 67. Bolton MM, Eroglu C (2009) Look who is weaving the neural web: glial control of synapse formation. Curr Opin Neurobiol 19:491–497 68. Araque A, Carmignoto G, Haydon PG (2001) Dynamic signaling between astrocytes and neurons. Annu Rev Physiol 63:795–813 69. Wang X, Lou N, Xu Q, Tian G-F, Peng WG, Han X, Kang J, Takano T, Nedergaard M (2006) Astrocytic Ca2+ signaling evoked by sensory stimulation in vivo. Nat Neurosci 9:816–823 70. Amiri M, Montaseri G, Bahrami F (2011) On the role of astrocytes in synchronization of two coupled neurons: a mathematical perspective. Biol Cybern 105:153–166 71. Amiri M, Bahrami F, Janahmadi M (2012) Modified thalamocortical model: A step towards more understanding of the functional contribution of astrocytes to epilepsy. J Comput Neurosci 33:285–299 72. Postnov DE, Brazhe NA, Sosnovtseva OV (2012) Functional modeling of neural-glial interaction. In: Biosimulation in biomedical research, health care and drug development. Springer, pp 133–151 73. Postnov DE, Koreshkov RN, Brazhe NA, Brazhe AR, Sosnovtseva OV (2009) Dynamical patterns of calcium signaling in a functional model of neuron–astrocyte networks. J Biol Phys 35:425–445 74. Kell DB, Oliver SG (2004) Here is the evidence, now what is the hypothesis? The complementary roles of inductive and hypothesis-driven science in the post-genomic era. BioEssays 26:99–105 75. Postnov DE, Ryazanova LS, Sosnovtseva OV (2007) Functional modeling of neural–glial interaction. BioSystems 89:84–91 76. Postnov DE, Ryazanova LS, Brazhe NA, Brazhe AR, Maximov GV, Mosekilde E, Sosnovtseva OV (2008) Giant glial cell: New insight through mechanism-based modeling. J Biol Phys 34:441–457 77. Amiri M, Hosseinmardi N, Bahrami F, Janahmadi M (2013) Astrocyte-neuron interaction as a mechanism responsible for generation of neural synchrony: a study based on modeling and experiments. J Comput Neurosci 34:489–504 78. De Pittà M, Brunel N (2016) Modulation of synaptic plasticity by glutamatergic gliotransmission: a modeling study. Neural plasticity 2016 79. De Pittà M, Volman V, Berry H, Parpura V, Volterra A, Ben-Jacob E (2012) Computational quest for understanding the role of astrocyte signaling in synaptic transmission and plasticity. Front Comput Neurosci 6:98 80. Fellin T, Ellenbogen JM, De Pittà M, Ben-Jacob E, Halassa MM (2012) Astrocyte regulation of sleep circuits: experimental and modeling perspectives. Front Comput Neurosci 6:65 81. Gordleeva SY, Stasenko SV, Semyanov AV, Dityatev AE, Kazantsev VB (2012) Bi-directional astrocytic regulation of neuronal activity within a network. Front Comput Neurosci 6:92

42

N. Iannella and M. Condemine

82. Tewari SG, Majumdar KK (2012) A mathematical model of the tripartite synapse: astrocyteinduced synaptic plasticity. J Biol Phys 38:465–496 83. Nadkarni S, Jung P (2004) Dressed neurons: modeling neural–glial interactions. Phys Biol 1:35 84. Nadkarni S, Jung P (2005) Synaptic inhibition and pathologic hyperexcitability through enhanced neuron-astrocyte interaction: a modeling study. J Integr Neurosci 4:207–226 85. Nadkarni S, Jung P (2007) Modeling synaptic transmission of the tripartite synapse. Phys Biol 4:1 86. Izhikevich EM et al (2003) Simple model of spiking neurons. IEEE Trans Neural Networks 14:1569–1572 87. Reato D, Cammarota M, Parra LC, Carmignoto G (2012) Computational model of neuronastrocyte interactions during focal seizure generation. Frontiers Comput Neurosci 6 88. Porter JT, McCarthy KD (1997) Astrocytic neurotransmitter receptors in situ and in vivo. Prog Neurobiol 51:439–455 89. Kawabata S, Tsutsumi R, Kohara A, Yamaguchi T, Nakanishi S, Okada M (1996) Control of calcium oscillations by phosphorylation of metabotropic glutamate receptors. Nature 383:88– 92 90. Henneberger C, Papouin T, Oliet SHR, Rusakov DA (2010) Long-term potentiation depends on release of D-serine from astrocytes. Nature 463:232–236 91. Gordon GRJ, Iremonger KJ, Kantevari S, Ellis-Davies GCR, MacVicar BA, Bains JS (2009) Astrocyte mediated distributed plasticity at hypothalamic glutamate synapses. Neuron 64:391 92. Pascual O, Casper KB, Kubera C, Zhang J, Revilla-Sanchez R, Sul J-Y, Takano H, Moss SJ, McCarthy K, Haydon PG (2005) Astrocytic purinergic signaling coordinates synaptic networks. Science 310:113–116 93. Stellwagen D, Malenka RC (2006) Synaptic scaling mediated by glial TNF-α. Nature 440:1054–1059 94. Kaneko M, Stellwagen D, Malenka RC, Stryker MP (2008) Tumor necrosis factor-α mediates one component of competitive, experience-dependent plasticity in developing visual cortex. Neuron 58:673–680 95. Beattie EC, Stellwagen D, Morishita W, Bresnahan JC, Ha BK, Von Zastrow M, Beattie MS, Malenka RC (2002) Control of synaptic strength by glial TNFα. Science 295:2282–2285 96. Chiu C-S, Brickley S, Jensen K, Southwell A, Mckinney S, Cull-Candy S, Mody I, Lester HA (2005) GABA transporter deficiency causes tremor, ataxia, nervousness, and increased GABA-induced tonic conductance in cerebellum. J Neurosci 25:3234–3245 97. Bult CJ, Eppig JT, Kadin JA, Richardson JE, Blake JA, Group MGD et al (2008) The mouse genome database (MGD): mouse biology and model systems. Nucleic Acids Res 36: D724– D728 98. Doengi M, Hirnet D, Coulon P, Pape H-C, Deitmer JW, Lohr C (2009) GABA uptakedependent Ca2+ signaling in developing olfactory bulb astrocytes. Proc Natl Acad Sci 106:17570–17575 99. Gomeza J, Hülsmann S, Ohno K, Eulenburg V, Szöke K, Richter D, Betz H (2003) Inactivation of the glycine transporter 1 gene discloses vital role of glial glycine uptake in glycinergic inhibition. Neuron 40:785–796 100. Filosa A, Paixão S, Honsek SD, Carmona MA, Becker L, Feddersen B, Gaitanos L, Rudhard Y, Schoepfer R, Klopstock T et al (2009) Neuron-glia communication via EphA4/ephrin-A3 modulates LTP through glial glutamate transport. Nat Neurosci 12:1285–1292 101. Carmona MA, Murai KK, Wang L, Roberts AJ, Pasquale EB (2009) Glial ephrin-A3 regulates hippocampal dendritic spine morphology and glutamate transport. Proc Natl Acad Sci 106:12524–12529 102. Klein R (2009) Bidirectional modulation of synaptic functions by Eph/ephrin signaling. Nat Neurosci 12:15–20 103. Omrani A, Melone M, Bellesi M, Safiulina V, Aida T, Tanaka K, Cherubini E, Conti F (2009) Up-regulation of GLT-1 severely impairs LTD at mossy fibre–CA3 synapses. J Physiol 587:4575–4588

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

43

104. Rothstein JD, Patel S, Regan MR, Haenggeli C, Huang YH, Bergles DE, Jin L, Hoberg MD, Vidensky S, Chung DS et al (2005) β-Lactam antibiotics offer neuroprotection by increasing glutamate transporter expression. Nat 433:73–77 105. Holtmaat A, Svoboda K (2009) Experience-dependent structural synaptic plasticity in the mammalian brain. Nat Rev Neurosci 10:647–658 106. Trachtenberg JT, Chen BE, Knott GW, Feng G, Sanes JR, Welker E, Svoboda K (2002) Long-term in vivo imaging of experience-dependent synaptic plasticity in adult cortex. Nat 420:788–794 107. Christopherson KS, Ullian EM, Stokes CCA, Mullowney CE, Hell JW, Agah A, Lawler J, Mosher DF, Bornstein P, Barres BA (2005) Thrombospondins are astrocyte-secreted proteins that promote CNS synaptogenesis. Cell 120:421–433 108. Meyer-Franke A, Kaplan MR, Pfieger FW, Barres BA (1995) Characterization of the signaling interactions that promote the survival and growth of developing retinal ganglion cells in culture. Neuron 15:805–819 109. Elmariah SB, Hughes EG, Oh EJ, Balice-Gordon RJ (2004) Neurotrophin signaling among neurons and glia during formation of tripartite synapses. Neuron Glia biol 1:339–349 110. Eroglu C, Allen NJ, Susman MW, O’Rourke NA, Park CY, Özkan E, Chakraborty C, Mulinyawe SB, Annis DS, Huberman AD et al (2009) Gabapentin receptor α2δ-1 is a neuronal thrombospondin receptor responsible for excitatory CNS synaptogenesis. Cell 139:380–392 111. Xu J, Xiao N, Xia J (2010) Thrombospondin 1 accelerates synaptogenesis in hippocampal neurons through neuroligin 1. Nat Neurosci 13:22–24 112. Goritz C, Mauch DH, Pfrieger FW (2005) Multiple mechanisms mediate cholesterol-induced synaptogenesis in a CNS neuron. Mol Cell Neurosci 29:190–201 113. Mauch DH, Nägler K, Schumacher S, Göritz C, Müller E-C, Otto A, Pfrieger FW (2001) CNS synaptogenesis promoted by glia-derived cholesterol. Sci 294:1354–1357 114. Garcia O, Torres M, Helguera P, Coskun P, Busciglio J (2010) A role for thrombospondin-1 deficits in astrocyte-mediated spine and synaptic pathology in Down’s syndrome. PLoS ONE 5:e14200 115. Pacey LKK, Doering LC (2007) Developmental expression of FMRP in the astrocyte lineage: implications for fragile X syndrome. Glia 55:1601–1609 116. Jacobs S, Doering LC (2010) Astrocytes prevent abnormal neuronal development in the fragile x mouse. J Neurosci 30:4508–4514 117. Chahrour M, Zoghbi HY (2007) The story of Rett syndrome: from clinic to neurobiology. Neuron 56:422–437 118. Chahrour M, Jung SY, Shaw C, Zhou X, Wong STC, Qin J, Zoghbi HY (2008) MeCP2, a key contributor to neurological disease, activates and represses transcription. Sci 320:1224–1229 119. World Health Organization and others (2013) Fact sheet N 999: epilepsy. Avalaible http:// www.who.int/mediacentre/factsheets/fs999/en 120. Lowenstein DH (2009) Epilepsy after head injury: an overview. Epilepsia 50:4–9 121. Annegers JF, Hauser WA, Coan SP, Rocca WA (1998) A population-based study of seizures after traumatic brain injuries. N Engl J Med 338:20–24 122. DeLorenzo RJ, Sun DA, Deshpande LS (2005) Cellular mechanisms underlying acquired epilepsy: the calcium hypothesis of the induction and maintainance of epilepsy. Pharmacol Ther 105:229–266 123. Bhalla D, Godet B, Druet-Cabanac M, Preux P-M (2011) Etiologies of epilepsy: a comprehensive review. Expert Rev Neurother 11:861–876 124. Choi J, Koh S (2008) Role of brain inflammation in epileptogenesis. Yonsei Med J 49:1–18 125. Uhlmann EJ, Wong M, Baldwin RL, Bajenaru ML, Onda H, Kwiatkowski DJ, Yamada K, Gutmann DH (2002) Astrocyte-specific TSC1 conditional knockout mice exhibit abnormal neuronal organization and seizures. Ann Neurol 52:285–296 126. Coulter DA, Eid T (2012) Astrocytic regulation of glutamate homeostasis in epilepsy. Glia 60:1215–1226 127. Tanaka K, Watase K, Manabe T, Yamada K, Watanabe M, Takahashi K, Iwama H, Nishikawa T, Ichihara N, Kikuchi T et al (1997) Epilepsy and exacerbation of brain injury in mice lacking the glutamate transporter GLT-1. Sci 276:1699–1702

44

N. Iannella and M. Condemine

128. Vezzani A, French J, Bartfai T, Baram TZ (2011) The role of inflammation in epilepsy. Nat Rev Neurol 7:31–40 129. Janzer RC, Raff MC (1987) Astrocytes induce blood–brain barrier properties in endothelial cells. Nat 325:253–257 130. Abbott NJ, Rönnbäck L, Hansson E (2006) Astrocyte–endothelial interactions at the blood– brain barrier. Nat Rev Neurosci 7:41–53 131. Hawkins BT, Davis TP (2005) The blood-brain barrier/neurovascular unit in health and disease. Pharmacol Rev 57:173–185 132. Ballabh P, Braun A, Nedergaard M (2004) The blood–brain barrier: an overview: structure, regulation, and clinical implications. Neurobiol Dis 16:1–13 133. Cotzias GC, Papavasiliou PS, Gellene R (1969) Modification of parkinsonism—chronic treatment with L-dopa. N Engl J Med 280:337–345 134. Barbeau A (1969) L-dopa therapy in Parkinson’s disease: a critical review of nine years’ experience. Can Med Assoc J 101:59 135. Fahn S (1986) Recent developments in Parkinson’s disease. Raven Pr 136. Jackson-Lewis V, Vila M, Tieu K, Teismann P, Vadseth C, Choi DK, Ischiropoulos H, Przedborski S et al (2002) Blockade of microglial activation is neuroprotective in the 1methyl-4-phenyl-1, 2, 3, 6-tetrahydropyridine mouse model of Parkinson disease. J Neurosci 22:1763–1771 137. Hirsch EC, Breidert T, Rousselet E, Hunot S, Hartmann A, Michel PP (2003) The role of glial reaction and inflammation in Parkinson’s disease. Ann N Y Acad Sci 991:214–228 138. McGeer PL, McGeer EG (2008) Glial reactions in Parkinson’s disease. Mov Disord 23:474– 483 139. Hunot S, Bernard V, Faucheux B, Boissiere F, Leguern E, Brana C, Gautris PP, Guerin J, Bloch B, Agid Y et al (1996) Glial cell line-derived neurotrophic factor (GDNF) gene expression in the human brain: a post mortem in situ hybridization study with special reference to Parkinson’s disease. J Neural Trans 103:1043–1052 140. Sian J, Dexter DT, Lees AJ, Daniel S, Agid Y, Javoy-Agid F, Jenner P, Marsden CD (1994) Alterations in glutathione levels in Parkinson’s disease and other neurodegenerative disorders affecting basal ganglia. Ann Neurol 36:348–355 141. Jenner P, Dexter DT, Sian J, Schapira AHV, Marsden CD (1992) Oxidative stress as a cause of nigral cell death in Parkinson’s disease and incidental Lewy body disease. Ann Neurol 32:S82–S87 142. Desagher S, Glowinski J, Premont J (1996) Astrocytes protect neurons from hydrogen peroxide toxicity. J Neurosci 16:2553–2562 143. Desagher S, Glowinski J, Prémont J (1997) Pyruvate protects neurons against hydrogen peroxide-induced toxicity. J Neurosci 17:9060–9067 144. Kastner A, Anglade P, Bounaix C, Damier P, Javoy-Agid F, Bromet N, Agid Y, Hirsch EC (1994) Immunohistochemical study of catechol-O-methyltransferase in the human mesostriatal system. Neurosci 62:449–457 145. Anglade P, Vyas S, Javoy-Agid F, Herrero MT, Michel PP, Marquez J, Mouatt-Prigent A, Ruberg M, Hirsch EC, Agid Y (1997) Apoptosis and autophagy in nigral neurons of patients with parkinson’s disease. Histol Histopathol 12:25–32 146. Mogi M, Harada M, Kondo T, Riederer P, Inagaki H, Minami M, Nagatsu T (1994) Interleukin1β, interleukin-6, epidermal growth factor and transforming growth factor-α are elevated in the brain from parkinsonian patients. Neurosci Lett 180:147–150 147. Mogi M, Harada M, Riederer P, Narabayashi H, Fujita K, Nagatsu T (1994) Tumor necrosis factor-α (TNF-α) increases both in the brain and in the cerebrospinal fluid from parkinsonian patients. Neurosci Lett 165:208–210 148. Mogi M, Harada M, Narabayashi H, Inagaki H, Minami M, Nagatsu T (1996) Interleukin (IL)-1β, IL-2, IL-4, IL-6 and transforming growth factor-α levels are elevated in ventricular cerebrospinal fluid in juvenile parkinsonism and Parkinson’s disease. Neurosci Lett 211:13–16 149. McGeer PL, Schwab C, Parent A, Doudet D (2003) Presence of reactive microglia in monkey substantia nigra years after 1-methyl-4-phenyl-1, 2, 3, 6-tetrahydropyridine administration. Ann Neurol 54:599–604

Neurons and Plasticity: What Do Glial Cells Have to Do with This?

45

150. Dugas B, Mossalayi MD, Damais C, Kolb J-P (1995) Nitric oxide production by human monocytes: evidence for a role of CD23. Immunol Today 16:574–580 151. Strittmatter WJ, Saunders AM, Schmechel D, Pericak-Vance M, Enghild J, Salvesen GS, Roses AD (1993) Apolipoprotein E: high-avidity binding to beta-amyloid and increased frequency of type 4 allele in late-onset familial Alzheimer disease. Proc Natl Acad Sci 90:1977–1981 152. Tanzi RE, Gusella JF, Watkins PC, Bruns GA, St George-Hyslop P, Van Keuren ML, Patterson D, Pagan S, Kurnit DM, Neve RL (1987) Amyloid beta protein gene: cDNA, mRNA distribution, and genetic linkage near the alzheimer locus. Sci 235: 880–884 153. Hardy J, Selkoe DJ (2002) The amyloid hypothesis of Alzheimer’s disease: progress and problems on the road to therapeutics. Sci 297:353–356 154. Rocchi A, Pellegrini S, Siciliano G, Murri L (2003) Causative and susceptibility genes for alzheimer’s disease: a review. Brain Res Bull 61:1–24 155. Roßner S, Lange-Dohna C, Zeitschel U, Perez-Polo JR (2005) Alzheimer’s disease β-secretase BACE1 is not a neuron-specific enzyme. J Neurochem 92:226–234 156. Heneka MT, O’Banion MK (2007) Inflammatory processes in alzheimer’s disease. J Neuroimmunol 184:69–91 157. Meda L, Cassatella MA, Szendrei GI, Otvos L, Baron P, Villalba M, Ferrari D, Rossi F (1995) Activation of microglial cells by β-amyloid protein and interferon-γ. Nat 374:647–650 158. Meda L, Baron P, Scarlato G (2001) Glial activation in Alzheimer’s disease: the role of Aβ and its associated proteins. Neurobiol Aging 22:885–893 159. Deane R, Wu Z, Zlokovic BV (2004) RAGE (Yin) versus LRP (Yang) balance regulates Alzheimer amyloid β-peptide clearance through transport across the blood–brain barrier. Stroke 35:2628–2631 160. Farfara D, Lifshitz V, Frenkel D (2008) Neuroprotective and neurotoxic properties of glial cells in the pathogenesis of alzheimer’s disease. J Cell Mol Med 12:762–780 161. Mattson MP, Chan SL (2003a) Calcium orchestrates apoptosis. Nat Cell Biol 5:1041–1043 162. Mattson MP, Chan SL (2003b) Neuronal and glial calcium signaling in Alzheimer’s disease. Cell Calcium 34:385–397

Nicolangelo Iannella I received the B.Sc., B.Sc.(Hons) and M.Sc. degrees in Theoretical Physics from the University of Adelaide and Flinders University of South Australia in 1990, 1991 and 1995, respectively, and the Ph.D. in Computational Neuroscience from the University of ElectroCommunications, Japan in 2009. From 2009, he was a Postdoctoral Researcher in RIKEN BSI. In 2010, he won the prestigious Australian Research Council (ARC) Australian Postdoctoral Award (APD) fellowship, based at the University of Adelaide from 2010–2014. In 2012 he received the Grad Cert in Education (Higher Education) (GCEHE) from the University of Adelaide. From 2014–2017 he was an adjunct research fellow at the University of South Australia. From 2016– 2018, he was a Cascade (Marie Curie) Research Fellow in Mathematical Sciences at the University of Nottingham. From 2018, he is a research fellow at the University of Oslo. His research interests include synaptic plasticity, neuronal dynamics and neuromorphic engineering. Dr. Iannella is a member of OCNS, INNS, SFN and IEEE. Michel Condemine received the Master’s degree in Engineering from École Nationale Supérieure d’Arts et Métiers (ENSAM—Paris, France) in 1991. During his Master’s degree he applied neurofuzzy techniques to the manufacture of automotive parts. In 1992, he participated in a DEA (Postgraduate Diploma) in Neuroscience, focusing on building a vision system for autonomous robots based upon the visual system of the fly. From 1993–1998, he has worked in industry focusing on the development and implementation of neural network solutions for predictive control, predictive maintenance and visual defect detection in the nuclear power sector. Since 1998, his main activity has been in consultancy domain by providing SAP software solutions integrating Demand,

46

N. Iannella and M. Condemine

Planning, Supply and Logistics, Production and Scheduling. Since 2008, he has been involved in research and development of spiking neural networks and their application to develop AI inspired business solutions and their integration into software applications/platforms.

Molecular Brain Imaging

Transcranial Dynamic Fluorescence Imaging for the Study of the Epileptic Seizures Vyacheslav Kalchenko, Alon Harmelin, David Israeli, Babak Kateb, Igor Meglinski, Qinggong Tang, Nitish V. Thakor, Alla Ignashchenkova, Anna Volnova, and Vassiliy Tsytsarev

V. Kalchenko (B) Department of Veterinary Resources, Weizmann Institute of Science, 234 Herzl St., Rehovot 7610001, Israel e-mail: [email protected] A. Harmelin Vice President for Administration and Finance, Weizmann Institute of Science, 234 Herzl St., Rehovot 7610001, Israel e-mail: [email protected] D. Israeli Psychiatric Array, Kaplan Hospital, The Hebrew University of Jerusalem, Jerusalem, Israel e-mail: [email protected] B. Kateb Founding Chairman of the Board of Directors, CEO and Scientific Director, President & Scientific Director, Society for Brain Mapping & Therapeutics (SBMT), Brain Mapping Foundation, 860 Via De La Paz, Suite E-1 Pacific Palisades, Pacific Palisades 90272-3668, CA, USA e-mail: [email protected] I. Meglinski School of Engineering and Applied Science, Aston University, Birmingham B4 7ET, United Kingdom e-mail: [email protected] Q. Tang Stephenson School of Biomedical Engineering, University of Oklahoma, Norman, OK 73019, USA e-mail: [email protected] N. V. Thakor Department of Biomedical Engineering, John Hopkins University, Baltimore, USA e-mail: [email protected] A. Ignashchenkova · A. Volnova Saint Petersburg State University, Translational Research Institute, Saint Petersburg, Russia e-mail: [email protected] © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_3

49

50

V. Kalchenko et al.

1 Introduction Accurate localization of epileptic seizures has significant importance in advancing the antiepileptic therapy. About 2 in 100 people in the United States experience an unprovoked seizure at least once in their lives. Variable types of the epilepsy is a neurological disorder in which the neurons activity is disturbed, causing a seizure during which patient may experience different pathological symptoms, including abnormal behavior, sensations, loss of consciousness and convulsion. There are many types of epilepsy but all of them results from pathological neural synchronization inside the brain that cause recurring seizures. There are many approaches to monitoring and assaying epileptic seizures. These include simultaneous recording of clinical electroencephalography (EEG) along with behavioral, biochemical analysis and eventual verification through histological changes. These methods provide limited understanding about the fundamental mechanisms of altered brain function after the seizures onset [1]. Over the past years these approaches have often been combined with different imaging methods in epilepsy research. Low-resolution electromagnetic tomography (LORETA), functional magnetic resonance imaging (fMRI), single photon emission spectroscopy (SPECT), positron-emission tomography (PET), optical imaging of intrinsic signals (IOS) and near-infrared spectroscopy (NIRS) were used regularly [2–4]. Voltagesensitive dye optical imaging (VSDi) and photoacoustic (PA) were successfully applied to animal models of epilepsy [5, 6]. The combination of different optical imaging methods in clinical and preclinical research along with electrophysiological, histological and behavioral methods hold the potential to revolutionize epilepsy research studies. Such combination methods need coordinated use of the technology, experimentation and data analysis. Here we review the coordinated approach to optical imaging methods, including laser speckle contrast imaging in translational research with a specific focus on epilepsy research. Use of animal model for neuroscience diseases like epilepsy is a key area of focus in brain optical imaging translational research. Animal models enable both development and testing of original methods and obtaining fundamental research A. Volnova e-mail: [email protected] V. Tsytsarev Department of Anatomy and Neurobiology, University of Maryland, 20 Penn st, HSF-2, Baltimore 21201, MD, USA e-mail: [email protected] B. Kateb Society of Brain Mapping and Therapeutics, Santa Monica, CA, USA Brain Mapping Foundation, West Hollywood, CA, USA National Center for NanoBioElectronics, Los Angeles, CA, USA Brain technology and Innovation Park, Los Angeles, CA, USA

Transcranial Dynamic Fluorescence Imaging for the Study …

51

results. Animal research can also set the stage for subsequent clinical translation and application. Animal model research aims to expedite the translation of scientific discovery into new methods of neural disorders therapy. It further promotes a wide-ranging exchange between, preclinical, clinical and fundamental neuroscience research. Cortical seizures can be induced by chemoconvulsants—pilocarpine, 4Aminopyridine, kainic acid and some others [7]. These models intend to mimic temporal lobe epilepsy. Kainic acid is an L-glutamate analog, being injected systematically or intracerebral causes neuronal depolarization and seizures. Another commonly use epileptogenic agent is pilocarpine, a muscarinic acetylcholine receptor agonist. Systemic or intracerebral injection of pilocarpine also causes seizures in the animal model. 4-Aminopiridine (potassium channel inhibitor), pentylenetetrazol (PTZ)—causes calcium channels to loose selectivity and conduct sodium ions and depolarize the neuron, and many other specific convulsants are widely used as acute seizure models, and not as animal models of epilepsy [7]. Various forms of brain optical imaging methods are now also being performed as research tools in animal model. Physiological base of the optical imaging of the epileptic seizures is the neurovascular coupling concerns the relationship between neuronal activity, local oxygenation, and blood circulation [8]. Focal epileptic seizures have been shown to elicit changes in the tissue oxygenation, utilization of glucose and increases in local blood circulation [9]. Transient local changes in the cerebral oxygenation may be excellent markers for the focus of the epileptic activity in the brain and may even precede the onset of the epileptic seizures [8, 10]. Optical imaging methods can localize epileptic foci and significantly identify preictal optical signals from both human epilepsy and animal model of the epileptic seizures. Perhaps they can also can help with the antiepileptic therapy based on the closed-loop device to predict and terminate seizures [8]. Imaging of the epileptic seizures can be based on the monitoring of the cerebral blood flow, local oxygenation and other physiological process including neural firing and glucose consumption. Monitoring the cerebral blood flow (CBF) and brain metabolism is critical for many clinical and fundamental neuroscience studies, and it can be performed by variable methods. First, laser—Doppler flowmetry is a prevalent method which provides data about CBF from a limited number of isolated areas in the brain. Change of this method to the scanning mode can help to obtain spatially resolved CBF data, but both spatial and temporal resolution remain limited [11, 12]. Second, diffuse optic tomography (DOT) successfully localized the seizure onset zone in rat neocortex [13]. In combination with the electrophysiological recording the DOT identified the seizure focus very accurately at different locations and different depths in an acute model of focal epilepsy [13]. Another promising technology employed a fluorescent 2-DG analog, 2-(N-(7nitrobenz-2-oxa-1,3-diazol-4-yl)amino)-2-deoxyglucose (2-NBDG), for visualization of in vivo pharmacologically induced epileptic seizures [14]. The increased uptake rate of the fluorescence glucose substitute, 2-NBDG reflected elevated local metabolism of the neural tissue caused by the epileptic activity [15]. The results

52

V. Kalchenko et al.

demonstrate that fluorescent glucose substitutes as well as some other biomarkers of the local metabolism have the potential to be used as a contrast agent to visualize epileptic activity in vivo [14]. Optical imaging of the epileptic seizures plays an important role not only in the animal model studies but also in clinical neuroscience. Surgical removal of epileptogenic cortical areas in which seizures originate offers patients the chance of being seizure free. Precise localization of the epileptogenic areas is critical and many screening technologies based on the occurrence of epileptic activities have been used during the preoperative evaluation phase for antiepileptic surgeries. These technologies include, but are not limited to, electroencephalograms (EEG) (fMRI), positron emission tomography (PET), and single-photon emission computed tomography (SPECT). However, all of these techniques have serious limitations. Electrophysiological recording with implanted long-term electrodes elevating the risks of brain damage, like infection or hemorrhage. Unfortunately, the cost of use of MRI or fMRI is still very high and only a limited number of research centers have the financial and infrastructural capabilities to offer these methods for many patient. When practically used IOS, the temporal resolution is lower and spatial properties of the imaging data is more valuable. To determine the spatial representation of the optical signals, we raw data can be thresholded using a percentage of the maximum pixel value [4, 16, 17]. The threshold can be set at 25–50% of the maximal amplitude, and all pixels above this threshold can be defined as activated area [17]. The geometrical center of the threshold area can be defined as epicenter [17, 18]. During the last decade, a type of intrinsic optical imaging, called dynamic intrinsic optical signal imaging (DIOSI) has been considered useful as an alternative intraoperative technique to distinguish epileptogenic cortical areas [19]. DIOSI is based on the acquisition of the brain surface images under 500 and 700 nm simultaneously. Green light images (500 nm) reflect changes in the hemodynamic, while red and near-infrared light (700 nm and more) is sensitive for the changes in the oxygenation and deoxygenation due to the difference in the light absorption between oxyand deoxy-hemoglobin at this part of spectrum. Both cerebral blood flow and local oxygenation can be affected by neural activity that offer a window of opportunity to separate normal and epileptogenic cortical areas [19]. Regarding the practical application, monitoring of the CBF could provide valuable information for clinical as well as basic neuroscience research, but existing CBF imaging methods are frequently limited by physical hurdles. Methods, based on the infusion into the blood stream isotopic tracers provide three-dimensional (3D) information about CBF with acceptable temporal resolution [12, 20]. Positron emission tomography (PET) with water, marked by oxygene-15 became a gold standard for in vivo imaging of CBF, but is has some limitation in temporal and spatial resolution [21]. Methods based on fMRI and positron provide functional brain maps of CBF but the resolutions remain limited by the scanner’s technical characteristics [22, 23]. Use of ionizing radiation in the case of PET, costs of investigations frequently complicate the use of PET and fMRI in both preclinical studies and clinical practices. Other methods, which are appropriate for the transcranial monitoring of the

Transcranial Dynamic Fluorescence Imaging for the Study …

53

CBF, are ultrasonography and Laser Doppler flowmetry (LDF). Unfortunately, the first method is limited to observations of only large vessel flow velocities, while LDF is invasive [24]. A relatively cheap method for CBF monitoring is based on the near-infrared spectroscopy (NIRS) that derives concentration changes of oxy- and deoxyhemoglobin [24], but the spatial resolution and sensitivity frequently remains insufficient. The main goal of the optical imaging of transcranial blood flow recording or imaging was to create an alternative to MRI, CT and PET to overcome these methods’ cost and complexity. It is especially important to track the blood circulation in the branches of the particular blood vessels. For this purpose optical imaging methods employed both infrared (IR) and near-infrared (NIR) fluorescence probes [25, 26]. Thus, because of noninvasive and relatively low cost nature of near-infrared spectroscopy (NIRS), it shows tremendous promise for a range of preclinical and clinical applications. NIRS could also contributes to the diagnosis and treatment of cerebrovascular diseases and many neurological disorders including depression, epilepsy and Alzheimer’s disease. In animal experiments, multi-channel NIRS has been developed to continuously monitor the concentration change of oxy- and deoxyhemoglobin to elucidate physiological changes in response to the different strength impaction and it was shown that the status of traumatic brain injury can be clinically evaluated by detecting by the monitoring of local oxygenation and blood circulation [27]. Combination of VSDi and IOS imaging has been successfully used for precision study of the neurovascular coupling during the epileptic seizures [28]. On the other hand, the combination of VSDi, based on the fluorescence in the red part of spectrum, and NIRS, is technically difficult. VSDi reflects mainly synaptic activity in the upper layer of the neocortex but carries little information about activity of the glial cells or neurons’ action potentials, which play important role in the neurovascular coupling [16]. Regardless of the experimental condition, raw imaging data can be converted to fractional change in fluorescence by dividing the change in fluorescence by the background fluorescence (F/F) [29, 30]. To decrease biological noise caused by the heart rhythm and respiration, data can be high-pass filtered with 0.5–1.5 Hz half amplitude cutoff [4, 30]. Information about temporal changes of the VSDi signal can be extracted from the imaging data by using a distinguished region of interest centered on the pixel with maximum activation within 15–30 ms after stimulus onset [4, 30]. First introduced a couple of decades ago, laser speckle contrast imaging (LSCI) is a simple but very suitable tool for wide-field imaging of the cerebral blood flow and oxygenation [11] (Fig. 1). From the physical point of view, speckle arises from the random interference of coherent light [11, 31]. In the brain parenchyma coherent light interacts with a random scattering substrate and a CCD camera will receives light that has scattered from varying positions within the brain parenchyma and have traveled a distribution of distances that varies with the arrangement oxygenated and deoxygenated hemoglobin in the blood cells with respect to the CCD camera [31, 32]. If blood cells are moving, this causes fluctuations in the interference, which appears as intensity variations at the pixels of the CCD camera [11]. This scattering causes the speckle pattern to blur.

54

V. Kalchenko et al.

Fig. 1 Schematic of the low-cost, compact laser speckle imaging system. Laser illumination generates speckle from the lights scattered by red cells in the blood vessels in the brain causing intensity fluctuations. From these fluctuations (standard deviation and mean), blood flow parameter K is calculated and a laser speckle contrast image is constructed. The image is rendered on the brain by functional augmented reality brain mapping

In contrast with 2-photon scanning imaging methods, LSCI is unable to provide depth resolved images, but on the other hand the spatial resolution (~10 μm) and temporal resolution (10 ms to 10 s) of LSCI is relatively high [33]. Thus, in the brain imaging, erythrocytes are the main source of moving scatters and works as a natural contrast agent [33]. The spatial resolution is in proportion to the erythrocytes size and can reach 10 μm, while the temporal resolution mainly determined by the hardware properties and can reach tens of milliseconds. The speckle pattern can be modulated by the motion of scattering centers in the area of the imaging. The properties of the time-varying speckle pattern can be used for measuring the scattering object’s speed [34]. Thus, potentially, this method can significantly expand the capabilities of preclinical studies due to its simplicity, low cost and ease of applicability. Since the method is dye-free and continuous, it also has a great potential in fundamental neuroscience to study stimulus response of the brain and neuro-vascular coupling.

Transcranial Dynamic Fluorescence Imaging for the Study …

55

2 Current Method for the Transcranial Cbf Imaging The optical- and optical-ultrasonic based imaging methods, optical coherence tomography (OCT) and photoacoustic tomography (PAT) have been successfully applied for the imaging of the cereal blood flow within last couple of decades. OCT provides microanatomical reconstruction of brain tissues with micrometer resolution and relatively high speed [35]. Functional OCT enables label-free monitoring of hemodynamic and metabolic changes in the brain parenchyma [35]. In contrast with computed tomography (CT) and magnetic resonance angiography, the optical imaging methods provides higher temporal resolution. Besides that, optical imaging methods usually are portable and cost effective. The application of the optical imaging methods in clinical neuroscience is limited due to very high light absorption and scattering of the skull and scalp tissue, but these limitations are not so critical in the animal modeling studies. Transcranial application of the near infrared spectroscopy (NIRS) was successfully applied for the monitoring of the cerebral tissue oxygenation and cerebral blood volume in the human infants [36]. Without any fluorescence probe, it was used in diffusion correlation spectroscopy for the measurement of the cerebral blood flow index and the to test the validity of the CBV-CBF relationship in premature neonates and to cerebral metabolic rate of oxygen was estimated with and without the cerebral blood flow index [36, 37]. In animal experiments noninvasive monitoring of the oxygenated and deoxygenated hemoglobin concentration was realized by photoacoustics – the hybrid, probe-less method of the imaging based on the direct transducing of the light energy into the energy of the ultrasonic waves [6, 38, 39]. Functional mapping of the brain activity based on the endogenous fluorescence of mitochondrial flavoprotein also can be employed for the transcranial imaging in animal experiments. Mice skull is thin and sufficiently transparent, so this method was used to investigate auditory cortical plasticity [40, 41]. Among the numerous applications of medical brain optical imaging, the application of the above mentioned methods to the visualization of the epileptic seizures is very important. While currently, the localization of the epileptic foci generally is generally determined by EEG and Electrocorticographic recordings, future approaches may involve functional MRI– in combination with the optical imaging. The feasibility of using the optical characteristics of cortical seizures as a potential means to guide neurosurgery was shown [42] recently. One such system used for an imaging consisted of the light source and spectrometric light sensitive elements under central computer control [42]. The data was obtained from the cortical area affected by the epileptic seizures and from normal (control) cortex, and significant differences were found between them. The diffuse reflectance intensity of normal cortex was significantly lower than that of epileptic area in the 400–650 nm part of spectrum. We believe that the elevation in diffuse reflectance intensities between 400 and 600 nm in the area is affected by the seizure-related changes of the local cerebral blood

56

V. Kalchenko et al.

flow. This conclusion is in agreement with the MRI data [43]. Both oxygenationrelated process and the imaging of the cerebral blood flow can be used for the monitoring of the seizures–related activity. The migration of leukocytes from the blood vessels to the neural tissue plays a very important role in the inflammatory-related neurological diseases [44]. The leukocyte trafficking was monitored by both singleand 2 photon imaging [44]. Nevertheless, the cellular mechanism of leukocyte trafficking in the brain remains unclear in spite of serious achievements of the last decade. The spatially and timely resolved monitoring of the cerebral vasomotion is critical for both clinical and experimental neuroscience. The noninvasive optical imaging of brain vascular vasculature using a numbering of the fluorescence probes has been considered as showing very little promise. However, recently a robust noninvasive optical-based imaging approach that allows monitor cerebral vasodynamics at the high temporal and spatial resolution has been described [45]. Briefly, this method utilizes standard fluorescent dyes in combination with imaging and image analysis procedures. Recently LSCI has been successfully used for the imaging of the epileptic seizures in animal model [46]. In addition to blood flow, deoxyhemoglobin saturation has been measured simultaneously with LSCI in the cerebral cortex to provide an extended set of measures characterizing physiological changes during the epileptic seizures [46]. Authors used 4-Aminoperedine (4-AP) model of the epileptic seizures, which is commonly used during last two decades [47]. Combining LSCI with other optical imaging modalities allows for simultaneous monitoring of changes in oxy- and deoxyhemoglobin concentration as well as in the cerebral blood flow, caused by the epileptic seizures, elicited by intracortical injection of 4-AP [46].

3 Biological Base Transcranial Fluorescence Imaging A recent study [48] demonstrates a through-scalp and through-skull fluorescence imaging of mouse cerebral vasculature without craniotomy, utilizing the intrinsic photoluminescence of single-walled carbon nanotubes in the 1.3–1.4 μm nearinfrared window. This impressive study presents a new imaging approach with a greater depth of penetration and spatial resolution than standard optical fluorescent imaging. Physically, this method is based on dynamic fluorescent (DF) imaging at a rate around 5.3 frames per second, which is considered as sufficient for real-time assessment of blood flow anomalies in a mouse middle cerebral artery occlusion (MCAO) stroke model. Many of the brain imaging methods are not adequate for high resolution transcranial functional brain mapping, but novel hybrid method of the brain imaging— photoacoustic is able to monitor local cerebral oxygenation noninvasively [6, 49]. We believe that this technology may serve as powerful method for functional neuroimaging studies in animal model of the brain deceases. Biologically, the functional brain optical imaging is based on the monitoring of the neurovascular coupling—the complicated relationship between local neural activity,

Transcranial Dynamic Fluorescence Imaging for the Study …

57

Fig. 2 (Left) A typical configuration for the data acquisition by transcranial optical vascular imaging fluorescence. (Right) The animal’s head is fixed in the stereotaxic frame, fluorophore is injected into blood stream. Vascular images are overlaid on the brain

glial cells and subsequent changes in the CBF. The magnitude and spatial location metabolic changes are tightly linked to the neural network, main components of which are shown on the Fig. 2. Pericytes play very important roles in the neurovascular coupling—these are highly specialized cells located on the capillaries, as well as on small venules and arterioles [50]. These cells are separated from the brain parenchyma by the basal lamina, a thin layer between the endothelial cells and pericyte [50, 51]. Pericytes controlled by neuronal terminals (Fig. 2) as well as by the blood flow. Many scientists believe that pericytes are particularly sensitive to damage during pathological processes including epilepsy, ischemia and Alzheimer’s disease [50]. It was reported that the change in capillary diameter produced by pericytes reduces vascular resistance sufficiently to contribute significantly to the local cerebral blood flow intensification evoked by normal or pathological neural activity [52]. Of particular relevance to brain optical imaging research, high density of capillaries means that the cerebral blood flow in the capillaries could allow relatively high spatially-restricted changes in the local metabolism on occurrence of a localized changes of the neuronal activity [50]. It was reported that genetically modified mice reveal structural differences between pericytes adjacent to arterioles versus those distributed in the capillary bed. The 3D view of pericyte distribution along the cortical tissue was obtained using optical clearing of brain [50, 53, 54]. Redistribution of pericytes have been demonstrated after epileptic status [55, 56]. Vasoconstriction of the pericyte-controlled capillary was observed in both genetic epilepsy and kainic acid model of epilepsy [56, 57]. We would like to emphasize that during brain vascular imaging, high temporal resolution is essential and provides crucial information about blood perfusion, especially in preclinical DF imaging of stroke. Stroke is known to be among the most

58

V. Kalchenko et al.

prevalent causes of death and adult disability in humans, especially in highly developed countries. The middle cerebral artery (MCA) territory is the most commonly affected during cerebral infarction. Unfortunately, preclinical research currently relies on a handful of approaches for the study and imaging of stroke in vivo. It is still a huge challenge for preclinical scientists to visualize the degree and character of hemodynamic changes after induction of MCAO. In view of that, it is extremely important to have a simple and robust approach that allows clear visualization of MCA topography and perfusion territory, as well as of other arteries and veins, at a sufficient spatial resolution but with high temporal resolution, even at the cost of reduced depth of penetration. Alzheimer’s disease (AD) patient are more likely to get epilepsy [58]. Recent evidence suggests that nonconvulsive neural network abnormalities including seizures may be more commonly found in patients with familial AD [59]. The beta-amyloid peptide has been identified as a possible link between AD and pathological synchronization of the neural network [59]. Near-infrared fluorescence (NIRF) molecular imaging recently has been successfully applied in animal model of Alzheimer’s disease (AD) [60]. It was demonstrated that analog of curcumin, CRANAD-3 is a suitable for the beta-amyloid detecting, which brain deposits have been identified as key players in the AD progression and AD—linked epileptic seizures [61–63]. Design, synthesis, and testing of a curcumin-derivatized nearinfrared (NIR) probes, CRANAD-2 and CRANAD-3 has been described recently [63]. CRANAD-3 displayed significant increases in fluorescence intensity at the near infrared part of spectrum for beta-amyloid and emission peak around 730 nm. Also CRANAD-3 exhibited strong binding with different types of beta-amyloid [63]. Koronyo et al. also have shown both in animal studies and clinical trials that beta-amyloid deposits could be imaged via retinal imaging [64]. The optic signal from CRANAD-3 is usable for in vivo NIRF transgenic mice AD (APP/PS1), which is the most studied transgenic AD mouse model [65] and indicates an early molecular pathology. Moreover, data suggested that CRANAD-3 could monitor the decrease in beta-amyloid after the specific anti-AD therapy [60]. The fluorescence signal in the CRANAD-17–treated animals was lower than that in the control mice, and the result was verified by immunohistochemistry. Authors believe that it was a first time that NIRF was successfully applied for the monitoring of the anti-AD therapy [60]. Imaging agents specifically targeting beta amyloid plaques in the brain tissue play a key role in the early diagnosis of AD and recent evidence strongly implicates a tight genetic association between beta amyloid overexpression, plaques and epilepsy [66]. Evaluation of a NIRF imaging probes with donor-acceptor architecture bridged by a conjugated electron chain for beta amyloid plaques was described recently [67–69], but the many questions about physical aspects of the transcranial imaging remain open.

Transcranial Dynamic Fluorescence Imaging for the Study …

59

4 Some Perspectives of the Transcranial Fluorescence Imaging in Experimental and Clinical Practice During the last few decades, the scientific community was sure that noninvasive brain optical imaging methods cannot be informative enough in the case of using standard fluorescence probes [48], but the possibilities of transcranial high-spatial resolution fluorescence imaging have been shown recently [45]. It was possible, using fluorescent probe, but imaging with the help of single-walled carbon nanotubes in the 1.3–1.4 μm near-infrared window was much more effective [45]. One application of the transcranial optical vascular imaging (TOVI) in preclinical research is a mapping of the epileptic seizures with high spatial and temporal resolution (Fig. 3). As well as other optical imaging methods such studies has great significance for both fundamental research on epileptic neurons and the clinical management of epilepsy. Compared with X-ray, CT, PET and MRI, fluorescence transcranial optical methods have a serious advantage: it allows experimentalists to get information

Fig. 3 Functional Augmented Neurophotonics Brain Mapping; Schematic of TOVI equipment, which includes a standard fluorescent zoom microscope, and lamp and an emission filter. Images captured by a CCD camera are saved as a raw stacked 16-bit tiff files on a PC-based workstation. Enhanced fluorescent images are acquired by the same camera during a few second interval after fluorescent material administration, All of these methods of brain optical and others which allow us successfully visualization neural activity. Imaging and novel methods of the control of the neural network are based on the combination of many achievements in the different areas of science and technology. It is nontrivial, if not impossible task to use all of the in the same lab

60

V. Kalchenko et al.

with a wide palette of contrast [70]. Natural contrast objects present in many parts of the animal’s body including the brain. It is also possible to use natural fluorescence, or autofluorescence, due to the fluorescence properties different of some peptides (elastin, keratin), physiologically active amino acids (tryptophan, tyrosine) and phospholipids [71]. Extrinsic optical agents including fluorescence probes, increase possibilities offered by the optical imaging methods. Usually 2-Dimensional imaging methods have been employed to make a fluorescence image of the vasculature, but fluorescence contrast 3D–images can be obtained using optical tomography [72]. Surgical removal of the skin and bone (cranial window) is a very common way to obtain high quality data about organization of cortical functional map and local neurosvascular coupling. Besides that, the noninvasive optical imaging of brain activity is also possible in small animals. In vivo functional imaging of the cortical and even subcortical structures was successfully performed on mice and rats using 2-photon as well as wide field optical imaging [25, 34, 73]. In the case of wide filed, transcranial imaging, however, failed to detect individual neurons or capillary. In vivo transcranial imaging in combination with fluorescence dextrans permits the visualization of connections [34] with accurate matching of functional and structural maps [34]. Using 2-photon imaging it was shown that Ca-sensitive fluorescence protein mice permit robust transcranial wide field imaging of auditory cortex [74]. Recently transcranial and open-brain high-speed OCT imaging of the mouse brain in combination with an appropriate mathematical algorithm was successfully applied for the CBF monitoring using spectral and time domain optical coherence tomography [73, 75]. The most perspective method for in vivo transcranial imaging of the CBF is the hybrid method that combines LSCI, fluorescent angiography and intensity-huesaturation (IHS) color model for the data analysis [25]. This method enables fast and accurate assessment of the CBF in particular cortical area as well as in the whole brain using portable and relatively cheap hardware [25]. Modern neuroscience have a huge arsenal of the imaging methods, adequate for visualization seizures in animal model of the epilepsy: intrinsic optical imaging (IOS), voltage-sensitive dye imaging (VSDi), optical coherence tomography (OCT), diffuse optic tomography (DOT), photoacoustic imaging (PA), and many others. All of these methods can be combined with each other’s and with electrophysiological methods, but usually any research team can use only one or two methods. The combination of different methods can improve scientific productivity, but due to many reasons usually each lab has only one It will be logical to use a huge arsenal of our imaging methods: angled fluorescence laminar optical tomography (aFLOT) approach [76, 77], fluorescent imaging of the glucose substitute [14, 15] Optical Coherence Tomography (OCT) [6]. Equipment for IOS and VSDi are different, usually involved in different projects but can be mounted on the same platform. It can decrease the cost a lot. Electrophysiological equipment, including systems for EEG recordings, multi-unit and single-unit recordings, and electrical stimulation usually working parallel to the optical imaging system. Basic electrophysiological equipment can be in use in combination with different imaging systems, so the total cost will be decreased. The data analysis is

Transcranial Dynamic Fluorescence Imaging for the Study …

61

Fig. 4 Principal organization of the optical imaging translational research center

important, and maybe even the most labor intensive process [4, 78]. Frequently most of analysis are performed by senior students or postdocs who have knowledge and experience in MatLab programming and data processing. Usually, they study biology or even medicine and have limited knowledge and very little experience in data analysis. Translational research centers (Fig. 4) allow having a one key person, who will be responsible for data analysis. Such person will teach students and perform analysis of both optical imaging and EEG data much more effectively, for many projects.

5 Conclusion Functional brain optical imaging could play a key role in the creation of the bridge between morphology and functional status in this particular part of the brain, and therefore contribute to more accurate diagnostics of the epilepsy and improved efficacy of the therapy. Coupling brain optical imaging with measurements of disease biomarkers and adding behavioral neuroscience techniques is making early diagnosis more feasible. This paper has provided overview such game changing technologies, which will impact the future of image guided diagnostics and intervention using neurophotonics. Translational research combines basic research in physics, chemistry, neuroscience, biomedical engineering and clinical expertise to develop new therapeutic strategies as well as diagnostic methods. Acknowledgement The images used in the Figs. 1, 2 and 3 are obtained by Dr. Vyacheslav Kalchenko, Weizmann Institute of Science

62

V. Kalchenko et al.

References 1. Lenkov DN, Volnova AB, Pope ARD, Tsytsarev V (2013) Advantages and limitations of brain imaging methods in the research of absence epilepsy in humans and animal models. J Neurosci Methods 212(2) 2. Pascual-Marqui RD et al (2011) Assessing interactions in the brain with exact low-resolution electromagnetic tomography. Philos Trans R Soc A Math Phys Eng Sci 369(1952):3768–3784 3. Yoshimura M et al (2018) Hyperactivation of the frontal control network revealed by symptom provocation in obsessive-compulsive disorder using EEG microstate and sLORETA analyses. Neuropsychobiology 1–10 4. Tsytsarev V, Bernardelli C, Maslov KI (2012) Living brain optical imaging: technology, methods and applications. J Neurosci Neuroengineering 1(2):13 5. Tsytsarev V, Premachandra K, Takeshita D, Bahar S (2008) Imaging cortical electrical stimulation in vivo: fast intrinsic optical signal versus voltage-sensitive dyes. Opt Lett 33(9) 6. Tsytsarev V, Rao B, Maslov KI, Li L, Wang LV (2013) Photoacoustic and optical coherence tomography of epilepsy with high temporal and spatial resolution and dual optical contrasts. J Neurosci Methods 216(2) 7. Kandratavicius L et al (2014) Animal models of epilepsy: use and limitations. Neuropsychiatr Dis Treat 10:1693 8. Patel KS, Zhao M, Ma H, Schwartz TH (2013) Imaging preictal hemodynamic changes in neocortical epilepsy. Neurosurg Focus 34(4):E10 9. Ingram J et al (2014) Oxygen and seizure dynamics: I. experiments. J Neurophysiol 112(2):205– 212 10. Schwartz TH (2007) Neurovascular coupling and epilepsy: hemodynamic markers for localizing and predicting seizure onset. Epilepsy Curr 7(4):91–94 11. Boas DA, Dunn AK (2010) Laser speckle contrast imaging in biomedical optics. J Biomed Opt 15(1):011109 12. Dunn AK, Bolay H, Moskowitz MA, Boas DA (2001) Dynamic imaging of cerebral blood flow using laser speckle. J Cereb Blood Flow Metab 21(3):195–201 13. Yang H, Zhang T, Zhou J, Carney PR, Jiang H (2015) In vivo imaging of epileptic foci in rats using a miniature probe integrating diffuse optical tomography and electroencephalographic source localization. Epilepsia 56(1):94–100 14. Tsytsarev V, Maslov KI, Yao J, Parameswar AR, Demchenko AV, Wang LV (2012) In vivo imaging of epileptic activity using 2-NBDG, a fluorescent deoxyglucose analog. J Neurosci Methods 203(1) 15. Yao J et al (2013) Noninvasive photoacoustic computed tomography of mouse brain metabolism in vivo. Neuroimage 64(1) 16. Ma H et al (2014) Wide-field in vivo neocortical calcium dye imaging using a convectionenhanced loading technique combined with simultaneous multiwavelength imaging of voltagesensitive dyes and hemodynamic signals. Neurophotonics 1(1):015003 17. Chen-Bee CH, Kwon MC, Masino SA, Frostig RD (1996) Areal extent quantification of functional representations using intrinsic signal optical imaging. J Neurosci Methods 68(1):27–37 18. Tsytsarev V, Pope D, Pumbo E, Yablonskii A, Hofmann M (2010) Study of the cortical representation of whisker directional deflection using voltage-sensitive dye optical imaging. Neuroimage 53(1):233–238 19. Song Y et al (2016) Intraoperative optical mapping of epileptogenic cortices during non-ictal periods in pediatric patients. NeuroImage Clin 11:423–434 20. Abraham T, Feng J (2011) Evolution of brain imaging instrumentation. Semin Nucl Med 41(3):202–219 21. Valotassiou V, Wozniak G, Sifakis N, Demakopoulos N, Georgoulias P (2008) Radiopharmaceuticals in neurological and psychiatric disorders. Curr Clin Pharmacol 3(2):99–107

Transcranial Dynamic Fluorescence Imaging for the Study …

63

22. Fantini S, Sassaroli A, Tgavalekos KT, Kornbluth J (2016) Cerebral blood flow and autoregulation: current measurement techniques and prospects for noninvasive optical methods. Neurophotonics 3(3):031411 23. Thanos PK, Wang G-J, Volkow ND (2008) Positron emission tomography as a tool for studying alcohol abuse. Alcohol Res Health 31(3):233–237 24. Kim MN et al (2010) Noninvasive measurement of cerebral blood flow and blood oxygenation using near-infrared and diffuse correlation spectroscopies in critically brain-injured adults. Neurocrit Care 12(2):173–180 25. Kalchenko V, Israeli D, Kuznetsov Y, Harmelin A (2014) Transcranial optical vascular imaging (TOVI) of cortical hemodynamics in mouse brain. Sci Rep 4(1):5839 26. Kateb B, Yamamoto V, Yu C, Grundfest W, Gruen JP (2009) Infrared thermal imaging: a review of the literature and case report. Neuroimage 47:T154–T162 27. Kuo J-R, Chang M-H, Wang C-C, Chio C-C, Wang J-J, Lin B-S (2013) Wireless near-infrared spectroscopy system for determining brain hemoglobin levels in laboratory animals. J Neurosci Methods 214(2):204–209 28. Ma H, Zhao M, Suh M, Schwartz TH (2009) Hemodynamic surrogates for excitatory membrane potential change during interictal epileptiform events in rat neocortex. J Neurophysiol 101(5):2550–2562 29. Grandy TH, Greenfield SA, Devonshire IM (2012) An evaluation of in vivo voltage-sensitive dyes: pharmacological side effects and signal-to-noise ratios after effective removal of brainpulsation artifacts. J Neurophysiol 108(11):2931–2945 30. Devonshire IM, Dommett EJ, Grandy TH, Halliday AC, Greenfield SA (2010) Environmental enrichment differentially modifies specific components of sensory-evoked activity in rat barrel cortex as revealed by simultaneous electrophysiological recordings and optical imaging in vivo. Neuroscience 170(2):662–669 31. Ringuette D, Nauenberg J, Monnier PP, Carlen PL, Levi O (2018) Data compression and improved registration for laser speckle contrast imaging of rodent brains. Biomed Opt Express 9(11):5615–5634 32. Wang L, Li Y, Li Y, Li K (2018) Improved speckle contrast optical coherence tomography angiography. Am J Transl Res 10(10):3025–3035 33. Senarathna J, Rege A, Li N, Thakor NV (2013) Laser speckle contrast imaging: theory, instrumentation and applications. IEEE Rev Biomed Eng 6:99–110 34. Wang Z, Hughes S, Dayasundara S, Menon RS (2007) Theoretical and experimental optimization of laser speckle contrast imaging for high specificity to brain microcirculation. J Cereb Blood Flow Metab 27(2):258–269 35. Men J et al (2016) Optical coherence tomography for brain imaging and developmental biology. IEEE J Sel Top Quantum Electron 22(4):1–13 36. Roche-Labarbe N et al (2010) Noninvasive optical measures of CBV, StO2, CBF index, and rCMRO2 in human premature neonates’ brains in the first six weeks of life. Hum Brain Mapp 31(3):341–352 37. Roche-Labarbe N, Wallois F, Ponchel E, Kongolo G, Grebe R (2007) Coupled oxygenation oscillation measured by NIRS and intermittent cerebral activation on EEG in premature infants. Neuroimage 36(3):718–727 38. Hu S, Maslov K, Tsytsarev V, Wang LV (2009) Functional transcranial brain imaging by optical-resolution photoacoustic microscopy. J Biomed Opt 14(4) 39. Wang X, Pang Y, Ku G, Xie X, Stoica G, Wang LV (2003) Noninvasive laser-induced photoacoustic tomography for structural and functional in vivo imaging of the brain. Nat Biotechnol 21(7):803–806 40. Takahashi K, Hishida R, Kubota Y, Kudoh M, Takahashi S, Shibuki K (2006) Transcranial fluorescence imaging of auditory cortical plasticity regulated by acoustic environments in mice. Eur J Neurosci 23(5):1365–1376 41. Tohmi M, Takahashi K, Kubota Y, Hishida R, Shibuki K (2009) Transcranial flavoprotein fluorescence imaging of mouse cortical activity and plasticity. J Neurochem 109:3–9

64

V. Kalchenko et al.

42. Oh S et al (2011) In vivo optical properties of cortical tubers in children with tuberous sclerosis complex (TSC): a preliminary investigation. Epilepsia 52(9):1699–1704 43. Widjaja E, Wilkinson ID, Griffiths PD (2007) Magnetic resonance perfusion imaging in malformations of cortical development. Acta Radiol 48(8):907–917 44. Zenaro E, Rossi B, Angiari S, Constantin G (2013) Use of imaging to study leukocyte trafficking in the central nervous system. Immunol Cell Biol 91(4):271–280 45. Kalchenko V, Israeli D, Kuznetsov Y, Meglinski I, Harmelin A (2015) A simple approach for non-invasive transcranial optical vascular imaging (nTOVI). J Biophotonics 8(11–12):897–901 46. Guevara E, Pouliot P, Nguyen DK, Lesage F (2013) Optical imaging of acute epileptic networks in mice. J Biomed Opt 18(7):076021 47. Tsytsarev V, Arakawa H, Borisov S, Pumbo E, Erzurumlu RS, Papkovsky DB (2013) In vivo imaging of brain metabolism activity using a phosphorescent oxygen-sensitive probe. J Neurosci Methods 216(2) 48. Hong G et al (2014) Through-skull fluorescence imaging of the brain in a new near-infrared window. Nat Photonics 8(9):723–730 49. Gottschalk S, Fehm TF, Deán-Ben XL, Tsytsarev V, Razansky D (2017) Correlation between volumetric oxygenation responses and electrophysiology identifies deep thalamocortical activity during epileptic seizures. Neurophotonics 4(1) 50. Hamilton NB, Attwell D, Hall CN (2010) Pericyte-mediated regulation of capillary diameter: a component of neurovascular coupling in health and disease. Front Neuroenergetics 2 51. Avsenik J, Bisdas S, Popovic KS (2015) Blood-brain barrier permeability imaging using perfusion computed tomography. Radiol Oncol 49(2):107–114 52. Peppiatt CM, Howarth C, Mobbs P, Attwell D (2006) Bidirectional control of CNS capillary diameter by pericytes. Nature 443(7112):700–704 53. Hill RA, Tong L, Yuan P, Murikinati S, Gupta S, Grutzendler J (2015) Regional blood flow in the normal and ischemic brain is controlled by arteriolar smooth muscle cell contractility and not by capillary pericytes. Neuron 87(1):95–110 54. Hartmann MJZ (2009) Active touch, exploratory movements, and sensory prediction. Integr Comp Biol 49(6):681–690 55. Milesi S et al (2014) Redistribution of PDGFRβ cells and NG2DsRed pericytes at the cerebrovasculature after status epilepticus. Neurobiol Dis 71:151–158 56. Kovács R et al (2018) Bioenergetic mechanisms of seizure control. Front. Cell. Neurosci. 12:335 57. Leal-Campanario R et al (2017) Abnormal capillary vasodynamics contribute to ictal neurodegeneration in epilepsy. Sci Rep 7(1):43276 58. Nicastro N, Assal F, Seeck M (2016) From here to epilepsy: the risk of seizure in patients with Alzheimer’s disease. Epileptic Disord 18(1):1–12 59. Born HA (2015) Seizures in alzheimer’s disease. Neuroscience 286:251–263 60. Zhang X et al (2015) Near-infrared fluorescence molecular imaging of amyloid beta species and monitoring therapy in animal models of Alzheimer’s disease. Proc Natl Acad Sci 112(31):9734– 9739 61. Kodam A et al (2018) A role for astrocyte-derived amyloid β peptides in the degeneration of neurons in an animal model of temporal lobe epilepsy. Brain Pathol 62. Obrig H (2014) NIRS in clinical neurology—a ‘promising’ tool? Neuroimage 85(Pt 1):535–546 63. Ran C et al (2009) Design, synthesis, and testing of difluoroboron-derivatized curcumins as near-infrared probes for in vivo detection of amyloid-β deposits. J Am Chem Soc 131(42):15257–15261 64. Koronyo-Hamaoui M et al (2011) Identification of amyloid plaques in retinas from Alzheimer’s patients and noninvasive in vivo optical imaging of retinal plaques in a mouse model. Neuroimage 54:S204–S217 65. Kantarci K et al (2012) Ante mortem amyloid imaging and β-amyloid pathology in a case with dementia with Lewy bodies. Neurobiol Aging 33(5):878–885 66. J. Noebels (Jan. 2011) A perfect storm: converging paths of epilepsy and Alzheimer’s dementia intersect in the hippocampal formation. Epilepsia 52 Suppl 1(suppl 1):39–46

Transcranial Dynamic Fluorescence Imaging for the Study …

65

67. Cui M, Ono M, Watanabe H, Kimura H, Liu B, Saji H (2014) Smart near-infrared fluorescence probes with donor-acceptor structure for in vivo detection of β-amyloid deposits. J Am Chem Soc 136(9):3388–3394 68. Chang WM, Dakanali M, Capule CC, Sigurdson CJ, Yang J, Theodorakis EA (2011) ANCA: a family of fluorescent probes that bind and stain amyloid plaques in human tissue. ACS Chem Neurosci 2(5):249–255 69. Watanabe H, Ono M, Matsumura K, Yoshimura M, Kimura H, Saji H (2013) Molecular imaging of β-amyloid plaques with near-infrared boron dipyrromethane (BODIPY)-based fluorescent probes. Mol Imaging 12(5):338–347 70. Hillman EMC et al (2011) In vivo optical imaging and dynamic contrast methods for biomedical research. Philos Trans A Math Phys Eng Sci 369(1955):4620–4643 71. Hillman EMC, Boas DA, Dale AM, Dunn AK (2004) Laminar optical tomography: demonstration of millimeter-scale depth-resolved imaging in turbid media. Opt Lett 29(14):1650–1652 72. Erickson SJ, Martinez SL, DeCerce J, Romero A, Caldera L, Godavarty A (2013) Threedimensional fluorescence tomography of human breast tissues in vivo using a hand-held optical imager. Phys Med Biol 58(5):1563–1579 73. Bukowska D et al (2012) Multi-parametric imaging of murine brain using spectral and time domain optical coherence tomography. J Biomed Opt 17(10):101515 74. Issa JB, Haeffele BD, Agarwal A, Bergles DE, Young ED, Yue DT (2014) Multiscale optical Ca2 +imaging of tonal organization in mouse auditory cortex. Neuron 83(4):944–959 75. Baran U, Wang RK (2016) Review of optical coherence tomography based angiography in neuroscience. Neurophotonics 3(1):010902 76. Tang Q et al (2017) High-dynamic-range fluorescence laminar optical tomography (HDRFLOT). Biomed Opt Express 8(4):2124–2137 77. Tang Q et al. (2016) In vivo mesoscopic voltage-sensitive dye imaging of brain activation. Sci Rep 6 78. Liao L-D et al (2013) Neurovascular coupling: in vivo optical techniques for functional brain imaging. Biomed Eng Online 12(1):38

Vyacheslav Kalchenko Obtained MD (1993), internship in Neurology (1994) and Ph.D. (1998) from the Chita State Medical Academy (Russia). Till 2001 he worked in industry and academia as an adjunct lecturer of the Department of Biomedical Engineering at the Chita State Technical University (Russia). After a postdoc at the Weizmann Institute of Science (Israel) (2002–2004), he become a Head of In Vivo Optical Imaging Unit at the Department of Veterinary Resources. For the period 2004– 2020, the Unit extended from the small core imaging facility to the world-leading Optical Imaging and Translational Bioengineering Center having wide-range state of the art imaging equipment for small animal imaging and new biomedical instruments development. Since 2015 he is also a Senior Staff Scientist at the Weizmann Institute. His research interests are focused on the conjunction of photonics, imaging science, digital neurology, and artificial intelligence. He has more than 20 years of R&D managing experience in the industry and academia. He pioneered many advanced imaging techniques, among them: TOVI–Transcranial Optical Vascular Imaging and Multimodal Fluorescent and Laser Speckle Imaging. He is author and co-author over 100 research papers in peer-reviewed scientific journals, proceedings of international conferences, patents, and book chapters. He has over 100 presentations at the major international conferences and symposia, including about 35 invited lectures and plenary talks. He is a Fellow of Royal Microscopical Society and Senior Member of SPIE and Honorary Professor of Chita State Medical Academy.

66

V. Kalchenko et al.

Vassiliy Tsytsarev holds a Ph.D. in Neuroscience from Saint Petersburg State University, Russia. Soon after graduation he moved to Japan and began working at the Brain Science Institute of RIKEN, and the Human Brain Research Center, Kyoto. Functional brain mapping, neural circuits and different types of brain optical imaging are his main scientific interests. In Japan, Vassiliy worked in the field of auditory neuroscience using intrinsic optical imaging (IOS) and voltagesensitive dye imaging. After seven years in Japan he moved to the United States, where he has worked at several universities; for the past six years, at the University of Maryland School. His current focus is on functional brain mapping, epileptic studies and neural network function in the rodent somatosensory system, which offers a perfect specimen for many types of neuroscience research, including models of neural diseases. Vassiliy is the author and co-author of more than 40 publications in peer-reviewed magazines, and several book chapters. He is a senior editor for the Journal of Neuroscience and Neuroengineering, serves on the board of directors of the Society for Brain Mapping and Therapeutics (SBMT), and on the editorial boards of other scientific journals.

Critical Elements for Connectivity Analysis of Brain Networks Jean Faber, Priscila C. Antoneli, Noemi S. Araújo, Daniel J. L. L. Pinheiro, and Esper Cavalheiro

1 Introduction In recent years, new and important perspectives were introduced in the field of neuroimaging with the emergence of the connectionist approach [1]. In this new context, it is important to know not only which brain areas are activated by a particular stimulus but, mainly, how these areas are structurally and functionally connected, distributed, and organized in relation to other areas. In addition, the arrangement of the network elements, i.e., its topology, and the dynamics they give rise to are also important. This new approach is called connectomics [2]. It brings together a series of techniques and methodologies capable of systematizing, from the different types of signals and images of the nervous system, how neuronal units to brain areas are connected. Through this approach, the different patterns of connectivity can be graphically and mathematically represented by the so-called connectomes [3]. The connectome uses quantitative metrics to evaluate structural and functional information from images of neural tracts and pathways or signals from the metabolic and/or electrophysiologic activity of cell populations or brain areas. Besides, with adequate treatment of this information, it is also possible to infer causal relationships. In this way, structural and functional evaluations are complementary descriptions which, together, represent the anatomic and physiologic neural properties, establishing a new paradigm for understanding how the brain functions by looking at brain connections [4–6]. J. Faber (B) · P. C. Antoneli · N. S. Araújo · D. J. L. L. Pinheiro · E. Cavalheiro Escola Paulista de Medicina (EPM), Department of Neurology and Neurosurgery, Discipline of Neuroscience, Lab. of Neuroengineering and Neurocognition, Universidade Federal de São Paulo—UNIFESP, São Paulo, Brazil e-mail: [email protected] E. Cavalheiro Centro Nacional de Pesquisa em Energia e Materiais CNPEM, Campinas, Brazil © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_4

67

68

J. Faber et al.

A connectionist approach allows us to evaluate how the anatomic organization of the brain relates to its functional dynamics and how structural or functional changes affect this relationship [7]. In order to perform a formal and quantitative analysis, Graph Theory is used in connectomics [8]. This method allows a systematic, consistent, and robust evaluation of the functional and structural neural networks. In this way, since the connectome incorporates all the mathematical properties of graphs, it naturally quantifies all the properties, similarities and differences among the different neural network configurations. Here, we highlight five critical elements of a network that allows an integrative analysis, focusing mainly on a functional description. These elements include; (i) the properties of its nodes; (ii) the metrics for connectivity and coupling between nodes; (iii) the network topologies; (iv) the network dynamics and (v) the interconnections between different domains and scales of network representations. The first element we must consider is the set of intrinsic properties of the nodes that comprise a network. When considering networks at the microscopic level, it can include, the type of neurons that are connected and the ways the neurons are activated (including metabolic and electrophysiological activities; [3]). At the mesoscopic level, the anatomical circuit properties matter, as well as the specific physiological signatures and activity patterns of each connected brain layer or subfield. Finally, at the macroscopic level, attention needs to be paid to all features associated with the anatomical composition (such as neuronal density, neural subfields, tracts and arrangement within each brain area). The second network element to highlight is the type of metric used to assess the connections or couplings between nodes [9]. Different metrics can be used to evaluate the connectivity in a brain network through quantification of the statistical dependencies, or even causal interactions, between node activities. Each one may reveal a linear or nonlinear relationship or describe a directed or undirected information flow through the network edges, defined according to what is being measured and studied. Here, we will explore three nonlinear metrics, Mutual Information, Kullback–Leibler, and Granger Causality, and one linear metric, given by the Pearson Coefficient. Furthermore, we will also discuss possible ways of coupling among the main features of the electrophysiological signal, such as amplitude, frequency, and phase [10]. The third element to be considered is the arrangement of the network nodes and vertices, i.e., topology [11]. The topology of a brain network is one of the most important aspects of its connectome. It may reveal how a particular neural activity, or even a brain area, is relevant for a specific brain state associated with a disease or stimulus, for instance, and also how a particular neural arrangement optimizes the flow of information in a complex brain network [12]. Furthermore, it also describes the importance of network hubs and assemblies by quantifying their interconnections and information storage [13]. The fourth element of a network is related to its change over time. The dynamic activity of the brain promotes different synchronizations among different areas and subfields at different periods of time [2]. Mainly, it is of interest to evaluate temporal phase transitions of neural activity associated with a particular cognitive state during

Critical Elements for Connectivity Analysis of Brain Networks

69

a behavioral task or associated with a neurocognitive disorder. For example, the phase transitions related to an epileptic seizure can be described by topological changes of a network over time [14]. Therefore, a connectome analysis can inform how and why a neural network fluctuates, repeats, reorganizes, stabilizes, or degenerates in time. Finally, the fifth network element addressed here is related to the most intricate brain characteristic: the way different levels of information, from biomolecules to brain areas, are integrated. From the connectome perspective, this problem might be assessed by inspecting how different networks, described at different scales, can be interconnected, and how the information storage and processing at one scale level interferes with them at another [15]. We are still far from having an answer on this issue, but the connectome approach allows us to propose mathematical models of integration yielding an objective formulation with a possible test of its consistency. Through a functional and effective connectivity analysis, the type of technique used to measure brain activities is fundamentally relevant since it defines the type, the scale, and the node features of a network. For instance, techniques of invasive electrophysiological recordings can register the activity of individual cells, such as action potentials and spike trains, or the activities of groups of cells, such as local field potentials [16]. Noninvasive techniques, like electroencephalography (EEG), functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG) and functional near-infrared spectrum (fNIRS), also allow a connectivity analysis at a large scale. In general, EEG recordings provide a description of the overall activity of the encephalon [17, 18]. But, despite lacking anatomical and/or physiological specificities, they can help to determine how certain cognitive or pathological mental states are associated with specific network topologies. Clinically, the connectome approach can be extremely powerful because it provides the functional and structural topologies, quantitative parameters of the brain’s activity that, in general, are not accessible using only traditional brain images [19]. For example, a functional/effective connectivity analysis can help to provide information about the stability or dysfunctionality of certain neural subnetworks associated with a brain disorder such as dementia, epilepsy, or Parkinson. It also allows for an evaluation of information flow in a brain area surrounded by a tumor or around an epileptic focus [20]. Another example of how connectome approach can help us to understand acute diseases or comorbidities is to study the evolution of the connection patterns over time. In addition, it brings a complete new look over the neural dynamics, addressing a truly integrated brain and not only its associated parts [21].

70

J. Faber et al.

2 Brain Networks 2.1 Graph Representation Network science is an interdisciplinary field that combines concepts and techniques from computational sciences, statistics, engineering, and mathematics among others [22]. Through these techniques, it is possible to construct graphical models that allow for a quantitative and systematic description of how the neural systems interact, organize themselves in different geometric patterns, evolve in time and stabilize to optimize the storage, flow, and processing of information. Additionally, these models allow for statistical inferences providing evaluation and visualization of the communication process among its units along time and space and with other networks in different scales. When network analyses are applied to brain circuits, they provide robust methods to forecast structural and functional brain changes associated with specific injuries or therapeutic interventions [23, 24]. This is called Connectomics [25]. Connectomics is a new approach that attempts to provide a solid road for the studies of connectomes, including the diversity of neural connectivity maps in different scales of time and space [2]. In a general way, connectome descriptions are based on Graph Theory [26]. More formally, Graph Theory is a theoretical field dedicated to the study of mathematical structures, called graphs, used to model pairwise relations between information units. As a mathematical object, a graph G can be defined by the relationship between the pair of sets for vertices V and edges E, i.e. G = {V , E}. Edges are also known as arcs or lines and can represent the mode, type, and intensity of link between pairs of vertices. The mode refers to the representation of the relationship between vertices, since, for instance, a graph might be displayed in 2 or 3 dimensions. The type refers to the direction of connections, i.e., undirected versus directed. Finally, one must consider the intensity that relates the strength between connections. Vertices are also known as dots or, more commonly, nodes [27, 28]. Typically, a graph is represented in a schematic geometric form composed of a set of nodes graphically represented by points joined by lines or curves. The latter can be directed or undirected when intended to represent the information flow, or it can be of different thicknesses when intended to represent the degree, cost, or connection probability between the nodes [29]. Historically, Graph Theory was introduced by Euler in 1735 when he proposed a problem known as “the seven bridges of Königsberg” [30]. In the German city of Königsberg, now the territory of Kaliningrad, Russia, there were seven bridges arranged in a very particular way as shown in Fig. 1. Euler asked if it would be possible to make a path passing through all of them by crossing each bridge once. By formalizing mathematically the idea of graphs, Euler showed that this solution was impossible [31]. A graph can be expressed using matrix notations, by means of the so-called adjacency matrices and incidence matrices, that contains information about the intensity

Critical Elements for Connectivity Analysis of Brain Networks

a

71

b

Fig. 1 The seven bridges of Krönigsberg. a An old map of Krönigsberg with the bridge’s configuration. b Graph representation of this map

and direction of the connections between nodes. For example, given a network with N = 4 nodes, {x, y, z, w}, the matrix representation, and its corresponding graphs are: Now, considering the same nodes set {x, y, z, w}, we may also represent the intensity of connections by adding weights to each edge, which will be given by the values obtained from the correlation metric like those shown in Fig. 2b. Later, we will discuss, in more detail, different metrics and approaches to quantify statistical dependencies, correlations and couplings between nodes and their correspondent intensities. However, for symmetrical connections given by symmetrical coupling models or metrics of statistical dependence, the edges are undirected, i.e., the links do not include arrows. These networks are represented by undirected graphs and square symmetric matrices as all the edges are bidirectional. Graph representations enable the characterization of network patterns by evaluating how nodes are connected to each other in order to display specific topological structures. Undirected graphs, for example, are well designed for systems that incorporate symmetric coupling interactions and symmetric metrics of correlations/associations. In contrast, if one pretends to represent the flow of information between two brain areas, a directed graph describing asymmetrical connections should be considered. In order to approach the diversity of possible connections and contact-points, graph representations can be displayed in different colors, styles, and sizes, as shown in Fig. 2b. The different intensities of connections can be illustrated by an index representing the strength nodes interaction, for example, using a weighted graph with different sizes of nodes and edges. It is also possible to represent graphs with multiple edges between the same pair of nodes and, also, nodes with self-connections as illustrated in Fig. 2c. The nonlinear and multiscale of brain dynamics are implicated in a diversity of physical couplings and statistical dependencies that are usually classified into three types: (i) functional connectivity, (ii) effective or causal connectivity, and (iii) structural connectivity [12, 32].

72

J. Faber et al.

a

b

c Fig. 2 General matrix and visual representations for different type of graphs. a Matrix and visual representation for non-weighted graphs. Figure shows a visual representation of the adjacency matrix M, where each node is represented by a circle and each vertex by a directed arrow. In this representation, the vertices directions are read from row to column. (A1) Shows the number of vertices from each node for the same graph. (A2 and A3) Exhibits a graph representation highlighting the nodes with more confluences. b Matrix and visual representation for weighted directed graphs. Visual representation for the distance matrix D, where each node is represented by a circle and each vertex by a directed arrow. In this representation, the vertices directions are read from row to column. (B1) shows the number of vertices from each node for the same graph. (B2) exhibit a graph representation emphasizing the nodes with more confluences. (C) Matrix and visual representation for undirected graphs with self-connections. It shows a visual representation for the symmetric matrix U, where each node is represented by a circle and each vertex by an undirected arrow (or nodes by circles and vertices by undirected arrows)

Scans of MRI and fMRI configure the most common measurements to describe patterns of brain connections [12]. However, when considering evaluations of functional connectivity, other brain signals can also provide useful information to describe neural network configurations. Actually, by evaluating functional connectivity we gain a totally different perspective of brain dynamics since it allows a detailed look on how the brain uses different ways to signalize and processes information [33, 34]. Currently, there are different methods for recording brain signals of different physical nature and spatiotemporal scales. Besides the fMRI technique, that essentially measures variations of blood flow associated with neural activity [35], during cognitive and/or behavioral tasks [36, 37], other non-invasive techniques such as EEG, magnetoencephalography (MEG) or functional near-infrared spectrum (fNIRS) also provide information from the brain activity [38, 39]. Since these techniques are not invasive, most of the networks yielded by these signals are related to global

Critical Elements for Connectivity Analysis of Brain Networks

73

aspects of brain dynamics and functions. There are some computational approaches, which typically use multivariate statistics, and that allow one to infer structural signal sources from deep brain areas through EEG recordings [40]. Although these approaches allow for the reconstruction of some deep brain areas and the generation of useful 3D functional network pictures, they lack the spatial accuracy of fMRI scans because they are statistical estimations [41, 42]. Other invasive techniques are also used to assess the electrophysiological activity of deep brain regions, such as the hippocampus, thalamus, or cerebellum, among others. This is usually done during (pre) surgical procedures, chronically implanted in patients with deep brain stimulators (DBS), animal models or voluntary subjects from brain-computer interface projects [43–45]. To calculate adjacency/incidence matrices reflecting functional or effective connectivity, pairs of time series are “linked” by a mathematical metric able to capture a statistical dependence (causal or not) between those sources of brain activity. In this way, the network edges in a functional connectivity description are labeled by numerical values measured by a specific statistical dependence metric that expresses a linear or nonlinear, symmetric or asymmetric correlation/association; and each network node represents the source of a measured brain signal.

2.2 Brain Network Nodes The specification of a node depends on what we want to know in order to select and measure a specific biological/physical feature that will be used to perform the quantitative analysis. Consider, for instance, a culture of neurons; we can construct a structural connectivity network by considering single neurons as nodes and all related structural connections among other neurons through axons and dendrites as the edges. To measure the nodes and edges in this network, we can apply biomolecular and histological techniques to mark and identify all the neurons and their respective physical connections [46]. Considering another scale, different brain regions can be considered as nodes in a structural representation. These regions can be the visual cortex, thalamus, hippocampus, motor cortex, etc., and the edges can be the neural pathways such as nerve fibers and tracts. In this case, the best approach to measure them is tractography which comprises a 3D imaging modeling technique that represents neural tracts by using diffusion-weighted images (DWI) recorded from MRI in parallel to computer-based image analysis [47]. However, to evaluate and construct functional networks the possibilities are even higher. A functional connectivity analysis will be based on the type of feature or signal being measured from different neural structures. Figure 3 provides a short list of the main biological structures considered as nodes in structural network analysis and their possible corresponding features to be measured by means of different techniques. In summary, there are many different brain (bio)physical quantities that can be measured with a technique that reads one or more features and these measures can be

74

J. Faber et al.

Fig. 3 General framework for different type of structural and functional nodes. The first column of the table presents a series of different possible structural nodes, according to its biological nature, information, and spatial scale. From the top to down it is listed as a possible network node: (I) a gene, (II) biomolecular, membrane and synapses, (III) cell unity (neurons and astrocytes), (IV) subfields or layers of a specific brain region, such as CA1, CA2, CA3, and CA4 of hippocampus, and (V) brain regions, such as motor cortex, visual cortex, thalamus and etc. The second column lists the main biological features evaluated in each correspondent structural node. These features are chosen according to the scientific field of investigation and each one can represent a functional node. The third column lists the main techniques used to measure the associated features of each structural node

characterized and compared using a wide variety of metrics. For instance, considering again a culture of cells or sliced brain tissues recorded in an in vitro procedure. We can record the extracellular potentials of the electrophysiological activity from particular populations of neurons or from a specific brain subfield [48]. Therefore, the individual neural activity recorded from each electrode during a right period can be considered a node in a functional network and the edges, some possible statistical dependency among them. As it will be described later, the statistical dependencies will be totally determined by a mathematical metric and the signal feature (phase, frequency, amplitude, time, etc.) being considered (Fig. 4). Under the Graph Theory point of view, a node is a redistribution point or a communication endpoint. A node is defined as an active system attached to a network, capable of creating, receiving, or transmitting information over a communication channel known as an edge in this case [29]. Any passive distribution point, such as a distribution frame or patch panel in computer networks, is consequently not a node.

Critical Elements for Connectivity Analysis of Brain Networks

75

a

c

b

Fig. 4 Different kind of nodes and correlation metrics in structural and functional networks. a In vitro culture of neurons on a multielectrode array (MEA) to record electrophysiological extracellular activity and two examples of two different topological configurations for a functional connectivity analysis. b A signal template cut from an electrophysiological recording is selected and two statistical features were calculated, amplitude histogram and power spectral density. Each one of these features, and any other, can be interpreted as a functional node of a network. c By means of different mathematical metrics (linear and nonlinear), it is possible to establish how all nodes are correlated. It is shown four possible metrics: KL(P|Q)—Kullback–Leibler or Divergent Entropy, I(X,Y)—Mutual Information, KS—Kolmogorov-Smirnov and ρ XY —Pearson’s Coefficient

A node can be also be classified according to their trespassed information flow as a place in a network where a message can be created (called source or server), received (sink or client), or transmitted (repeaters or peers), see Fig. 5. A peer node sometimes may work as a client- or a server-node [49]. Furthermore, a peer-to-peer node or an overlay network that actively routes data to other network structures in a different spatial or temporal scale can be called supernodes [50]. Others information associated with network nodes are listed in Table 1. To characterize the entire network, it is crucial to know the characteristics of a node in a network. It helps to define its topology which determines its efficiency in transmitting and storing information. An important concept is the modular nodes (Fig. 5f) characterized by clusters of nodes or nodes more densely connected to other nodes [11]. These special nodes in neural circuits or structural networks may represent the nucleus with a specialized

a

b

c

d

e

f

Fig. 5 General scheme of types of nodes. a A source or client node where all point out from the node to other nodes. This type of node represents the creation of information. b A peer or repeater node, c a terminal or isolate node, d a sink or client node and e a self-reference node and f modular nodes

76 Table 1 Node information

J. Faber et al. Degree

Number of edges associated with a node

Nearest

Nearest neighbors within radius

Indegree

In-degree: number of oriented edges entering the node

Outdegree

Out-degree: number of oriented edges incoming the node

Predecessors Number of predecessors in an oriented network Successors

Number of successors in an oriented network

function, where different modules may work in parallel to support different neurophysiological processes [13]. In addition, a network having a distributed “core-toperiphery” configuration has a set of central nodes that are interconnected with all other nodes in the network and a set of nodes on the periphery that are sparsely interconnected with all other nodes in the network [11]. This type of network architecture represents a process of information integration through neural assemblies, neural circuits, or functional nuclei, characterizing a control point [6, 51]. All these characterizations of edges and nodes are critical and configure the first elements of a network and provide the main base to any connectivity analysis.

3 Measures of Connectivity 3.1 Couplings and Correlations After defining the graph nodes and what they represent, it is necessary to quantify the interactions between them. These interactions are measures that denote the information provided by the graph, such as its topology, architecture, and complexity. Essentially, the edges indicate the nodes that are currently linked and the strength of this link when the graph is weighted. Thus, a graph can be symmetrical or not depending on the metrics used to calculate its connectivity. Besides structural connectivity, one can also consider functional and effective connectivity. Functional connections refer to any form of statistical dependence between the activities of two nodes [52], in general, without any assumption of causal influences. A statistical dependence between two variables can be defined in terms of Bayes’s rule. This rule states that two variables X and Y are dependent when, at least, the probability to get one of the outcomes from either of the two variables depends on whether we conditioned it to some knowledge on obtained outcomes from the other [53]. Mathematically, this can be expressed as P(X|Y) = P(X), where P(X|Y) = P(X,Y)/P(Y) refers to the probability distribution of X when its measure is conditioned to some knowledge from the outcomes of Y, and P(X,Y) is the joint probability distribution of X and Y. Inversely, X and Y are independent when P(X|Y) = P(X). The same statements apply if one exchanges X with Y in these expressions.

Critical Elements for Connectivity Analysis of Brain Networks

77

This rule represents the more general form to measure a relationship between two random variables. Correlations are special cases of statistical dependencies, where we consider them as a mathematical metric that measures an increasing or a decreasing trend, linear or nonlinear, parametric or non-parametric. We say two variables X and Y are correlated with increasing trend when the values of Y increase according to the positive increase of X [53]. Similarly, a correlation with a decreasing trend occurs when the values of Y decrease according to the positive increase of X. Although a correlation metric can be linear or nonlinear, linear metrics are more commonly used in the literature, like the Pearson’s correlation coefficient. In this way, functional connectivity evaluates only statistical dependencies of node outputs (Fig. 6a). On the other hand, there are other types of connections that refer to causal influences between nodes [52]. It means that the activity of one node X directly influences a

b

Fig. 6 Differences between functional and effective interaction. a When there are only output signals from two or three systems it is difficult to infer causality and correlations can be more adequate to describe interaction among these systems. The measures of the output from a system in specific time windowing, for instance, only represents a statistical dependence between the variables been evaluated, performed through linear or non-linear correlation metrics. b When there is a mathematical model or experimental protocol that supports the manipulation of inputs of a system in function to its outputs and outputs of other systems, it is possible to analyze their effective interactions, inferring causality

78

J. Faber et al.

the activity of the other connected node Y. Thus, node links are directionally represented with an arrow, indicating the direction of information flow from the source node to the receiver node. They can also be bidirectional, where both nodes are reciprocally coupled [54]. In connectivity analysis, two main definitions of causality to approach this situation are used. The first one is based on control theory, where the input of one node Y is influenced by the output of another node X. Correlations cannot infer causality since it impossible to determine if the statistical dependence being evaluated is from one of the two nodes or from a third node, or if it occurs by chance (see Fig. 6a). Any causal influence between any pair of nodes could happen not only directly between them but also through a third (or more) intermediate one. In this case, a node X is influenced by an intermediate node Z (or more) which is influenced by another node Y (Fig. 6b). By perturbing the system intentionally and observing the effects of the perturbation, as already mentioned, it is possible to evaluate possible causality effects. In this case, we can, for example, block chemically or remove a tract, subregion, or nerve that connects two nodes and see if the observed statistical dependencies between their activities are maintained or modified. The second definition is commonly used to describe causal influences between two nodes by temporal influences of part of an output X onto an output Y [55–57]. This definition reads [58] “a signal X is said to be the cause of a second signal Y when the information about the past of X helps to determine the presence and/or future states of Y beyond and above the information from only the past of Y ”, [54]. In this case, it is necessary to have a mathematical model that describes this temporal influence. There are some disagreements on the use of these approaches as real metrics for causality measurements because they can only provide evaluations of directed functional connectivity since they are defined in terms of time-lagged statistical dependencies [59], Fig. 7. Therefore, by using these metrics, it is possible to establish a well-defined criterion to describe edges among different nodes in a network. In this way, once a correlation or a coupling metric that determine all node links is chosen, it is necessary to verify if there are spurious correlations. For this procedure, different methods can be applied as thresholds using surrogate or baseline signals, for example, in order to decide statistically significant network edges. In this way, by using the matrix representation of a graph, there are two possibilities: (i) the edges can be ‘digitized’ with those edges bigger than the threshold valued as 1 s and other edges valued as 0 s, or (ii) they can be weighted with continuous or discrete values for each edge [60].

3.2 Functional Connectivity 3.2.1

Undirected Metrics

The first aspect of undirected metrics is the equilibrate flow of information between two nodes. It means that (i) the physical interaction between two nodes is totally

Critical Elements for Connectivity Analysis of Brain Networks

79

Fig. 7 Diagram of temporal causality and Transfer Entropy. The definition of causality by statistical dependencies assuming temporal lags between two signals. In this definition, there is a time-windowing t − τ in both signals that cause the variations in y(t). On the right side, there is Venn diagram showing the relationship among each variable and their intersection. As case of a possible temporal causality, the Transfer Entropy metric is one of the most used in the scientific literature

symmetrical and static without flow of information from one to another node [61]; in this case, any metric provides a good measurement of the relationship between nodes; or (ii) there is an information flow from one node to another, but we use a symmetric statistical metric, unable to describe/detect the asymmetry between nodes. As described in Fig. 2, this type of edge is represented by a symmetrical adjacent matrix [62]. A commonly used metric for this type of link is a measure of correlation given by the parametric Pearson’s coefficient that calculates the ratio of covariance between two variables X and Y normalized by the square root of the product of their own variance: ρ=

COV (X , Y )  σx2 σy2

(1)

COV (X,Y ) represents the covariance of the variables X with Y, and σx and σy are their respective standard deviation. The coefficient ρ ranges between −1 and 1 and depends only on the spread of X and Y, capturing only their linear correlation. In the context of connectivity, it has been applied to evaluate the degree of linear correlation of a signal’s amplitude among different nodes [63]. A non-parametric correlation metric is also possible, such as Kendal’s and Spearman coefficient [60, 64, 65]. Analogously, the coherence is also a symmetric method that evaluates the spectral correlation of two signals X(t) and Y (t) in the frequency domain [66]:

80

J. Faber et al.

Pxy (f ) Cxy (f ) =  Pxx (f )Pyy (f )

(2)

where Pxx (f ), Pyy (f ) are respectively the power spectrum of X(t) and Y (t), and Pxy (f ) is the cross-spectrum between them. The spectrum can be calculated by applying the Fourier or Wavelet transforms on the signals [67]. The magnitude of coherence Cxy (f ) can be normalized to values between 0 and 1, representing the intensity of the correlation power at specific frequencies [68]. However, in order to use coherence as an interaction coefficient between two nodes, it is necessary to have an operation to summarize an index that represents the general coherence spectrum between X and Y, such as the magnitude of total coherence spectrum. It occurs because a coherence spectrum cannot be described by only one point, invalidating a network edge representation [69, 70]. The neural oscillations show specific patterns that are sometimes evidenced by the increasing or decreasing power along specific frequency bands. These fluctuations can be explored as relevant information to establish functional connectivity between nodes. In this way, spectral coherence can be used to detect aspects of phase synchronization among specific rhythms of the signals X(t) and Y (t), which may provide information on the communication dynamics among neurons [71]. Some researchers applied this technique to study functional networks using EEG, with electrodes placed on the scalp of a patient with a neurological disease, in order to detect its influence on the performance of different cognitive and behavioral tasks [72]. Spectral coherence measures an important effect of oscillations, when two or more signals have the same phase difference at a given frequency since most of the neural communication is directly related to the phase relationship of neural populations [73]. Another way to calculate symmetrical nonlinear interaction between two nodes, X and Y, is by using the mutual information between them calculated through I(X,Y ). This technique comes from the Information Theory and uses the concept of Shannon’s entropy to evaluate the information shared by two or more random variables [74]. Considering X and Y as two random variables with specific states, {x 1 , x 2 , x 3 ,…, x n } and {y1 , y2 , y3 ,…, yn } associated with their probability distributions {p(x 1 ), p(x 2 ), p(x 3 ),…, p(x n )} and {p(y1 ), p(y2 ), p(y3 ),…, p(yn )}, the mutual information is defined as:      p xi , yj   p xi , yj log I (X , Y ) = (3) p(xi )p yj i j   values associated with a right state x i where p(xi ) and p yj are the probability  and yj , respectively, and p xi , yj is the conditional probability between these two states. Considering the definition of Shannon’s entropy being the expected value of the amount of information given by some random variable, the mutual information I(X,Y) represents the average amount of information shared by two systems X and Y

Critical Elements for Connectivity Analysis of Brain Networks

81

[75]. It is important to emphasize that this metric can be used for any physical or statistical feature associated with the signals X and Y, but its representativeness is directly affected by the empirical probability distribution created to stipulate its information [76]. One of the advantages of using I(X,Y) is that, since it is not a linear function, it allows generalized measurements for symmetrical statistical dependencies between two or more random variables [9].

3.2.2

Directed Metrics

When the interaction of any two nodes X and Y presents a privileged pathway of information flow and unsatisfied one of the two previous conditions of symmetry, the use of directional metrics is preferable to represent the network links. In general, these metrics aim to capture a statistical clue of causation or, at least, a direction of the information dynamics, considering the different degree of dependence between nodes X and Y [9]. Here, we will present three mathematical metrics that evaluate directed interaction between nodes: Kullback–Leibler (KL), Transfer Entropy (TE) and Phase Slope Index (PSI). However, it is important to mention that there are many other, more or less, adequate metrics according to what is intended to describe. The Divergence of Kullback–Leibler, also known as relative entropy, is based on the concepts of information theory which can be roughly interpreted as a measure of the cost to turn a right probability distribution, P(X), into another, Q(X), under the same set of states or alphabet. Similarly, we may ask which probability distribution, P(X) or Q(X), will minimize the number of bits used to represent all the states of the random variable X = {x 1 , x 2 , x 3 ,…, x n }? [77, 78]. Kullback–Leibler is, therefore, an asymmetrical measurement defined as: DKL (P|Q) = −



p(xi ) log

i

p(xi ) q(xi )

(4)

where p(xi ) and q(xi ) represent the probability values from P(X) and Q(X), respectively. In a general way, the main objective to use directional metrics is to describe, in some way, statistical causations. As described previously, Transfer Entropy (TE) can be a metric that captures possible temporal causations between two signals (Fig. 7). It appears as a new approach to contrast with the time delayed mutual information [79], measuring the information transferred between to random process, X(t) and Y (t), considering part of their past and current states [80]. Mathematically it is defined as: TE(X → Y ) =



p(yn+1 , yn(k) , xn(l) ) log

p(yn+1 |yn(k) , xn(l) ) p(yn+1 |yn(k) )

(5)

where yn+1 it is a future state of Y; yn(k) is an vector of k previous possible states of Y (yn(k) = (y1 , y2 , . . . , yn−k+1 )); xn(l) is l previous states of X with the minimum of 1

82

J. Faber et al.

and maximum of k (xn(l) = (x1 , x2 , . . . , xl )). TE can, therefore, represent the directed information flow from X(t) to Y (t), or it can be interpreted as a degree of dependence between X and Y [81]. Although Wiener had initially used the concept of causality to interpret TE, it has currently been claimed that TE is an approach to quantify the predictive information flow [82]. Finally, an alternative technique to infer asymmetrical information flow from two signals is the phase slope index (PSI). PSI is also a nonlinear metric designed to measure the frequency-average of the slope phase of the spectral coherence [83]. As any interaction requires time and different interactions have, in general, different communication speeds between the sender and receiver, the phase difference between the sent and received messages should be assessed by the frequency. It means that PSI can detect positive or negative slopes on phase frequency-range that indicates the direction of information flow. If this relationship is negative, then the information flow occurs in the opposite direction [84, 85]. The PSI is defined as: ⎛ ⎞  Cxy (f )Cxy (f + δf )⎠ Ψ˜ ij = ζ ⎝

(6)

f ∈F

where ζ represents the use of just the imaginary part of the complex number; P Cxy (f ) = √ xy the spectrum coherence between X and Y and δ represents Pxx (f )Pyy (f )

the frequency resolution of the spectrum. Therefore, PSI estimates the degree of coherent communication between two or more nodes. Worthy of emphasis is that if X has impact on Y, it does not imply that Y has no impact on X.

3.3 Effective Connectivity As already discussed, in different references, a measure of correlation does not imply a measure of causality. This relationship is at the heart of the difference between the concept of functional connectivity and that of effective connectivity. It is not the purpose of this chapter to describe all the metrics of effective or structural connectivity, but to discuss some aspects of effective connectivity. The most common applications of effective connectivity are found in macroscopic networks through EEG, MEG, and fMRI recordings, although there is no formal restriction for using this approach [59]. Effective connectivity is an alternative measure or extension of functional connectivity; thus, one can also minimally infer causal relationships between them [52]. The techniques used in the characterization of effective connectivity are extremely general and allow a description of connection at any scale of time and space in the neural domain.

Critical Elements for Connectivity Analysis of Brain Networks

83

There are currently two main mathematical approaches to describe effective connectivity between two or more nodes: The Granger Causality Modeling (GCM) and the Dynamic Causal Modeling (DCM) [86–88]. The technique for measuring effective connectivity described by GCM was described by Granger in 1969 as a particular case of the Wiener definition given in 1956. Subsequently, other variations emerged [54, 55, 57]. The causality of Granger considers only information shared by linear interactions between the signals being evaluated. Technically, the GCM uses linear multivariate autoregressive models (MVARs) from a discrete set of differential equations [89, 90]. The GCM metric has been widely used in different approaches; however, caution should be taken when interpreting results because the technique requires many restrictions and conditions. The first of these is independence from random fluctuations in signal relative to past events. The second is inherent in the linearity of its expression. Dynamic Causal Modeling (DCM) [88] can be understood as a more general measure than GCM because it makes less demands on the description of signal interaction models. In addition to using a continuous time formulation, it can also use a bilinear approximation (or Taylor expansion) in its interaction model [91]. The equations described in the DCM interaction models can also be associated with a second set of time-independent equations, which map the variable X, associated with neural states, into recorded signals such as EEG, MEG or BOLD. From this projection, a generalization is made about the linearities imposed by the model, thus overcoming the restrictions found in the GCM. However, a level of caution is also important since these same projections may also insert false positives about the relationships between two nodes [59, 90, 91].

4 Graph Measures Once nodes and the connections are defined and characterized, we can use mathematical metrics to quantify network properties that define how the graph elements interact and how they are organized in time and space, forming complex mathematical structures and, finally, how these structures express storage, processing, transmission and organization of the neural information. There are two main types of measures that characterize a graph; topological and geometric. Topological measures quantify the association between nodes irrespective of their physical location. A node can connect to another node, close to or far from it. Geometric measures, instead, evaluate how nodes are physically associated in a geographic space. Since geometric measures are usually related with the distance between connected nodes, they are frequently used in structural connectivity descriptions. For this reason, we will not discuss geometric metrics in this text, but it can find in Bullmore and Basset.

84

J. Faber et al.

4.1 Topological Measures Topological measure comprises a set of metrics used to quantify different network information. To choose the most informative metrics, it is necessary to consider the type of network and the type of information being studied. Essentially, topological metrics in neuronal networks measure how functional integration, segregation, efficiency, resilience, and motifs describe the information storage, flow, and assessment in the brain and how the network integrity is maintained. Some basic network characteristics greatly affect many topological measures, like number of nodes, number of edges, clustering coefficient, path length, and degree (which is considered a fundamental measure—Fig. 8). The degree distribution, i.e., the probability distribution of the number of edges connected to a node, can determine the complexity of arranged network architectures.

4.1.1

Functional Segregation

In general, functional brain connections present features of complex networks with non-random connections and shared relationships [92]. The study of brain functional segregation analyzes how networks are organized in specialized cores for information storage and processing (Fig. 8). The best example of this organization is the brain cortex that, despite its apparent homogeneity, is composed by many distinct functional areas such as the visual and somatosensory-motor regions [93]. Modules are characterized by a large number of connections between elements inside them, also called “community”, and lower numbers among elements of other communities. Metrics able to detect these communities are known as modularity and clustering Fig. 8 Basic topological features and segregation measures. a Node, the basic connectivity unity of a graph. b Edges, the connection representation between nodes c A node degree is the number of connections that a specific node makes with other nodes. d Segregation measures like clustering coefficient calculate the existence of specialized modules of information storage and processing

a

b

d

c

Critical Elements for Connectivity Analysis of Brain Networks

85

or yet module coefficients (C, [11, 94]). All mathematical metrics and definitions presented in this section can be also consulted in Rubinov and Sporns [11, 12]. In general, brain networks present higher C, most probably due to the network arrangement that divides and compartmentalize the information flow to optimize the mechanism of information processing. The study of how these modules are connected leads to the understanding of the integration between them. For example, how the cortical infrastructure supports a single function involving specialized areas linked by the functional integration between them. Therefore, segregation only makes sense in functional integration context and vice versa [52, 95].

4.1.2

Functional Integration

Functional integration is associated with the network capacity to involve global interactions transcendent to the limits of modules. A cortical structure supporting or dedicated to a special function is made of many segregated information, implying that there is a coordinated activation of many neurons in different regions. Complex dynamics of cognitive or behavioral control requires efficient communication among diverse modules and a high capacity to integrate distributed information [6]. For any connectivity analysis, measures of these attributes are associated with paths and hubs of a network. As mentioned in Session 3, in functional connections of networks, paths are sequences of statistical dependencies that does not necessarily correspond to structural connections [11]. Pathlengths indicate the efficiency between connections of different modules. A shorter path implies a stronger integration since information transmission is faster. The average shortest path length (L) plays an important role in the characterization of a graph and is an important measure of integration and efficiency. Therefore, the global efficiency of a network is given by the average of the inverse shortest path length [30]. Based on this measure, highly disconnected networks present paths tending to infinity and, consequently, the efficiency tends to zero. In addition, there are two more network attributes that indicate the integration of information: network hubs and interconnection propensity of hubs [94]. Hubs are associated with central nodes in the network and are very important due to their high connectivity density with other nodes. They can be identified by different metrics that measure degree, number of connections between specific nodes (Fig. 9), and centrality or participation in modular connectivity. Global hubs are responsible for intermodular communication and integration, while hubs inside a module promote the cohesion of their own communities [11, 12, 94].

86

J. Faber et al.

a

b

Fig. 9 Integration Measures. a Hubs are network elements that integrate different modules and can be calculated by metrics like degree, centrality or participation in modular connectivity. b Paths are sequences of minimum links between distinct nodes and are associated with the network efficiency by calculating the shortest path

4.1.3

Network Resilience

An important network characteristic is its reliability, measured by its resilience. In brain networks, reliability is associated with the capacity of brain systems to overcome a pathological attack by a disease or an aberrant development [96]. Functional damages are closely related to damages in biological network structures since many neuropathologic lesions can affect functional brain activity. For instance, depending on the injured areas, some brain functions can be lost after a stroke [97]. Another example is the disconnection hypothesis in schizophrenia that suggests that an impaired neuromodulation of synaptic plasticity results in an abnormal functional integration of particular neural systems [98]. The brain network resilience, therefore, is its capacity to adapt and maintain its functionality over physiological adversities, and it can be directly or indirectly characterized through topological measurements. Network resilience can be measured, for example, by using assortativity or average neighbor degree [11, 24, 26] before and after an insult to test the network vulnerability or the network ability to recover from that insult [99]. Moreover, the insult can be computationally simulated by removing random or targeted nodes. Thus, the effects of the lesion may be measured and compared with the structural, functional, and effective connectivity [11].

4.1.4

Complex Network Architectures

Natural networks present architectural features that reflect their construction or development processes and function. As mentioned in Sect. 4.1.2, k i is the number of vertices linked with the node i. Thus, the probability pk is defined as the fraction of

Critical Elements for Connectivity Analysis of Brain Networks

87

vertices that have degree k in the network. In other words, pk is the probability that a node chosen at random has a degree k and it is given by the distribution function P(k) [27, 29, 30]. The degree distribution can be presented as a histogram of the degrees of vertices and described by the function that fit the histograms as shown in Fig. 10. In networks architectures, random graphs are associated with the disordered nature of the links between nodes. In random graphs all connections are equally probable, resulting in a Gaussian degree distribution [30, 101]: Pk = e−

< k >k k!

(7)

where is the average degree of the network and Pk gives the probability to randomly select a node with exactly k edges. In this kind of graph, all edges are randomly designated as nodes pairs. Because of this, its C and L are very small and very short, respectively [95]. As opposed to random graphs, regular lattice graphs have a very ordered pattern of connection between nodes, such as in ring or grid lattice (Fig. 10), where the connected nodes tend to have the same neighbors but the path lengths between them vary greatly and the shortest paths are compounded by many intermediate nodes. Hence, lattice graphs have bigger C and longer L values [26]. Many natural networks have an uneven distribution with a much skewed and slower decaying than a Poisson distribution. One example of this dynamics is given by a power law decaying [102]: Pk ∼ k −γ

(8)

These networks are called scale-free [26]. A general characteristic of this kind of network is the existence of hubs, since some nodes are highly connected while others have few connections [27]. Some examples of scale-free networks are metabolic networks [103], gene regulatory networks [104], World-Wide Web [105], etc. Some studies have investigated a possible scale-free organization of functional connectivity in human brains, but the results of these studies have been inconclusive [106, 107]. Conversely, van den Heuvel et al. [108] suggested that the functional connectivity of the human brain is a combination of the scale-free and small-world organization. Small-world networks combine high levels of local clustering among network nodes and short paths that globally connect all nodes of the network, promoting integration between clusters [96]. As shown in Fig. 10, both degree distributions of small-world and random networks can be fitted and modeled by a Poisson function. But the difference between them is that C is much higher in a small-world network than in a random network, whereas L is similar in both networks, given that they have the same size [27]. Then, these criteria are used to determine a network with a small-world architecture evaluated by the expression below [12]:

88 Fig. 10 Network architectures and examples of P(k) distributions. a Random network, nodes were random connected by the edges. b Example of a P(k) distribution of a random network with k varying from 0 to 100. c Lattice network, nodes were ordered connected and all of them have the same node degree. d Example of a P(k) distribution of the lattice network presented in item C. e Scale-free network, the clusters formation can be observed. f Example of a P(k) distribution of a scale-free network, the function of distribution follows a power law. g Small-world network, the formation of clusters can be visualized, and there is integration among them. h Example of a P(k) distribution of a small-world network with k varying from 0 to 50, the P(k) distribution may be similar to the random network distributions, what will differentiate it will be its segregating properties. All the P(k) distributions are represented by an empirical histogram and by an analytical model (Adapted from [100, 27])

J. Faber et al.

a

b

c

d

e

f

g

h

Critical Elements for Connectivity Analysis of Brain Networks

S=

c crand L Lrand

89

(9)

where C and L are the clustering coefficient and characteristic path length, respectively, being compared to the C and L of their corresponding random network (modeled computationally). Watts and Strogatz [109] demonstrated the presence of small-world topology in the nervous system of C. elegans [32, 110, 111]. According to Basset and Bullmore [96], there are some reasons for neural networks to be small-world, since the brain is composed by a complex network with multiple spatial and time scales. In the macroscale network information is segregated and distributed and then is integrated to form a unique function. Similarly, small-world architecture comprises a high clustering coefficient and a short path length indicative of segregated and distributed processing and information integration, respectively. In addition, during brain development, the network is optimized to minimize costs and maximize the efficiency of information processing. These characteristics can be found in small-world networks with high global and local efficiency, which can indicate parallel information processing, low wiring costs, and sparse connectivity between nodes. Considering all concepts together, when analyzing a functional and effective network architecture, it is important to pay attention to some critical points. For instance, depending on the feature being considered as a node and the metric to compute the communication among them, a totally different topology may arise. In this way, it is critical to be careful with the classification of specific regions of the brain, but mainly what physical/signal features, and also which metric are been considered. A functional network described by one specific feature and specific metric can present one kind of architecture that is impossible to generalize to all functional brain networks associated with other features and metrics since the node degrees can change completely. Because of the popularity of connectomics, many studies have presented strong claims about brain networks; however, a bit of caution and conservatism is needed since it is very difficult to reduce all aspects of the brain to simple features and simple interactions.

4.1.5

Network Motifs

Significant and recurrent patterns of node interconnections are known as network motifs. Usually, the connection patterns of a network are compared with a random network to find patterns that appear in numbers significantly higher than those in a randomized network [112]. In this way, network motifs are well-defined connectivity blocks that appear in a right network with equal or greater probability when compared with a random network simultaneously lower than a cutoff value. It is important to mention that there are patterns without any statistical significance that are still important for the network [112].

90

J. Faber et al.

a

b

c

Fig. 11 Motifs examples. Three recurring connection patterns are show: a three undirected edges form a triangle linking three nodes, b three directional edges link three nodes forming a feedback loop and c four edges form two parallel ways that leave and arrive at the same node

As shown in Fig. 11, the functional network topology can exhibit, for instance, triangles and feedback loops or biparallel blocks that represent specific mechanisms of the network such as information protection, processing, and storage. The network motifs can also be measured by its frequency of occurrence, normalized as the motif z-score [11]. Measure

Description

Node degree

Measures the number of connections of node i. It is a basic measure used by many others measures

Clustering coefficient

Measures the degree that the graph nodes tend to cluster together

Global clustering coefficient

Mathematical definition  ki = i,j∈N aij aij is the connection status between the node i and the node j (when i and j are neighbors). When the link exists the aij value is 1 otherwise is 0 i Ci = ki (kΓi −1)

i is edges between neighbors

 Measures the clustering C = 1n i∈N Ci coefficient of the entire network n is the number of nodes in the network and C i is the Cluster coefficient (continued)

Critical Elements for Connectivity Analysis of Brain Networks

91

(continued) Measure

Description

Modularity

Measures the strength that a Q= network is divided into modules 1 

Mathematical definition

l

i,j∈N



k −k aij − iin T jout δij

l is the total number of edges, aij is the element of the adjacency matrix, kiin is the degree of node i, kjout is the degree of node j, δij is the Kronecker delta (1 if nodes i and j are in the same module and zero, otherwise)  Shortest path length Measure of the shortest path dij = akj ∈g(k↔j) akj length between two nodes akj is the the connection status between nodes in the shortest path (geodesic distance) between nodes k and j (g(k ↔ j), akj = 1 if there is connection and 0 otherwise)  1 Average shortest path length Measure the average shortest L = n(n−1) i,j∈N dij path length between all nodes in n is the total number of nodes a network in the network  1 1 Global efficiency Measure of how efficiently the E = n(n−1) i,j∈N dij network globally exchanges dij is the shortest path length information and n is the total number of nodes in the network Closeness centrality

Betweenness centrality

Measure of centrality that indicates the average length of the shortest path between a node i and all other nodes in the network

L−1 i =

Measure of centrality that indicates how many times a node acts as a bridge along the shortest path between two other nodes

bi =

 n−1

j∈N ,j =1 dij

n is the total number of nodes in the network. (Normalized form) 1 (n−1)(n−2)

 k,j∈N

ρhj (i) ρhj

ρ hj is the number of shortest paths between h and j, and ρ hj (i) is the number of shortest paths between h and j that pass through i

From the point of view of brain complexity, where structural connectivity generates different functional states with different features, Sporns and Kotter [5] hypothesized that brain networks have a large number of motifs in functional connectivity to maximize their number and diversity of cognitive states, keeping the same number of structural motifs.

92

J. Faber et al.

5 Network Dynamics and Multilayer Networks As previously described, the structural brain connectome can be modeled as networks at different spatial scales [3, 13, 113]. However, how the structural and functional connectivity interplay with and within other neural networks in space and time remains unclear. The investigation of this question embraces multiple spatiotemporal scales and demand several modalities of experimental techniques for data recording. The multilayer network approach allows for the merging of datasets in a consistent way and bring to light the hidden features of the complex organization of the brain networks.

5.1 From Static to Dynamic Networks As described in Sect. 2, a network is defined as a graph (G), an abstract representation corresponding to a network arrange. To capture further information in the graph, such as temporal information, new edges need to be included along additional non-nodal dimensions. The extended mathematical definition of a graph (see Sect. 2.1) is: G = {V, E, D}

(10)

where D is a set of additional non-nodal dimensions [114]. The Eq. 10 is sometimes referred in mathematics as a multigraph network and, in network theory, as a multilayer or multiplex network [115]. A graph is said to be a dynamical network (or a temporal network) when D contains an ordered set of temporal indices representing time. In this way, D could be a set containing discrete temporal indices in seconds, minutes, hours, days or even years:t = {1, 2, . . . , T}. In this case, a temporal network can be expressed by a group of static graphs G = Gt of size NxN nodes, corresponding to a series of “photographs” of the network at each time t (Fig. 12 [114]). The dynamical networks theory can be applied to investigate the oscillatory behavior of the resting state networks, since it allows us to assess the intrinsic dynamics of the network nodes and the couplings between all nodes of a network across time, revealing the functional network dynamics [116]. The alterations in brain connectivity over time can be assessed by multiple approaches [117], such as: (i) the dynamic functional network connectivity: characterization of temporal coupling changes between fixed spatial networks [118]; (ii) the time-varying spatial connectivity: evaluation of changes in spatial patterns of correlated networks over time [119]; and (iii) the time-varying graph metrics: quantification of graph measures which can reconfigure over time [117]. Thompson et al. [114] presented the principal measures from dynamical network theory (briefly described in Table 2). Most of the mathematical metrics are applied to binary, non-

Critical Elements for Connectivity Analysis of Brain Networks

93

Fig. 12 Scheme of dynamical networks in the brain. In a dynamic functional connectivity analysis, the nodal time series are time-windowed and the relationship between pairs of nodes is given by connectivity metrics (see the measures described in Sects. 3.2 and 3.3). The static layers for each time-window (Gt ) are concatenate to a N × N × T array representing the changes in functional connectivity between nodes as a function of time

directed graphs, although several of them can be fitted for non-binary, directed, and continuous time graphs. The burstiness coefficient (Bij ) in Table 2 calculates the number of bursts per edge but can also be applied to a given node by the summation of the burstiness coefficient of all edges associated with its node. Similarly, the definitions of fluctuability, volatility, and temporal efficiency can be extended to a nodal level. The set of metrics described in Table 2 summarize the connectivity information over short- and long-term time-scales, allowing to identify groups of edges that have similar temporal evolution or investigate how different tasks evoke different network configurations [120, 121]. However, it is necessary to evaluate which dynamic network metrics are more appropriate for each research problem. Although dynamical network theory allows access to several metrics, it is not advisable to apply all the available measures to a given dataset. A hypothesis about a possible network state should first be considered and after a measure that will help to quantify this network configuration and why it is considered. Besides, for the interpretation of a specific measure, an account of the data temporal resolution and the level of network organization under analysis must be taken, globally or locally [114]. Investigations are still needed to validate existing models, build improved models, and develop high-level summary metrics. In addition, there are still several aspects

94

J. Faber et al.

Table 2 Principal dynamical network metrics Measure

Description

Mathematical definition

Influence of a node i in the dynamical network, calculated by the sum of the edges for the node and the number of the edges across time

DiT =

Local level measures Temporal centrality (DT )

N T j=1

t t=1 Gi,j

T : number of time points N : number of nodes t : time-graphlet Gi,j

N 1 T = 1 Temporal closeness centrality Time between connections, Ci,t j=1 d t N −1 i,j T given by the inverse sum of the (C ) t shortest paths across all time di,j : average shortest path points, between nodes i and j Burstiness (Bij )

σ (τ )−μ(τ ) Measure of distribution of Bij = σ τij +μ τij ( ij ) ( ij ) subsequent connections per edge. B > 0 indicates that the τij : distribution of intercontact temporal connectivity is bursty times between nodes i and j through time σ : standard deviation, μ: mean

Global level measures Fluctuability (F)

Volatility (V )

Reachability latency (Rr )

Temporal efficiency (E)

 

U (GI ,j ) Ratio of number of edges F = i j  Gt present in G over the all edges i j t I ,j   T t of G t . Quantify the temporal U Gi,j = 1, if t Gi,j > 0 variability of connectivity. F is  T t 1 when every edge is unique 0, if t Gi,j = 0 and occurs only once in time  t t+1  1 T −1 Rate of consecutive change of V = T −1 t=1 D G , G graphlets over time D: Hamming distance that

Quantifies the average time it takes for a dynamical network to reach an a priori defined reachability ratio r

quantifies the difference between a graphlet at t and the graphlet at t + 1 1   t Rr = TN t i di

dit : ordered vector of length N of the shortest temporal paths for node i at time point t. k: [rN ] th element of dit  1 Inverse average shortest E = T N 21−N , ( ) I ,j,t dIt,j temporal path. This measure is calculated at each time point i = j through the inverse of the average shortest path length for all nodes to obtain an estimate of global temporal efficiency

Critical Elements for Connectivity Analysis of Brain Networks

95

of time-varying brain connectivity that need to be studied, which include: (i) mathematical or physical models that can capture both spatial and temporal couplings; (ii) approaches to capture both static and dynamic connectivity; and (iii) application of the existing tools to large datasets to identify predictor parameters.

5.2 Multilayer Networks The multilayer network formalism allows us also to include many other dimensions of information, encoding its different network layers [15, 122, 123], including: (i) the activity across multiple spatial scales; (ii) the activity in different frequency-bands [124]; (iii) the multi-modal networks connectivity [125]; and (iv) the relationship of structural and functional/effective networks [126]. Therefore, a multilayer network can be described as a network of networks [127], or a network that contains different layers, in which the edges in a given layer represent a different type of relationship in another layer [128], Fig. 13. The traditional network models have provided key insights into the structure and function of the brain through the assessment of descriptive and inferential network measures [32]. However, single networks provide a limited representation of the brain structure by excluding or aggregating the multiple connection types between its components [29, 115]. The multilayered network approach for modelling brain

a

b

Fig. 13 Schematic representation of multilayer networks. The multilayer network framework allows to represent systems that consist of networks at multiple levels or with multiple types of edges (e.g., the functional brain activity in different spatial scales). In (a), each layer of a multilayer network corresponds to a different type of interaction between nodes and is represented by a different adjacency matrix. As shown in (b), the layers can also be interconnected. In the multilayer network approach, the interlayer edges can embrace pairwise connections between all possible combinations of nodes and layers and it is possible generalize this framework to consider hyperedges that connect more than two nodes (for details, see [115])

96

a

J. Faber et al.

b

Fig. 14 Representation of a supra-adjacency matrix. a A multilayer network constitutes of three different layers, each one represented by a (possible directed and/or weighted) adjacency matrix. b The supra-adjacency matrix of multilayer network shown in (a), representing intra- and inter-layer connectivity

organization allows the incorporation of multiple structural relationships, known as multiplexity [129, 130], that go beyond the statistical dependencies or correlations between network elements [115, 122]. A fundamental aspect of describing multilayered networks is defining and quantifying the interconnectivity between different categories of connections. This amounts to switching between layers in a multilayered system, and the associated inter-layered connections in a network are responsible for the emergence of new phenomena in multilayered networks. The structure of a multilayer network can be represented by a supra-adjacency matrix [115], as shown in Fig. 14. This approach allows the application of numerous tools and methods that have been developed for matrices in the investigation of multilayer networks. Additionally, the supra-adjacency matrix representation is useful to describe walks on multilayer networks and provide a way to depict multilayer networks that are not node-aligned without added empty nodes [115]. However, to construct the supra-adjacency matrix, we must flatten the multilayer network and some of the information can be lost. To overcome this issue Kivelä et al. [115] suggested that the edge set of the multilayer network should be grouped with the intralayered edges. The interlayered and coupling edges construct the respective supra-adjacency matrix.

5.3 Edges Between Networks In a multilayer network, establishing how each subnet communicates and connects is a job that depends on experimental factors or mathematical models that describe how

Critical Elements for Connectivity Analysis of Brain Networks

97

each scale transmits information to another. In situations where information levels are the same between different layers of networks, describing this link is reduced to describing the communication between different modules in a complex network. However, adequately describing this link may not be such a simple task when there are different layers processing different information through nodes with different modalities. Certainly, experimental criteria and protocols that guide the type and intensity of these communications offer great help in describing these connections. However, depending on the types of networks being connected, an experimental protocol may need levels of control that are difficult to achieve, especially when evaluating functional and effective networks in human brains. Today, this topic is one of the great frontiers in network theories due to the large number of variables and parameters to be considered as information in the description of isolated and, mainly, connected networks. Thus, for a better description of connectivity patterns and brain activity, there is still a long way to go in order to better understand all the relationships and effects of each characteristic and each metric in different network topological measures.

6 Clinical Applications Technological developments of non-invasive neuroimaging techniques associated with powerful computational processing and big data have made possible the study of brain function in normal and diseased conditions. Many multinational groups have been formed with the aim of combining technologies to study the brain, such as the Human Brain Project (HBP—http://www.humanbrainproject.eu/en/), Brain Initiative (https://www.braininitiative.nih.gov/), Human Connectome (HCP— https://www.neuroscienceblueprint.nih.gov/connectome/), Virtual Brain (VB— http://www.thevirtualbrain.org/tvb/zwei), Brain Minds (http://brainminds.jp/en/cen tral/mission), Blue Brain Project (BBP—https://bluebrain.epfl.ch/), China Brain Project (https://www.ncbi.nlm.nih.gov/pubmed/27809999), as well as many other research groups. These initiatives promote discoveries in neurological and neuropsychiatric diseases that can facilitate our understanding of their behavior and consequently their treatment. Network theory has made it possible to understand how brain networks develop and how structures, from genes to brain areas, interact to form architectures that show universal features like hierarchical modularity and small-world organization, and how it is associated with functional cognitive and intelligence. Understanding the healthy dynamics of the brain can lead to an understanding of what has been damaged in neurological and neuropsychiatric diseases. Hence, network theory has been explored in clinical neurophysiology and has impacted clinical concepts. A good example of this is the idea that Alzheimer’s disease and epilepsy can be explained in terms of “hub failure” [22].

98

J. Faber et al.

Since, in principle, a disease alters the regular functioning of the brain, we can assume that functional connectivity will also be altered. For this reason, many studies were designed to understand functional and effective connectivity under normal and diseased conditions. Huang et al. [28], using resting-state fMRI (RS-fMRI), studied the functional connectivity pre-to-post-operative after total knee arthroplasty with general anesthesia. They observed that 48 h after surgery at least one fourth of the sample showed significant functional network decline, indicating that the integrity of the brain is disturbed after general anesthesia. Supekar et al. [131] compared network topological properties of children and young-adults and concluded that apart from the fact that both networks have small-world organization, they differ significantly in hierarchical organization and interregional connectivity, suggesting the existence of key principles underlying functional brain maturation. A very studied neuropsychiatric disease using network theory is schizophrenia. Van den Heuvel et al. [108] examined a topology structure of rich club in patients with schizophrenia and its role in global functional brain dynamics. They noticed a reduction of rich club connections in patients and associated this reduction with lower levels of global communicative capacity. Lynall et al. [132] measured aspects of functional network topology using RS-fMRI time series of schizophrenics. They used interregional correlation matrices to construct weight graphs and observed that the schizophrenic group showed weaker integration and more diverse connections in functional connectivity. Ganella et al. [133] used RS-fMRI in treatment resistant schizophrenics (TRS) and unaffected first-degree family members (UFM) to study risk or resilience endophenotypes in schizophrenia associated with functional brain connectivity. They infer that both the TRS and UFM groups had functional connectivity deficits representing a risk endophenotype. Nevertheless, the UFM functional connectivity is more topologically resilient than that in TRS, which may explain the absence of schizophrenia despite familial liability. Network theory is also widely used in the study of epilepsy. Zhang et al. [134] hypothesized a decoupling of structural and functional connectivity in epilepsy. Using fMRI images in idiopathic generalized epileptics, they corroborated their hypothesis and suggested that this decoupling can be used as a biomarker of subtle brain abnormalities in epilepsy. Hogan [135] studied the effects of local insults on brain development. They found indications of correlation between severity of topological network reconfiguration and time of insult during corticogenesis. Tecchio et al. [136] also studied patients with drug-resistant epilepsy (DRE), investigating changes in functional connectivity caused by cathodal transcranial direct current stimulation (ctDCS) using eLORETA analysis in EEG data. They verified a correlation between epileptic seizure reduction and increase of functional connectivity in the epileptic focus after ctDCS in DRE patients, which can contribute to understanding the underlying mechanisms of ctDCS treatment. Lately, Alzheimer’s disease (AD) has been studied with the aid of network theory. Supekar et al. [131] studied if the small-world brain properties are lost in AD. They noticed that the clustering coefficient distinguished health subjects into patients with sensitivity of 72% and specificity of 78%, indicating that it may be a

Critical Elements for Connectivity Analysis of Brain Networks

99

potential biomarker of AD. Stam et al. [137] also investigated the functional brain networks disruption in AD patients. They used beta band–filtered EEG and observed that AD patients presented a characteristic path length longer than that of healthy subjects, while clustering coefficient did not present significant changes. These findings suggest a less optimal organization and a loss of complexity in AD. Kabbara et al. [138] reported that AD patients showed less functional integration and more functional segregation compared with healthy subjects. They also found an association between cognitive scores of AD patients and their alterations in functional brain networks. These applications exemplify the breadth, robustness, diversity, and effectiveness that an analysis of brain dynamics based on a network approach can provide. There is still a great road to drive, and it is not yet known if graph-based network analysis will be only a new technique for quantifying patterns or whether it will be a true theory with a new perspective on how the brain represents information through its biological structures and how it processes information in time and space, integrating different modalities and providing cognitive emergent states.

References 1. Williams N, Henson RN (2018) Recent advances in functional neuroimaging analysis for cognitive neuroscience. Brain Neurosci Adv 2:239821281775272. https://doi.org/10.1177/ 2398212817752727 2. Bassett DS, Sporns O (2017) Network neuroscience. Nat Neurosci 20(3):353–364. https:// doi.org/10.1038/nn.4502 3. Sporns O, Tononi G, Kötter R (2005) The human connectome: a structural description of the human brain. PLoS Comput Biol 1(4):0245–0251. https://doi.org/10.1371/journal.pcbi.001 0042 4. Avena-Koenigsberger A, Misic Bratislav, Sporns Olaf (2017) Communication dynamics in complex brain networks. Nat Rev Neurosci 19(1):17–33. https://doi.org/10.1038/nrn.201 7.149 5. Sporns O, Kötter R (2004) Motifs in brain networks. PLoS Biol 2(11). https://doi.org/10. 1371/journal.pbio.0020369 6. Sporns O (2002) Network analysis, complexity, and brain function. Complexity 8(1):56–60. https://doi.org/10.1002/cplx.10047 7. Honey CJ, Thivierge JP, Sporns O (2010) Can structure predict function in the human brain? NeuroImage 52(3):766–776. https://doi.org/10.1016/j.neuroimage.2010.01.071 8. Fornito A, Zalesky A, Breakspear M (2013) Graph analysis of the human connectome: promise, progress, and pitfalls. NeuroImage 80:426–444. https://doi.org/10.1016/j.neuroi mage.2013.04.087 9. Bastos AM, Schoffelen J-M (2016) A tutorial review of functional connectivity analysis methods and their interpretational pitfalls. Front Syst Neurosci 9(January):1–23. https://doi. org/10.3389/fnsys.2015.00175 10. Canolty RT, Knight RT (2010) The functional role of cross-frequency coupling. Trends Cogn Sci 14(11):506–515. https://doi.org/10.1016/j.tics.2010.09.001 11. Rubinov M, Sporns O (2010) Complex network measures of brain connectivity: uses and interpretations. NeuroImage 52(3):1059–1069. https://doi.org/10.1016/j.neuroimage.2009. 10.003

100

J. Faber et al.

12. Kaiser M (2011) A tutorial in connectome analysis: topological and spatial features of brain networks. NeuroImage 57(3):892–907. https://doi.org/10.1016/j.neuroimage.2011.05.025 13. Sporns O (2013) The human connectome: origins and challenges. NeuroImage 80:53–61. https://doi.org/10.1016/j.neuroimage.2013.03.023 14. Van Diessen E, Diederen SJH, Braun KPJ, Jansen FE, Stam CJ (2013) Functional and structural brain networks in epilepsy: what have we learned? Epilepsia 54(11):1855–1865. https:// doi.org/10.1111/epi.12350 15. Betzel RF, Bassett DS (2017) Multi-scale brain networks. NeuroImage 160(November):73– 83. https://doi.org/10.1016/j.neuroimage.2016.11.006 16. Buzsáki G, Anastassiou CA, Koch C (2012) The origin of extracellular fields and currentsEEG, ECoG, LFP and spikes. Nat Rev Neurosci 13(6):407–420. https://doi.org/10.1038/nrn 3241 17. Babiloni C, Pizzella V, Del Gratta C, Ferretti A, Romani GL (2009) Chapter 5 Fundamentals of electroencefalography, magnetoencefalography, and functional magnetic resonance imaging. Int Rev Neurobiol 86 (1st edn., Elsevier Inc.) https://doi.org/10.1016/S0074-7742(09)860 05-4 18. Pizzagalli DA (2007) Electroencephalography and high-density electrophysiological source localization. Handbook Psychophysiol 3:56–84. https://doi.org/10.1109/TNSRE. 2007.903919 19. Bullmore ET, Bassett DS (2011) Brain graphs: graphical models of the human brain connectome. Ssrn. https://doi.org/10.1146/annurev-clinpsy-040510-143934 20. Douw L, van Dellen E, de Groot M, Heimans JJ, Klein M, Stam CJ, Reijneveld JC (2010) Epilepsy is related to theta band brain connectivity and network topology in brain tumor patients. BMC Neurosci 11. https://doi.org/10.1186/1471-2202-11-103 21. Baliki MN, Geha PY, Apkarian AV, Chialvo DR (2008) Beyond feeling: chronic pain hurts the brain, disrupting the default-mode network dynamics. J Neurosci 28(6):1398–1403. https:// doi.org/10.1523/JNEUROSCI.4123-07.2008 22. Stam CJ, van Straaten ECW (2012) The organization of physiological brain networks. Clin Neurophysiol 123(6):1067–1087. https://doi.org/10.1016/j.clinph.2012.01.011 23. Stam CJ (2014) Modern network science of neurological disorders. Nat Rev Neurosci 15(10):683–695. https://doi.org/10.1038/nrn3801 24. Bullmore ED, Sporns O (2009a) Complex brain networks: graph theoretical analysis of structural and functional systems. Nat Rev Neurosci 10(3):186–198. https://doi.org/10.1038/nrn 2575 25. Fornito A, Zalesky A, Pantelis C, Bullmore ET (2012) Schizophrenia, neuroimaging and connectomics. NeuroImage 62(4):2296–2314. https://doi.org/10.1016/j.neuroimage.2011. 12.090 26. Sporns O (2011) The human connectome: a complex network. Ann N Y Acad Sci 1224(1):109–125. https://doi.org/10.1111/j.1749-6632.2010.05888.x 27. Costa LF, Rodrigues FA, Travieso G, Villas Boas PR (2007) Characterization of complex networks: a survey of measurements 56 (February 2007): 167–242. https://doi.org/10.1080/ 00018730601170527 28. Huang H, Tanner J, Parvataneni H, Rice M, Horgas A, Ding M, Price C (2018) Impact of total knee arthroplasty with general anesthesia on brain networks: cognitive efficiency and ventricular volume predict functional connectivity decline in older adults. J Alzheimer’s Dis 62(1):319–333. https://doi.org/10.3233/JAD-170496 29. Newman MEJ (2003) The structure and function of complex networks. SIAM Rev 45(2):167– 256. https://doi.org/10.1137/S003614450342480 30. Boccaletti S, Latora V, Moreno Y, Chavez M, Hwang DU (2006) Complex networks: structure and dynamics. Phys Rep 424(4–5):175–308. https://doi.org/10.1016/j.physrep.2005.10.009 31. Amaral LAN, Ottino JM (2004) Complex networks. Eur Phys J B—Condens Matter 38(2):147–162. https://doi.org/10.1140/epjb/e2004-00110-5 32. Sporns O, Chialvo DR, Kaiser M, Hilgetag CC (2004) Organization, development and function of complex brain networks. Trends Cogn Sci 8(9):418–425. https://doi.org/10.1016/j.tics. 2004.07.008

Critical Elements for Connectivity Analysis of Brain Networks

101

33. Allen EA, Damaraju E, Eichele T, Wu L, Calhoun VD (2018) EEG signatures of dynamic functional network connectivity states. Brain Topogr 31(1):101–116. https://doi.org/10.1007/ s10548-017-0546-2 34. Shine JM, Aburn MJ, Breakspear M, Poldrack RA (2018) The modulation of neural gain facilitates a transition between functional segregation and integration in the brain. ELife 7:1–16. https://doi.org/10.7554/eLife.31130 35. Li K, Guo L, Nie J, Li G, Liu T (2009) Review of methods for functional brain connectivity detection using FMRI. Comput Med Imaging Graph 33(2):131–139. https://doi.org/10.1016/ j.compmedimag.2008.10.011 36. Barch DM, Burgess GC, Harms MP, Petersen SE, Schlaggar BL, Corbetta M, Glasser MF et al (2013) Function in the human connectome: task-FMRI and individual differences in behavior. NeuroImage 80:169–189. https://doi.org/10.1016/j.neuroimage.2013.05.033 37. Rissman J, Gazzaley A, D’Esposito M (2004) Measuring functional connectivity during distinct stages of a cognitive task. NeuroImage 23(2):752–763. https://doi.org/10.1016/j.neu roimage.2004.06.035 38. Baker JM, Bruno JL, Gundran A, Hadi Hosseini SM, Reiss L (2018) FNIRS measurement of cortical activation and functional connectivity during a visuospatial working memory task, pp 1–22. https://doi.org/10.1371/journal.pone.0201486 39. Zhang H, Zhang YJ, Chun Ming L, Ma SY, Zang YF, Zhu CZ (2010) Functional connectivity as revealed by independent component analysis of resting-state FNIRS measurements. NeuroImage 51(3):1150–1161. https://doi.org/10.1016/j.neuroimage.2010.02.080 40. Grech R, Cassar T, Muscat J, Camilleri KP, Fabri SG, Zervakis M, Xanthopoulos P, Sakkalis V, Vanrumste B (2008) Review on solving the inverse problem in EEG source analysis. J NeuroEng Rehabil 5. https://doi.org/10.1186/1743-0003-5-25 41. Bradley A, Yao J, Dewald J, Richter CP (2016) Evaluation of electroencephalography source localization algorithms with multiple cortical sources. PLoS ONE 11(1):e0147266. https:// doi.org/10.1371/journal.pone.0147266 42. Pascual-Marqui RD (1999) Review of methods for solving the EEG inverse problem. Int J Bioelectromagnetism 1(1):75–86. https://www.citeulike-article-id:5020586 43. Chen XL, Xiong YY, Xu GL, Liu XF (2012) Deep brain stimulation. Intervent Neurol 1(3– 4):200–2012. https://doi.org/10.1016/B978-0-12-385157-4.00740-5 44. Engel AK, Moll CKE, Fried I, Ojemann GA (2005) Invasive recordings from the human brain: clinical insights and beyond. Nat Rev Neurosci 6(1):35–47. https://doi.org/10.1038/nrn1585 45. Lal TN, Hinterberger T, Widman G, Schroder ¨ M, Hill J, Rosenstiel W, Elger CE, Scholkopf ¨ B (2005) Methods towards invasive human brain computer interfaces. Adv Neural Inf Proc Syst 17:737–744. https://doi.org/10.1.1.64.8486 46. Sporns O (2016) Connectome networks: from cells to systems. In: Micro-, Meso- and macroconnectomics of the brain, pp 107–127 47. Craddock RC, Jbabdi S, Yan CG, Vogelstein JT, Xavier Castellanos F, Di Martino A, Kelly C, Heberlein K, Colcombe S, Milham MP (2013) Imaging human connectomes at the macroscale. Nat Methods 10(6):524–539. https://doi.org/10.1038/nmeth.2482 48. Poli D, Pastore VP, Massobrio P (2015) Functional connectivity in vitro neuronal assemblies. Front Neural Circ 9(October):57. https://doi.org/10.3389/fncir.2015.00057 49. Yoneki E, Hui P, Crowcroft J (2008) Distinct Types of hubs in human dynamic networks. In: Proceedings of the 1st workshop on social network systems—SocialNets ’08, pp 7–12. https://doi.org/10.1145/1435497.1435499 50. Navlakha S, Rastogi R, Shrivastava N (2008) Graph summarization with bounded error. In: Proceedings of the 2008 ACM SIGMOD international conference on management of data—SIGMOD ’08, p 419. https://doi.org/10.1145/1376616.1376661 51. Park HJ, Friston K (2013) Structural and functional brain networks: from connections to cognition. Science 342(6158). https://doi.org/10.1126/science.1238411 52. Friston KJ (2011) Functional and effective connectivity: a review. Brain Connectivity 1(1):13– 36. https://doi.org/10.1089/brain.2011.0008

102

J. Faber et al.

53. Altman N, Krzywinski M (2015) Points of significance: association, correlation and causation. Nat Methods 12(10):899–900. https://doi.org/10.1038/nmeth.3587 54. Valdes-Sosa PA, Roebroeck A, Daunizeau J, Friston K (2011) Effective connectivity: influence, causality and biophysical modeling. NeuroImage 58(2):339–361. https://doi.org/10. 1016/j.neuroimage.2011.03.058 55. Akaike H (1968) On the use of a linear model for the identification of feedback systems. Ann Inst Stat Math 20(1):425–439. https://doi.org/10.1007/BF02911655 56. Granger CWJ (1969) “Investigating causal relations by econometric models and crossspectral methods author(S):” Ecometrica 37:424–438. http://ir.obihiro.ac.jp/dspace/handle/ 10322/3933 57. Schweder T (1970) Composable Markov processes. J Appl Probab 7 (Aug 1970):400–410. https://doi.org/10.1111/j.1365-2672.2007.03484.x 58. Wiener N (1956) The theory of prediction. Modern mathematics for engineers. New York, pp 165–90 59. Razi A, Friston KJ (2016) The connected brain. IEEE Sig Process Mag 33(3):14–35 60. Papo D, Zanin M, Buldú JM (2014) Reconstructing functional brain networks: have we got the basics right? Front Human Neurosci 8(February):8–11. https://doi.org/10.3389/fnhum. 2014.00107 61. Ahuja RK, Magnanti TL (2018) Network flows: theory, algorithms, and applications. In: Pearson Education (ed) Network flows 62. Wilson RJ (1979) Introduction to graph theory. Pearson Education India 63. Lee Rodgers J, Nicewander WA (1988) Thirteen ways to look at the correlation coefficient. Am Stat 42(1):59–66 64. Kelley DJ, Farhoud M, Meyerand ME, Nelson DL, Ramirez LF, Dempsey RJ, Davidson RJ (2007) Creating physical 3D stereolithograph models of brain and skull. PLoS ONE 2(10):e1119. https://doi.org/10.1371/journal.pone.0001119 65. Meskaldji DE, Vasung L, Romascano D, Thiran JP, Hagmann P, Morgenthaler S, Van De Ville D (2015) Improved statistical evaluation of group differences in connectomes by screening-filtering strategy with application to study maturation of brain connections between childhood and adolescence. NeuroImage 108:251–264. https://doi.org/10.1016/j.neuroimage. 2014.11.059 66. Shaw JC (1984) Correlation and coherence analysis of the EEG: a selective tutorial review. Int J Psychophysiol 1(3):255–266. https://doi.org/10.1016/0167-8760(84)90045-x 67. Sifuzzaman M, Islam M, Ali MZ (2009) Application of wavelet transform and its advantages compared to Fourier transform 68. Pesaran B, Vinck M, Einevoll GT, Sirota A, Fries P, Siegel M, Truccolo W, Schroeder CE, Srinivasan R (2018) Investigating large-scale brain dynamics using field potential recordings: analysis and interpretation. Nat Neurosci 1–17. https://doi.org/10.1038/s41593-018-0171-8 69. Baccala L, Sameshima K (2001) Partial directed coherence: a new concept in neural structure determination. Biol Cybern 84(1):463–474. https://doi.org/10.1007/PL00007990 70. Sun FT, Miller LM, D’Esposito M (2004) Measuring interregional functional connectivity using coherence and partial coherence analyses of FMRI data. NeuroImage 21(2):647–658. https://doi.org/10.1016/j.neuroimage.2003.09.056 71. Bowyer Susan M (2016) Coherence a measure of the brain networks: past and present. Neuropsychiatric Electrophysiol 2(1):1. https://doi.org/10.1186/s40810-015-0015-7 72. Carmona J, Suarez J, Ochoa J (2017) Brain functional connectivity in Parkinson’s disease— EEG resting analysis. In: VII Latin American congress on biomedical engineering CLAIB 2016, Bucaramanga, Santander, Colombia, 26–28 Oct 2016, pp 185–188 73. Engel AK, Fries P (2016) Neuronal oscillations, coherence, and consciousness. Neurol Conciousness (Elsevier Ltd.). https://doi.org/10.1016/B978-0-12-800948-2.00003-0 74. Vergara JR, Estévez PA (2014) A review of feature selection methods based on mutual information. Neural Comput Appl 24(1):175–186. https://doi.org/10.3762/bjnano.7.4 75. MacKay JC (2003) Information theory, inference, and learning algorithms. Cambridge University Press. https://doi.org/10.1016/S0020-7063(14)00055-7

Critical Elements for Connectivity Analysis of Brain Networks

103

76. Cover TM, Thomas JA (2012) Elements of information theory. Wiley 77. Polani D (2013) Kullback-Leibler divergence. Springer 15:1087–1088. https://doi.org/10. 1016/0378-4754(88)90061-4 78. Shlens J (2014) Notes on Kullback-Leibler divergence and likelihood. ArXiv 1–4. https://doi. org/10.1108/JKM-06-2014-0253 79. James RG, Barnett N, Crutchfield JP (2016) Information flows? A critique of transfer entropies. Phys Rev Lett 116(23):1–5. https://doi.org/10.1103/PhysRevLett.116.238701 80. Schreiber Thomas (2006) Measuring information transfer. Phys Rev Lett 85(2):461. https:// doi.org/10.1103/PhysRevLett.85.461 81. Bossomaier T, Barnett L, Harré M, and Lizier JT (2016) An introduction to transfer entropy. Springer 82. Lindner M, Vicente R, Priesemann V, Wibral M (2011) TRENTOOL: a matlab open source toolbox to analyse information flow in time series data with transfer entropy. BMC Neurosci 12. https://doi.org/10.1186/1471-2202-12-119 83. Nolte G, Ziehe A, Kramer N, Popescu F, Muller K-R (2008) Comparison of granger causality and phase slope index. Nips 6:267–276 84. Cohen MX (2015) Effects of time lag and frequency matching on phase-based connectivity. J Neurosci Methods 250:137–146. https://doi.org/10.1016/j.jneumeth.2014.09.005 85. Maris E, Fries P, van Ede F (2016) Diverse phase relations among neuronal rhythms and their potential function. Trends Neurosci 39(2):86–99. https://doi.org/10.1016/j.tins.2015.12.004 86. Friston KJ, Litvak V, Oswal A, Razi A, Stephan KE, Van Wijk BCM, Ziegler G, Zeidman P (2016) Bayesian model reduction and empirical Bayes for group (DCM) studies. NeuroImage 128:413–431. https://doi.org/10.1016/j.neuroimage.2015.11.015 87. Kuznetsov YA (2013) Elements of applied bifurcation theory, vol 112. Springer Science & Business Media 88. Marreiros G, Santos R, Ramos C, Neves J (2010) Context-aware emotion-based model for group decision making. IEEE Intell Syst 25(2):31–39. https://doi.org/10.1109/MIS.2010.46 89. Brockwell PJ, Davis RA (1998) Time series: theory and methods. J Am Stat Assoc 92 (Springer, New York). https://doi.org/10.2307/2965440 90. Haufe S, Nikulin VV, Müller KR, Nolte G (2013) A critical assessment of connectivity measures for EEG data: a simulation study. NeuroImage 64(1):120–133. https://doi.org/10. 1016/j.neuroimage.2012.09.036 91. Stephan KE, Friston KJ (2010) Analyzing effective connectivity with functional magnetic resonance imaging. Wiley Interdisc Rev: Cogn Sci 1(3):446–459. https://doi.org/10.1002/ wcs.58 92. Wig GS, Schlaggar BL, Petersen SE (2011) Concepts and principles in the analysis of brain networks. Ann N Y Acad Sci 1224(1):126–146. https://doi.org/10.1111/j.1749-6632.2010. 05947.x 93. Honey CJ, Tter RK, Breakspear M, Sporns O (2007) Network structure of cerebral cortex shapes functional connectivity on multiple time scales. Proc Natl Acad Sci 104(24):10240– 10245. https://doi.org/10.1073/pnas.98.2.676 94. Sporns O (2013) Network attributes for segregation and integration in the human brain. Curr Opin Neurobiol 23(2):162–171. https://doi.org/10.1016/j.conb.2012.11.015 95. Stam CJ, Reijneveld JC (2007) Graph theoretical analysis of complex networks in the brain. Nonlinear Biomed Phys 1:1–19. https://doi.org/10.1186/1753-4631-1-3 96. Bassett DS, Bullmore ED (2006) Small-world brain networks. Neuroscientist 12(6):512–523. https://doi.org/10.1177/1073858406293182 97. Khadem A, Hossein-Zadeh GA, Khorrami A (2016) Long-range reduced predictive information transfers of autistic youths in EEG sensor-space during face processing. Brain Topogr 29(2):283–295. https://doi.org/10.1007/s10548-015-0452-4 98. Stephan KE, Friston KJ, Frith CD (2009) Dysconnection in Schizophrenia: from abnormal synaptic plasticity to failures of self-monitoring. Schizophr Bull 35(3):509–527. https://doi. org/10.1093/schbul/sbn176

104

J. Faber et al.

99. Barker K, Ramirez-Marquez JE, Rocco CM (2013) Resilience-based network component importance measures. Reliab Eng Syst Saf 117:89–97. https://doi.org/10.1016/j.ress.2013. 03.012 100. Vella, D, Zoppis I, Mauri G, Mauri P, Di Silvestre D (2017) From protein-protein interactions to protein co-expression networks: a new perspective to evaluate large-scale proteomic data. Eurasip J Bioinf Syst Biol 2017(1). https://doi.org/10.1186/s13637-017-0059-z 101. Wang XF (2002) Complex networks: topology, dynamics and synchronization. Int J Bifurcat Chaos 12(05):885–916 102. Strogatz SH (2001) Exploring complex networks. Nature 410(6825):268–276. https://doi.org/ 10.1038/35065725 103. Rajula HS, Reddy MMM, Fanos V (2018) Scale-free networks in metabolomics. Bioinformation 14(03):140–144. https://doi.org/10.6026/97320630014140 104. Ouma WZ, Pogacar K, Grotewold E (2018) Topological and statistical analyses of gene regulatory networks reveal unifying yet quantitatively different emergent properties. PLoS Comput Biol 14(4):1–17. https://doi.org/10.1371/journal.pcbi.1006098 105. Broder A, Kumar R, Maghoul F, Raghavan P, Rajagopalan S, Stata R, Tomkins A, Wiener J (2000) Graph structure in the web. Comput Netw 33(1):309–320. https://doi.org/10.1016/ S1389-1286(00)00083-9 106. Eguíluz VM, Chialvo DR, Cecchi GA, Baliki M, Vania Apkarian A (2005) Scale-free brain functional networks. Phys Rev Lett 94(1):1–4. https://doi.org/10.1103/PhysRevLett. 94.018102 107. He BJ, Zempel JM, Snyder AZ, Raichle ME (2010) The temporal structures and functional significance of scale-free brain activity. Neuron 66(3):353–369. https://doi.org/10.1016/j.neu ron.2010.04.020 108. van den Heuvel MP, Stam CJ, Boersma M, Hulshoff Pol HE (2008) Small-world and scalefree organization of voxel-based resting-state functional connectivity in the human brain. NeuroImage 43(3):528–539. https://doi.org/10.1016/j.neuroimage.2008.08.010 109. Watts DJ, Strogatz SH (1998) Watts-1998-collective dynamics of ‘small-world’ 393(June):440–442. https://doi.org/10.1038/30918 110. Hallquist MN, Hillary FG (2018) We Thank Zach Ceneviva, Allen Csuk, Richard Garcia, Melanie Glatz, and Riddhi Patel for their work collecting, organizing, and coding references for the literature review and manuscript. Netw Neurosci 111. Sporns O, Zwi JD (2004) The small world of the cerebral cortex. Neuroinformatics 2(2):145– 162. https://doi.org/10.1385/NI:2:2:145 112. Milo R, Shen-Orr S, Itkovitz S, Kashtan N, Chlovskii D, Alon U (2002) Network motifs : simple building blocks of complex networks. American Association for the Advancement of Science Stable. http://www.jstor. Science 298(5594):824–827 113. Horn Andreas, Ostwald Dirk, Reisert Marco, Blankenburg Felix (2014) The structuralfunctional connectome and the default mode network of the human brain. NeuroImage 102(P1):142–151. https://doi.org/10.1016/j.neuroimage.2013.09.069 114. Thompson WH, Brantefors P, Fransson P (2017) From static to temporal network theory: applications to functional brain connectivity. Netw Neurosci 1(2):69–99. https://doi.org/10. 1162/NETN_a_00011 115. Kivelä M, Arenas A, Barthelemy Marc, Gleeson JP, Moreno Y, Porter MA (2014) Multilayer networks. J Complex Netw 2(3):203–271. https://doi.org/10.1093/comnet/cnu016 116. Deco G, Jirsa VK, McIntosh AR (2011) Emerging concepts for the dynamical organization of resting-state activity in the brain. Nat Rev Neurosci 12(1):43–56. https://doi.org/10.1038/ nrn2961 117. Calhoun VD, Miller R, Pearlson G, Adali T (2014) The chronnectome: time-varying connectivity networks as the next frontier in FMRI data discovery. Neuron 84(2):262–274. https:// doi.org/10.1016/j.neuron.2014.10.015 118. Doron KW, Bassett DS, Gazzaniga MS (2012) dynamic network structure of interhemispheric coordination. Proc Natl Acad Sci 109(46):18661–18668. https://doi.org/10.1073/pnas.121640 2109

Critical Elements for Connectivity Analysis of Brain Networks

105

119. Ma S, Calhoun VD, Phlypo R, Adali T (2014) Dynamic changes of spatial functional network connectivity in healthy individuals and Schizophrenia patients using independent vector analysis. NeuroImage 90:196–206. https://doi.org/10.1016/j.neuroimage.2013.12.063 120. Cole MW, Reynolds JR, Power JD, Repovs G, Anticevic A, Braver TS (2013) Multi-task connectivity reveals flexible hubs for adaptive task control. Nat Neurosci 16(9):1348–1355. https://doi.org/10.1038/nn.3470 121. Davison EN, Schlesinger KJ, Bassett DS, Lynall ME, Miller MB, Grafton ST, Carlson JM (2015) Brain network adaptability across task states. PLoS Comput Biol 11(1). https://doi. org/10.1371/journal.pcbi.1004029 122. De Domenico M, Sasai S, Arenas A (2016) Mapping Multiplex hubs in human functional brain networks. Front Neurosci 10(Jul):1–14. https://doi.org/10.3389/fnins.2016.00326 123. Vaiana M, Muldoon SF (2018) Multilayer brain networks. J Nonlinear Sci (Sept 2017) 1–23. https://doi.org/10.1007/s00332-017-9436-8 124. Yu R, Chien YL, Wang HLS, Liu CM, Liu CC, Hwang TJ, Hsieh MH, Hwu HG, Tseng WYI (2014) Frequency-specific alternations in the amplitude of low-frequency fluctuations in Schizophrenia. Hum Brain Mapp 35(2):627–637. https://doi.org/10.1002/hbm.22203 125. Garcés P, Pereda E, Hernández-Tamames JA, Del-Pozo F, Maestú F, Pineda-Pardo JA (2016) Multimodal description of whole brain connectivity: a comparison of resting state MEG, FMRI, and DWI. Hum Brain Mapp 37(1):20–34. https://doi.org/10.1002/hbm.22995 126. Battiston F, Nicosia V, Chavez M, Latora V (2017) Multilayer motif analysis of brain networks. Chaos 27(4). https://doi.org/10.1063/1.4979282 127. D’Agostino G, Scala A (2014) Networks of networks: the last frontier of complexity, vol 97. https://doi.org/10.1007/978-3-319-03518-5 128. Bassett, DS, Khambhati AN, Grafton ST (2016) Emerging frontiers of neuroengineering: a network science of brain connectivity, pp 327–352 (March). https://doi.org/10.1146/annurevbioeng-071516-044511 129. De Domenico M, Solé-Ribalta A, Cozzo E, Kivelä M, Moreno Y, Porter MA, Gómez S, Arenas A (2014) Mathematical formulation of multilayer networks. Phys Rev X 3(4):1–15. https://doi.org/10.1103/PhysRevX.3.041022 130. Lee KM, Min B, Goh KI (2015) Towards real-world complexity: an introduction to multiplex networks. Eur Phys J B 88(2). https://doi.org/10.1140/epjb/e2015-50742-1 131. Supekar, K, Menon V, Rubin D, Musen M, Greicius MD (2008) Network analysis of intrinsic functional brain connectivity in Alzheimer’s disease. PLoS Comput Biol 4(6). https://doi.org/ 10.1371/journal.pcbi.1000100 132. Lynall M-E, Bassett DS, Kerwin R, McKenna PJ, Kitzbichler M, Muller U, Bullmore E (2010) Functional connectivity and brain networks in Schizophrenia. J Neurosci 30(28):9477–9487. https://doi.org/10.1523/JNEUROSCI.0333-10.2010 133. Ganella EP, Seguin C, Bartholomeusz CF, Whittle S, Bousman C, Wannan CM, Zalesky A (2018) Default-mode ntwork activity distinguishes Alzheimer’s disease from healthy aging: evidence from functional MR. Schizophr Res 193:284–292. https://doi.org/10.1016/j.schres. 2017.07.014 134. Zhang Z, Liao W, Chen H, Dante Mantini J, Ding R, Qiang X, Wang Z et al (2011) Altered functional-structural coupling of large-scale brain networks in idiopathic generalized epilepsy. Brain 134(10):2912–2928. https://doi.org/10.1093/brain/awr223 135. Hogan RE (2018) Malformations of cortical development: a structural and functional MRI perspective. Epilepsy Currents 18(2):92–94. https://doi.org/10.5698/1535-7597.18.2.92 136. Tecchio F, Cottone C, Porcaro C, Cancelli A, Di Lazzaro V, Assenza G (2018) Brain functional connectivity changes after transcranial direct current stimulation in epileptic patients. Front Neural Circ 12(May):1–7. https://doi.org/10.3389/fncir.2018.00044 137. Stam CJ, Jones BF, Nolte G, Breakspear M, Scheltens Ph (2007) Small-world networks and functional connectivity in Alzheimer’s disease. Cereb Cortex 17(1):92–99. https://doi.org/10. 1093/cercor/bhj127 138. Kabbara A, Eid H, El Falou W, Khalil M, Wendling F, Hassan M (2018) Reduced integration and improved segregation of functional brain networks in Alzheimer’s disease. J Neural Eng 15(2). https://doi.org/10.1088/1741-2552/aaaa76

106

J. Faber et al.

139. Heuvel MP, Den V, Sporns O, Collin G, Scheewe T, Mandl RCW, Cahn W, Goni J, Hulshoff Pol HE, Kahn RS (2013) Abnormal rich club organization and functional brain dynamics in Schizophrenia. JAMA Psychiatry 70(8):783–792. https://doi.org/10.1001/jamapsychiatry. 2013.1328

Jean Faber is graduated in Physics at the Federal University of Juiz de Fora (UFJF), Brazil, and he received his Master and Ph.D. degrees in Computational Modeling in the field of Quantum Information and Computation, at the National Laboratory for Scientific Computing (LNCCMCTI), Brazil. He starts his studies in Neuroscience as a postdoctoral fellow at the International Institute of Neuroscience of Natal—Edmond Lily Safra (IINN-ELS), investigating the relationship between memory consolidation and sleep, by evaluating neuro-electrophysiological patterns during wake-sleep cycles of rats and human beings. He also did a postdoc at the Commissariat à l’Energie Atomique et aux Energies Alternatives (CEA/LETI/CLINATEC), supported by the Fondation Nanoscience (Grenoble/France) in a project on Brain Computer Interface. Currently, Jean Faber is Adjunct Professor at the Federal University of São Paulo (UNIFESP), in the Department of Neurology and Neurosurgery/Discipline of Neuroscience. He has interests in the areas of Neuronal Signal Analysis, Neuro-connectivity, Neuro-Biofeedback and Brain Computer Interface. Priscila C. Antoneli is graduated in Biomedical Engineering from Federal University of Uberlândia (UFU), Brazil. She received her master’s degree in Electric Engineering in the Biomedical Engineering area from University of Campinas (UNICAMP), Brazil. Currently, she is Ph.D. student at Biocheminstry Department of Federal University of São Paulo (UNIFESP), Brazil, where she studies functional connectivity in in vitro neuronal networks. She has experience in signal processing, in vitro electrophysiology, biomedical instrumentation, animal experimentation and in the clinical engineering area. Noemi S. Araújo is graduated in Biomedical Engineering at the Instituto de Ciência e Tecnologia of the Universidade Federal de São Paulo (UNIFESP), Brazil. She received her master’s degree in Neurology and Neuroscience in the Escola Paulista de Medicina of the UNIFESP, in which performed recurrence analysis for the epileptiform-like activity recorded from hipocampal tissues of patients with mesial temporal lobe epilepsy (MTLE). Currently, she is Ph.D. student at Neurology and Neurosurgery Department of UNIFESP, Brazil, where she studies the relationship of structural and functional connectivity in the hippocampal subfields in experimental model of MTLE. She has experience in signal processing, in vitro electrophysiology, animal experimentation and statistics. Daniel J. L. L. Pinheiro is graduated in Biomedical Engineering at the Instituto de Ciência e Tecnologia of the Universidade Federal de São Paulo (UNIFESP), Brazil. He received his master’s degree in Neurology and Neuroscience in the Escola Paulista de Medicina of the UNIFESP, investigating the modulation of brain rhythms in epileptiform activities. Currently, he is Ph.D. student at Neurology and Neuroscience Graduate Program of UNIFESP, Brazil, where he studies functional and effective neural connectivity in patients with spinal cord injury that were submitted to neurostimulator transplant in peripheric nerves. He has experience in signal processing, biomedical images, in vivo electrophysiology, biomedical instrumentation and statistics. Esper Cavalheiro is full professor at the Department of Neurology and Neurosurgery of the Escola Paulista de Medicina da Universidade Federal de São Paulo. He is a full member of the Brazilian Academy of Sciences, of the International League Against Epilepsy, of the International Bureau of Epilepsy. He was president of the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) of Brazil and secretary of science and technology and programs of the

Critical Elements for Connectivity Analysis of Brain Networks

107

Ministry of Science and Technology in Brazil. He conducts research in neurosciences focusing on the mechanisms underlying major neurological disorders. After his retirement in the UNIFESP (September 2018), he assumed the position of researcher at the National Center for Research in Energy and Materials (CNPEM) in Campinas/SP, Brazil.

Brain Optical Imaging

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art with Emphasis on Pre-clinical and Clinical Studies Ron D. Frostig

1 Introduction and Background Intrinsic signal optical imaging (ISOI) remains one of the most exciting functional imaging techniques for functional mapping. It is routinely employed for basic research and has also been slowly adopted in recent years in preclinical and clinical research. For detailed recent reviews of ISOI see [1–3]. The current review, beyond some basic introduction and background, will only highlight recent developments and topics that were not covered in these reviews.

1.1 ISOI Advantages and Limitations ISOI is a relatively inexpensive, no-contact, wide-field imaging method that only employs light and therefore external contrast agents, probes, or indicators are not needed. There is also no need for a microscope as the camera can be directly mounted above the brain. Utilization of ISOI is based on the finding that when the brain is illuminated evoked neuronal activity causes characteristic changes in the intensity of the light that is reflected back from the brain. Accordingly, patterns of evoked activity from the living brain can be recorded or imaged by sensitive optical systems [4–6]. These stimulus-evoked reflectance changes are referred to as intrinsic signals to differentiate them from other optical signals obtained using extrinsic probes such as voltage sensitive dyes or calcium indicators. ISOI spatial resolution is excellent (below 100 mµ). Its temporal resolution is about 200 ms yet a recent report claimed R. D. Frostig (B) Departments of Neurobiology and Behavior and Biomedical Engineering, The Center for the Neurobiology of Learning and Memory, University of California, Irvine, CA 92697, USA e-mail: [email protected] © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_5

111

112

R. D. Frostig

it to be as fast as 80 ms [7]. A major, important advantage of ISOI in rodents is that it is non-invasive to the imaged brain as light can illuminate the brain through a thinned skull (rats) or intact skull (mice) enabling unperturbed, reliable imaging of neuronal activity. The ability to image through the skull also facilitates long term chronic ISOI imaging from the same rodent [8–10]. While ISOI imaging through a thinned skull has been also reported for cats [4], skull and dura removal are common in cats and non-human primates. Since ISOI is only based on measuring neuronal activity dependent changes in reflection from the illuminated brain, it entails that one can employ high illumination levels that are only shot-noise limited—the optimal way for illumination as shotnoise is uncontrollable and its relative contribution is progressively reduced with employment of higher illumination levels. This also entails that unlike contrast agent or indicator-based imaging experiments (e.g., voltage sensitive dye optical imaging, multi photon imaging) ISOI experiments are basically unlimited in their duration because there is no progressive bleaching of an indicator. One clear limitation of ISOI is the relatively shallow cortical depth that it can probe. This is due to the exponential loss of photons with cortical depth and a further exponential loss of photons in the return path from cortical depth to the surface. Therefore, and depending on the thickness of the cortex, ISOI is inherently biased for functional imaging of activity in the upper layers of the cortex. The depth of light penetration also depends on the wavelength of illumination as longer wavelengths penetrate deeper into the cortical tissue. However, the amplitude of intrinsic signals is stronger with short wave illumination and gets progressively weaker with longer wave illumination and this wavelength dependence should be considered when deciding which wavelength to employ; other considerations regarding illumination wavelength choice are described below.

1.2 ISOI Set-Up The basic components of a typical ISOI system are simple as compared to other functional imaging methods: a cooled camera system, lens, stable source of illumination and a computer. Camera systems have become progressively faster and sophisticated, enabling multiple imaging alternatives. Stable illumination, which used to necessitate expensive stabilizers, has been replaced by battery powered inexpensive bright LEDs. And computers have become much faster and progressively inexpensive. ISOI has been typically employed using scientific CCD camera systems that excelled in their spatial resolution, quantum efficiency, high signal/noise and exceptional linearity. The relatively slow operation of scientific CCD systems has not hindered ISOI as the evoked intrinsic signals are inherently slow and ~100 ms frames are adequate for most research questions and in many cases 100 ms frames were even further averaged to 500 ms frames prior to analysis. In recent years, however, cheaper than CCD, remarkably faster, and closing the gap in quantum efficiency, scientific

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

113

CMOS camera systems offer an attractive alternative to the long hegemony of scientific CCD camera systems, especially with the recent advance of back illuminated scientific CMOS camera systems. A definitive comparative test is still needed to settle the issue on whether scientific CMOS camera system could replace scientific CCD camera systems at the high illumination levels that are typically employed in ISOI experiments.

1.3 The Composition of an Intrinsic Signal Stimulus evoked ISOI response at the cortical level typically consist of a sequence of three clear and almost independent phases, described both in anesthetized and awake animals, and their presence serves as a good indicator for the health of the imaged brain (Figs. 1 and 2). These phases are named after similar phases in fMRI literature and have been thoroughly investigated in the rodent cortex by [11]. When applying red illumination (605–660 nm range) these include a short ‘initial dip’ phase below baseline level, followed by an ‘overshoot’ phase over baseline level (equivalent to the blood-oxygenated-level-dependent or BOLD fMRI signal), and followed up by a final, below baseline level, ‘undershoot’ phase. The entire sequence takes more than 10 s for completion but may sometimes have follow-up phases beyond the triphasic signal if no new stimulus is delivered (see example in Fig. 2; more examples can be seen in [11]). The initial dip typically peaks at ~1.5 s after stimulation, the overshoot after ~4 s and the undershoot after ~9 s. These latencies to peak could potentially change with changes in amplitude, frequency, duration and speed of the stimulation, but the potential effects of such changes on the triphasic signal have yet to be systematically investigated. In terms of amplitude the overshoot has the strongest amplitude followed by the undershoot and followed by the initial dip. Indeed, the typical initial dip peak evoked amplitude in rodents is only at the 10−4 –10−3 range of brain illumination. Despite its weaker amplitude, when it comes to mapping evoked neuronal activity in the brain, the initial dip is the most advantageous phase. Not only is it spatially correlated with the underlying evoked activity, it is also more robust to artifacts originating from evoked blood vessels activation that can distort functional mapping (see below). The presence of an additional evoked signal, a rapidly evoked scattering signal (peaking less than 100 ms following stimulation) has been documented both in rodents and in humans, for recent review see [12].

1.4 The Underlying Sources of the Intrinsic Signal ISOI has several underlying sources, but no clear consensus has been reached regarding their relative contributions at the different evoked phases and the potential effects of their spatiotemporal interactions. Underlying sources include

114

R. D. Frostig

Fig. 1 Single stimulus delivery and trial collection using long data collection to capture the triphasic intrinsic signal. Top: Schematic of visualizing intrinsic signal activity following single mechanical stimulation of the C2 whisker for 1 s, depicted as dark background. Post-stimulus frames are converted to fractional change (FC) values relative to pre-stimulus data. Bottom: Visualization and plot of intrinsic signal activity for a representative rat. A single stimulus delivery evoked a triphasic signal. Activity was relatively stable prior to stimulus onset, evidenced by a homogeneous pre-stimulus image with only subtle contributions from blood vessels (light or dark streaks). The post-stimulus intrinsic signal time course from the location of peak initial dip activity was extracted and plotted. In general data is acquired in 100 ms frames and averaged to 500 ms frames to improve signal/noise ratio. Fractional change (FC) values for a given post-stimulus 500-ms frame were calculated relative to the 500-ms frame collected immediately prior to stimulus onset (baseline frame) on a pixel-by-pixel basis. Then, to generate images of post-stimulus activity, an 8-bit linear grayscale mapping function was applied to the FC values, with an FC value of 0 (no change from baseline) mapped to middle gray shade and an arbitrary threshold of ±2.5 × 10−4 FC (±0.025%) from 0 used such that evoked ISOI initial dip (and also ISOI undershoot) and ISOI overshoot signal phase would appear as black or white, respectively, in generated images. Modified with permission from [23]. The identical grayscale fractional change bar also applies to Figs. 2, 3 and 4

evoked changes in hemodynamic sources that are connected to evoked neurovascular coupling processes. These include reduced hemoglobin (HbR), oxygenated hemoglobin (HbO2 ), total hemoglobin (HbT), blood volume and blood flow—all acting as internal contrast agents that can be utilized for functional mapping. Such evoked hemodynamic changes are dominant in the blue-green part of the spectrum due to strong hemoglobin absorption at this spectral range. In addition, changes in tissue scattering due to: ion and water movement in and out of active neurons, astrocytes and axons; capillary expansion; neuronal expansion; and neurotransmitter

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

115

Fig. 2 Single stimulus delivery and trial collection using long data collection to capture the triphasic intrinsic signal as acquired from 10 rats. Each image is based on 128 trials. Analysis and gray scale as in Fig. 1. Top: Despite temporal variance the triphasic signal can be visualized in all 10 rats. Note an additional ‘overshoot’ in rat #10. Bottom: Plot of the intrinsic signal (mean and standard error bars) obtained from the 10 rats. Note that the initial dip peaks ~1.5 s following stimulation; the overshoot following ~4.5 s and the undershoot following ~9 s. Modified with permission from [23]

release, have been described. The scattering signals are dominant in the near infrared part of the spectrum where they are not masked by strong hemoglobin absorption; for reviews see [3, 13]. Like the fMRI case [14], the underlying neuronal correlate of the intrinsic signal is evoked subthreshold (synaptic) activation, typically measured with microelectrodes as local field potential—LFP [15, 16] with some or no contribution from evoked suprathreshold (spiking) activity. However, claims that glia cells activity also contribute to the intrinsic signals have been reported [17, 18], but see [19]. The subthreshold dominance for ISOI is similar to its dominance in voltage-sensitive dyes (VSD) optical imaging [20], suggesting that evoked synaptic activity is underpinning these three popular functional imaging methods.

116

R. D. Frostig

1.5 Sources of Noise For ISOI the major noise sources are biological. Heart rate and respiration artifacts are minor in rodents as compared to higher mammals. The absence of heart rate artifacts in rodents is presumably due to their faster heart rate (relative to cats and monkeys) in combination with the lower temporal resolution (e.g., 500 ms frames) of a typical ISOI experiments. The reason respiration artifacts are minor when imaging in rodents is less clear; it may be due to the opportunity to image rodents through the thinned skull, thereby providing dampening of brain movement during breathing. In general, there are two types of biological noise whose magnitudes are much larger than that of stimulus-evoked intrinsic signals [11]: (1) global, spontaneous fluctuations and (2) local contributions overlying surface blood vessels. The amplitude of global, spontaneous fluctuations in intrinsic signals can be ten times stronger and occur on a slower time scale (oscillations of ~0.05–0.1 Hz or one complete cycle every 10–20 s) as compared to the stimulus-evoked intrinsic signals. Because stimulus delivery evokes only a small change in signal on top of the large spontaneous intrinsic signal fluctuations, successful imaging of stimulus-evoked intrinsic signals requires that these spontaneous fluctuations are somehow averaged out. As they are not time-locked to stimulus delivery, spontaneous intrinsic signals fluctuations can be minimized by averaging a set of stimulation trials. The number of imaging trials needed (32–64 trials) for sufficient capturing of the triphasic intrinsic signal is comparable to that of a typical single unit recording experiments. Any residual presence of spontaneous fluctuations can then be addressed at the level of data analysis. To better understand these spontaneous fluctuations, control trials have to be collected and analyzed in the same manner as stimulation trials. It was found that their presence can be substantial in magnitude and areal extent despite averaging across many trials. Fortunately, these spontaneous fluctuations are nonspecific both in the temporal and spatial domains, making them typically distinguishable from stimulus-evoked intrinsic signals, but not always [11]. Contributions from surface blood vessels within the imaged cortical region can also be time-locked to stimulus delivery and thus the averaging of stimulation trials is not effective in minimizing them. The degree of vessel contributions can depend on the illumination wavelength used, but vessel contributions typically follow a slower time course as compared to the initial dip. Thus, evoked vessel contributions can be minimized by limiting analysis to only data collected soon after stimulus onset [21]. An alternative approach for artifacts removal is the employment of fMRI-like data acquisition and Fourier data analysis where biological artifacts have clear signature at the frequency domain and can be filtered out as reviewed by [22].

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

117

1.6 Modes of Data Acquisition Data acquisition time typically depend on the desired signal/noise and contrast/noise ratios in addition to the research question. For example, if ISOI is only needed for obtaining a quick location of evoked activity, few trials could be sufficient. More trials are needed if a clear evoked functional representation is needed, one that its peak and areal extent are easily delineated and therefore quantifiable. Traditionally, experiments were designed to image all three phases of the evoked intrinsic signal (Figs. 1 and 2). Since a completion of the evoked triphasic intrinsic signal lasts more than 10 s, averaging experiments of 64 trials with incorporation of a random interstimulus interval could take about 20–25 min for completion. However, if only the initial dip is of interest, significantly shorter (×10 shorter) data acquisition times can be employed resulting in a clear increase in the efficiency of data acquisition [23], see Figs. 3 and 4. Further, the resulting evoked functional representation of such short experiments are typically stronger (amplitude) and larger (areal extent of cortical activation) compared to the longer triphasic experiments [23]. Therefore, a fast 2–4 min ISOI experiment with the addition of about 20 s for preliminary mapping, results in good signal/noise and contrast/noise cortical representation. This short ISOI data acquisition could be very important for intraoperative experiments in humans where there is a limited availability of time for functional imaging (see below). In

Fig. 3 Results from a representative rat when short data collection is applied to focus on the initial dip. Activity profile for multiple stimuli delivered at 4 s intervals. Plot: The post-stimulus intrinsic signal time course was extracted and plotted at the location of peak initial dip evoked by stim 1 delivery. Note that stim 1 of the 4-stimuli series evoked an initial dip plus onset of overshoot as typically observed for a single stimulus (compare to Figs. 1 and 2 plots). Upon delivery of stim 2–4, the signal was observed to reliably dip below baseline for each of stim 2–4 despite a strong rise in signal immediately prior to stimulus onset. Images: The same post-stimulus frames are visualized in reference to either the frame collected immediately prior to stim 1 (baseline 1; top row images) or the frame collected immediately prior to a particular stimulus delivery (stim 2, 3, or 4; baseline 2, 3, or 4, respectively; bottom row images). Note that an evoked dip in signal coinciding with a strong signal rise (as is the case for stim 2–4) can be effectively visualized as a dark activity area only when the frame immediately prior to each of stim 2–4 deliveries is used as baseline (bottom row images). Modified with permission from [23]

118

R. D. Frostig

Fig. 4 Activity profile for multiple stimuli delivered at 4 s intervals—summary of 10 rats. Average plot (means and standard errors bars) and visualization of the activity profile evoked by the 4-stimuli series are provided for 10 rats, each image is based on 64 trials. Notably, there was no change in image quality if jitter was added to the 4 s inter-stimuli intervals. Modified with permission from [23]

addition, from a behavioral perspective, delivery of stimuli with short inter-stimulus intervals is more akin to stimulation experienced by an awake, behaving subject, rather than a short stimulation every 10+ s. Finally, despite the enhanced amplitude and areal extent of a cortical representation in short imaging experiments, preliminary results show that only ~30% of the trials actually contribute to the final imaging results. Therefore, development of a rapid on-line algorithm for identification and removal of unsuccessful trials could definitely result in even faster and better results.

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

119

1.7 Imaging Preference Versus Imaging Spread Two complementary modes of evoked activity can be imaged when characterizing the functional organization of the cortex: preference versus spread. Imaging evoked preference allows addressing the question: what cortical regions respond preferentially to a given stimulus delivery? In addressing this question, ISOI is employed to map specific columnar systems (e.g., ocular dominance columns, orientation columns) because response preference to specific stimulus is typically organized in columns. Imaging evoked spread allows addressing a different question: what is the total cortical area activated by a specific stimulus delivery? In addressing such a question, the imaging target is the total cortical spread of evoked activity—also known as the functional representation of a specific stimulus. An important case of mapping the spread of cortical activity is when the stimulus delivery is spatially focused (‘point’ stimulation), e.g., single whisker, a pure tone, a discrete e.g., 1° × 1° visual stimulation. The resultant cortical activity spread following point-stimulation is known as the cortical point-spread. Characterizing the point-spread addresses the following question: how much cortex is activated by delivering a point-like stimulation to the peripheral sensory epithelium? For both modes of ISOI imaging, peak optical activation is always localized at the expected cortical location based on the topographic organization of sensory cortex (i.e., retinotopic, tonotopic and somatotopic organizations). Figure 5 shows 3-D ISOI visualization examples of single whisker functional

Fig. 5 Examples of 3-D ISOI visualization of whiskers functional representations in the rat somatosensory cortex. Left: C2 whisker functional representation (‘point spread’) following 5 Hz mechanical stimulation averaged from 37 rats. X-axis is 6 mm (3 mm to each side from the vertical black line). Y-axis is expressed in reflectance fractional change units (R/R) × 10−4 . Peak location is reliably located over the appropriate barrel. Note the areal extent of the spread of activation and the progressive smooth decline of evoked amplitude over cortical distance away from peak location. These results were confirmed by microelectrode array recordings. Right: Functional representation of 24 large whiskers (vibrissae) following 5 Hz simultaneous mechanical stimulation averaged from 10 rats. X and Y axes are identical to the left figure. Note the presence of only a single peak at the center of the functional representation and that the amplitude is comparable to the C2 whisker’s amplitude. These surprising characteristics were due to sublinear summation of each stimulated whisker point spread. These results were confirmed by microelectrode array recordings. Modified with permission from [15]

120

R. D. Frostig

representation (i.e., its point spread) and a case of simultaneous stimulation of 24 large whiskers (vibrissae). We advocate for increased awareness for the point-spread concept as it highlights different, fundamental structure–function properties of cortex, especially at the subthreshold (synaptic) level, that are not typically revealed when only cortical preference is targeted. A detailed review of the cortical point spread that highlights its cortical abundance, its electrophysiological and anatomical basis, and its major importance for cortical function has been recently published [2].

2 A Novel ISOI Frontier for Basic Research Beyond its continuous successful employment for studies devoted to mapping evoked cortical activity (either preference or spread), in recent years ISOI has also been employed to map large-scale correlated fluctuations in spontaneous cortical activity. The notion of ‘connectomics’ has expanded in recent years to encompass resting state functional connectivity. Resting state functional connectivity is defined in terms of temporally correlated intrinsic neural activity measured throughout the entire cortex as pioneered by fMRI researchers, for review see [24]. Such correlated spontaneous activity defines widely distributed topographies known as ‘resting state’ networks. The Culver lab has pioneered the use of ISOI for imaging and employing its results for analyzing resting state networks in the mouse cortex. ISOI-based detailed resting states maps of the mouse cortex have been described by this group [25], and have been used to investigate changes in such maps in pre-clinical animal models (see below).

3 Pre-clinical ISOI Studies Since its inception, ISOI has been mainly employed for basic research on the functional organization and plasticity of the developing and adult cortex—research that has revolutionized our understanding of cortical functional organization and its plasticity. In parallel, a growing number of pre-clinical studies have been successfully employed for various pre-clinical studies. Some successful examples are described here. The application of ISOI for pre-clinical epilepsy research, reviewed by [26–28] has been successful in precise localization of epileptic foci, mapping the spread of epilepsy in cortical tissue; and the discovery that blood flow level during epilepsy is inadequate to meet the metabolic demands of the epileptic tissue, resulting in a decrease of tissue oxygenation during epileptic attacks. In addition, resting state analysis in mice before and after the onset of epileptiform activity showed both decrease and increase in specific homologous correlations between cortical areas [29].

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

121

The application of ISOI for spreading depolarization (SD) research, for review see [30], has been successful in identifying the SD origin; in delineating its spread; in describing its dynamics; in describing how the presence of anatomical blocks such as sulci and pial vessels has changed the characteristics of the SD waves; and in identifying different phases of the SD by using different wavelengths of illumination. Regarding the spatiotemporal characteristics of SD propagation in cortex it was found that such spatiotemporal patterns include spiral and reverberating waves that can collide. The rat stroke model is an excellent example of ISOI employment for pre-clinical research. Specifically, the case of a permanent middle cerebral artery occlusion (pMCAo). It has been repeatedly demonstrated that the delivery of tactile stimulation that directly activates the MCA cortical ischemic territory during a 2 h window following pMCAo, results in functional and structural protection from impending ischemic stroke. Our findings [31–38] have revealed that such stimulus-based protection is achievable in anesthetized rats, unrestrained behaving rats, and old rats, and that such protection holds for at least 4 months (equivalent to 10–15 human years). This protection critically depends on functional pial collaterals (aka leptomeningeal collaterals) that connect the MCA with the other two major cortical arteries (anterior cerebral artery, and posterior cerebral artery) acting as alternative arterial blood supply sources for the permanently occluded MCA by supplying retrograde blood flow into the occluded MCA and therefore protect the ischemic MCA territory. For reasons that are not yet clear, such sensory-based protection fails in popular strains of mice (C57BL/6j,CD1) [39]. Without ISOI, discovering such unexpected protection from impending ischemic stroke following pMCAo would be nearly impossible. With intermittent whisker stimulation following pMCAo there was a clear functional recovery within 60– 90 min following pMCAo, as amplitude and areal extent of the stimulated whisker cortical representation were back to pre-pMCAo baseline levels (Fig. 6) [35]. In general, these ISOI findings were supported by microelectrode recordings, postmortem histology, and behavioral tests aimed to detect sensory-motor problems

Fig. 6 ISOI tracking functional recovery of an intermittently stimulated whisker functional representation following permanent MCA occlusion (pMCAo denoted by vertical, dashed line). X-axis time in minutes following pMCAo. Thirty minutes blocks of a stimulated whisker functional representation before (baseline) and after pMCAo are shown sequentially. Note minimal recovery within the first block (0–30) minutes after pMCAo. Full recovery to baseline level was imaged in the third block (60–90 min) after pMCAo. Dimensions: 5.7 mm × 5.7 mm; gray level bar scaled to fractional change (±0.025%) in reflectance. Modified with permission from [35]

122

R. D. Frostig

following stroke. Notably, having an ISOI map of a cortical functional representation of a given sensory stimulation (e.g., whisker) considerably enhances other techniques such as microelectrode recordings since the functional image is the best guide to accurate placement of microelectrodes. Detailed description of changes in resting state networks in a mouse model of transient MCAo have been described [40]. Regional changes in functional connectivity within the MCA territory were largely proportional to infarct volume. However, subcortical damage affected functional connectivity in the somatosensory cortex as much as larger infarcts of cortex and subcortex. The extent of injury correlated with cortical activations following electrical stimulation of the affected forelimb, and with functional connectivity in the somatosensory cortex. However, the authors have found that without correction for the change in hemodynamics following pMCAo, the resting state analysis could be distorted.

4 Intraoperative Applications for ISOI With its excellent spatial resolution and wide-field imaging abilities, relatively inexpensive functional imaging technology that is contact-free and label-free, it has been expected that the promising ISOI technique would be quickly adopted for intraoperative mapping of function. Intraoperative functional mapping is important for the ultimate goal of predicting when resection of cortical area will cause functional deficits— currently typically achieved by direct electrical stimulation mapping (ESM), the gold standard for intraoperative functional mapping. However, such adoption turned out to be slow and spotty; for review see [41–43]. The typical challenges that slow progress include: mechanical movements of the cortical tissue related to respiration and heart pulsation, evoked intrinsic signals from blood vessels; limited time for data acquisition; the lack of standard protocols for illumination, stimulation and analysis; the lack of a standard dedicated ISOI system; and the slow analysis time. Nevertheless, despite the limited number of intraoperative ISOI studies, and despite the relative low numbers of patient in these trials, partial or full solutions to these challenges have clearly advanced the feasibility of intraoperative ISOI. These include better registration and alignment algorithms to overcome movement artifacts originated from heart pulsation and respiration; implementation of faster algorithms for data acquisition and especially for data analysis; and active search for optimal illumination wavelengths and experimenting with different data acquisition times. An exemplary case is the paper by Sobottka et al., [44] demonstrating how issues with intraoperative ISOI can be resolved. The authors tested several CCD camera systems and chose an optimal one for their operating room; used large pool of patients with tumor lesions (N = 41); provided 2-D high-resolution maps of activity following peripheral stimulation to the surgeon within 12 min based on spectral analysis which improved the signal/noise ratio compared to relative difference algorithms; employed prolonged stimulation and rest periods that enabled reduction of cortical movement artifacts and glare artifacts; movement artifacts were further reduced by elastic registration

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

123

algorithms, local differences in direction and magnitude of the movements were compensated for; and determining that the use of 568 nm illumination light that highlights changes in blood volume (an isosbestic point for oxi- and deoxy-hemoglobin) was optimal for their trials. Notably, in this study the authors didn’t mount a transparent glass plate directly to the imaged cortical tissue, as employed by all previous intraoperative ISOI studies to dampen cortical movement artifacts, a procedure that might have affected the physiology of the cortex. Despite not employing the glass plate, the results showed specificity that was almost 100%. Indeed, one very encouraging general finding from many intraoperative experiments by different groups, has been the impressive validation of ISO with other techniques, e.g., electrophysiological recordings with microelectrodes, electrocortical stimulation of cortex (ESM), evoked potentials, and anatomical registration. In parallel to intraoperative studies aiming to image function to safely perform cortical operations, studies of intraoperative ISOI imaging for human cortical function have slowly progressed too, for review see [41–43]. The major limitation to progress of such studies is the limited time available for ISOI-based research in the operating room. Majority of such studies were performed in the somatosensory cortex by the Sato group with special emphasis on the representation of specific digits or parts of the face. The major finding of these studies has been that while each representation peaks at distinct cortical areas, there is clear overlap among these different representations, an overlap that is also imaged in the representation of digits in the somatosensory cortex of non-human primates [45] and in the representation of whiskers in barrel cortex of rats [15]. The importance of such overlap has been recently reviewed [2]. Indeed, such overlap could be detected even between different unimodal areas. A recent intraoperative ISOI was employed to image multisensory activation of the auditory cortex. The authors have demonstrated how the primary and secondary auditory cortex cold be activated by tactile stimulations [46]. Other cortical areas studied with ISOI were language areas where expansion the representation of task-related (object naming and word discrimination) was imaged. ISOI also highlighted the degree of overlap and separation of different languages in bilingual cortical representation, for review see [47]. Another area of intraoperative ISOI research field is the study of cortical epileptic activity, but only few studies have been reported, for review see [27, 48, 49]. These studies have demonstrated how the employment of ISOI contributes to better delineation of the epileptic foci; how hemodynamic changes could predict the onset of a seizures; and how drugs could significantly reduce the spread of cortical activation during epilepsy.

5 Summary and Outlook ISOI has been continuously improving its abilities with the adoption of progressively improved camera systems and new ways to acquire and analyze imaging data. Notwithstanding its continuous success in basic research and its clear promise

124

R. D. Frostig

for pre-clinical and clinical research, the adoption of ISOI in these areas has been slow. Despite major promising studies like the Sobottka et al. [44], and despite the impressive validation of the ISOI mapping results, the field of intraoperative ISOI remains spotty and experimental. A two-pronged approach is needed to turn this experimental situation into an operating room routine. On one hand more basic and preclinical animal research is needed to further improve our understanding of the optimal conditions for ISOI. For example, if fast ISOI imaging described above could be applied intraoperatively, a high-resolution quality map of a stimulus functional representation could be achieved in less than 5 min. On the other hand, more thorough, large-scale clinical studies like Sobottka et al. Sobottka et al. [44] are clearly needed. Intraoperative studies that involve collaboration between basic or preclinical researchers and the medical team would be ideal. An exciting recent ISOI research avenue is the employment of implantable CMOS sensors directly on top of the brain of rodents: rats [50] and mice [51]. These implantable CMOS chips are small (3.3 × 5.3 × 0.35 mm3 ), weigh only 0.02 gr and have built in LED light sources. Because they are in direct contact with the cortex there is no need for a lens. These chips were already tested and validated under anesthesia and their expected application in awake behaving conditions could improve our understanding of ISOI and optimize its application, as traditionally ISOI research has been mostly conducted in anesthetized animal models. As this review demonstrates, ISOI has repeatedly proven its unique advantages over other techniques and remains one of the most attractive and exciting tools for studying with high spatial resolution the mesoscale cortical domain as a standalone methodology or in combination with other techniques. We therefore continue to advocate further development and optimization of this technique for basic, preclinical and clinical research, as all these modes of research could inform each other for a more wide-spread future application of this promising technique. Acknowledgements The author would like to thank current and previous members of his lab, especially Cynthia Chen-Bee for her major contributions to the application and analysis of ISOI. Members include Susan Masino, Michael Kwon, Jonathan Bakin, Yehuda Dory, Neal Prakash, Daniel Polley, Barbara Brett-Green, Silke Penschuck, Eugene Kvasnak, Jimmy Stehberg, Ying Xiong, Christopher Lay, Melissa Davis, Aneeka Hancock, Yi Zhou, Nathan Jacobs, Brett Johnson, Ellen Wann, and Gabriel Hui. Supported by the Leducq Foundation grant (15CVD02).

References 1. Frostig RD, Chen-Bee CH (2012) The use of intrinsic signal optical imaging for mapping cortical function. In: Romain Brette AD (ed) Handbook of neuronal activity measurements. Cambridge University Press 2. Frostig RD, Chen-Bee CH, Johnson BA, Jacobs NS (2017) Imaging Cajal’s ‘neuronal avalanche’: How wide-field optical imaging of the point spread advanced the understanding of neocortical structure-function relationship. Neurophotonics 4(3):031217. https://doi.org/10. 1117/1.NPh.4.3.031217

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

125

3. Grinvald A, Sharon D, Omer D, Vanzetta I (2016) Imaging the neocortex functional architecture using multiple intrinsic signals: implications for hemodynamic-based functional imaging. Cold Spring Harbor protocols 2016 (3):pdb top089375. https://doi.org/10.1101/pdb.top089375 4. Frostig RD, Lieke EE, Ts’o DY, Grinvald A (1990) Cortical functional architecture and local coupling between neuronal activity and the microcirculation revealed by in vivo high-resolution optical imaging of intrinsic signals. Proc Natl Acad Sci USA 87(16):6082–6086 5. Grinvald A, Lieke E, Frostig RD, Gilbert CD, Wiesel TN (1986) Functional architecture of cortex revealed by optical imaging of intrinsic signals. Nature 324(6095):361–364 6. Ts’o DY, Frostig RD, Lieke EE, Grinvald A (1990) Functional organization of primate visual cortex revealed by high resolution optical imaging. Science 249(4967):417–420 7. Lu HD, Chen G, Cai J, Roe AW (2017) Intrinsic signal optical imaging of visual brain activity: tracking of fast cortical dynamics. NeuroImage 148:160–168. https://doi.org/10.1016/j.neuroi mage.2017.01.006 8. Masino SA, Frostig RD (1996) Quantitative long-term imaging of the functional representation of a whisker in rat barrel cortex. Proc Natl Acad Sci USA 93(10):4942–4947 9. Polley DB, Chen-Bee CH, Frostig RD (1999) Two directions of plasticity in the sensorydeprived adult cortex [see comments]. Neuron 24(3):623–637 10. Polley DB, Kvasnak E, Frostig RD (2004) Naturalistic experience transforms sensory maps in the adult cortex of caged animals. Nature 429(6987):67–71 11. Chen-Bee CH, Agoncillo T, Xiong Y, Frostig RD (2007) The triphasic intrinsic signal: implications for functional imaging. J Neurosci 27(17):4572–4586 12. Gratton G, Chiarelli AM, Fabiani M (2017) From brain to blood vessels and back: a noninvasive optical imaging approach. Neurophotonics 4(3):031208. https://doi.org/10.1117/1.NPh.4. 3.031208 13. Vanzetta I, Grinvald A (2008) Coupling between neuronal activity and microcirculation: implications for functional brain imaging. HFSP J 2(2):79–98. https://doi.org/10.2976/1.288 9618 14. Logothetis NK (2008) What we can do and what we cannot do with fMRI. Nature 453(7197):869–878 15. Chen-Bee CH, Zhou Y, Jacobs NS, Lim B, Frostig RD (2012) Whisker array functional representation in rat barrel cortex: transcendence of one-to-one topography and its underlying mechanism. Front Neural Circ 6:93. https://doi.org/10.3389/fncir.2012.00093 16. Frostig RD, Xiong Y, Chen-Bee CH, Kvasnak E, Stehberg J (2008) Large-scale organization of rat sensorimotor cortex based on a motif of large activation spreads. J Neurosci 28(49):13274– 13284 17. Gurden H, Uchida N, Mainen ZF (2006) Sensory-evoked intrinsic optical signals in the olfactory bulb are coupled to glutamate release and uptake. Neuron 52(2):335–345 18. Schummers J, Yu H, Sur M (2008) Tuned responses of astrocytes and their influence on hemodynamic signals in the visual cortex. Science 320(5883):1638–1643 19. Vincis R, Lagier S, Van De Ville D, Rodriguez I, Carleton A (2015) Sensory-evoked intrinsic imaging signals in the olfactory bulb are independent of neurovascular coupling. Cell Rep 12(2):313–325. https://doi.org/10.1016/j.celrep.2015.06.016 20. Grinvald A, Hildesheim R (2004) VSDI: a new era in functional imaging of cortical dynamics. Nat Rev 5(11):874–885 21. Chen-Bee CH, Polley DB, Brett-Green B, Prakash N, Kwon MC, Frostig RD (2000) Visualizing and quantifying evoked cortical activity assessed with intrinsic signal imaging. J Neurosci Meth 97(2):157–173 22. Kalatsky VA (2009) Fourier approach to optical imaging. In: Frostig RD (ed) In vivo optical imaging of brain function, 2nd edn. CRC Press, Boca Raton 23. Chen-Bee CH, Agoncillo T, Lay CC, Frostig RD (2010) Intrinsic signal optical imaging of brain function using short stimulus delivery intervals. J Neurosci Meth 187(2):171–182. S01650270(10)00031-2[pii], https://doi.org/10.1016/j.jneumeth.2010.01.009 24. Raichle ME (2015) The brain’s default mode network. Annu Rev Neurosci 38:433–447. https:// doi.org/10.1146/annurev-neuro-071013-014030

126

R. D. Frostig

25. White BR, Bauer AQ, Snyder AZ, Schlaggar BL, Lee JM, Culver JP (2011) Imaging of functional connectivity in the mouse brain. PLoS ONE 6(1):e16322. https://doi.org/10.1371/jou rnal.pone.0016322 26. Haglund MM, Hochman DW (2007) Imaging of intrinsic optical signals in primate cortex during epileptiform activity. Epilepsia 48(Suppl 4):65–74. https://doi.org/10.1111/j.15281167.2007.01243.x 27. Patel KS, Zhao M, Ma H, Schwartz TH (2013) Imaging preictal hemodynamic changes in neocortical epilepsy. Neurosurg Focus 34(4):E10. https://doi.org/10.3171/2013.1.FOCUS1 2408 28. Schwartz TH, Hong SB, Bagshaw AP, Chauvel P, Benar CG (2011) Preictal changes in cerebral haemodynamics: review of findings and insights from intracerebral EEG. Epilepsy Res 97(3):252–266. https://doi.org/10.1016/j.eplepsyres.2011.07.013 29. Guevara E, Pouliot P, Nguyen DK, Lesage F (2013) Optical imaging of acute epileptic networks in mice. J Biomed Opt 18(7):76021. https://doi.org/10.1117/1.JBO.18.7.076021 30. Zheng Z, Cao Z, Luo J, Lv Y (2018) Characterization of intrinsic optical signal during spreading depolarization. Neuropsychiatry 8(1):302–309 31. Davis MF, Lay C, Frostig RD (2013) Permanent cerebral vessel occlusion via double ligature and transection. J Vis Exp 77:1–8. https://doi.org/10.3791/50418 32. Davis MF, Lay CC, Chen-Bee CH, Frostig RD (2011) Amount but not pattern of protective sensory stimulation alters recovery after permanent middle cerebral artery occlusion. Stroke; J Cerebral Circ 42(3):792–798. STROKEAHA.110.607135[pii], https://doi.org/10.1161/STR OKEAHA.110.607135 33. Hancock AM, Lay CC, Davis MF, Frostig RD (2013) Sensory stimulation-based complete protection from ischemic stroke remains stable at 4 months post-occlusion of MCA. J Neurol Disorders 1(4):135. https://doi.org/10.4172/2329-6895.1000135 34. Lay CC, Davis MF, Chen-Bee CH, Frostig RD (2010) Mild sensory stimulation completely protects the adult rodent cortex from ischemic stroke. PLoS ONE 5(6):e11270. https://doi.org/ 10.1371/journal.pone.0011270 35. Lay CC, Davis MF, Chen-Bee CH, Frostig RD (2011) Mild sensory stimulation reestablishes cortical function during the acute phase of ischemia. J Neurosci 31(32):11495–11504. 31/32/11495[pii], https://doi.org/10.1523/JNEUROSCI.1741-11.2011 36. Lay CC, Davis MF, Chen-Bee CH, Frostig RD (2012) Mild sensory stimulation protects the aged rodent from cortical ischemic stroke after permanent middle cerebral artery occlusion. J Am Heart Assoc 1(4):e001255. https://doi.org/10.1161/JAHA.112.001255,jah355[pii] 37. Lay CC, Frostig RD (2014) Complete protection from impending stroke following permanent middle cerebral artery occlusion in awake, behaving rats. Eur J Neurosci 40(9):3413–3421. https://doi.org/10.1111/ejn.12723 38. Lay CC, Jacobs N, Hancock A, Zhou Y, Frostig RD (2013) Early stimulation treatment provides complete sensory-induced protection from ischemic stroke under isoflurane anesthesia. Eur J Neurosci 38:2445–2452. https://doi.org/10.1111/ejn.12217 39. Hancock AM, Frostig RD (2017) Testing the effects of sensory stimulation as a collateralbased therapeutic for ischemic stroke in C57BL/6J and CD1 mouse strains. PLoS ONE 12(9):e0183909. https://doi.org/10.1371/journal.pone.0183909 40. Bauer AQ, Kraft AW, Wright PW, Snyder AZ, Lee JM, Culver JP (2014) Optical imaging of disrupted functional connectivity following ischemic stroke in mice. NeuroImage 99:388–401. https://doi.org/10.1016/j.neuroimage.2014.05.051 41. Morone KA, Neimat JS, Roe AW, Friedman RM (2017) Review of functional and clinical relevance of intrinsic signal optical imaging in human brain mapping. Neurophotonics 4(3):031220. https://doi.org/10.1117/1.NPh.4.3.031220 42. Sato K, Nariai T, Momose-Sato Y, Kamino K (2017) Intraoperative intrinsic optical imaging of human somatosensory cortex during neurosurgical operations. Neurophotonics 4(3):031205. https://doi.org/10.1117/1.NPh.4.3.031205 43. Seth SA, Yanamadala V, Eskandar EN (2012) Intraoperative human functional brain mapping using optical intrinsic signal imaging. In: Advances in brain imaging INTECH

Intrinsic Signal Optical Imaging (ISOI): State-of-the-Art …

127

44. Sobottka SB, Meyer T, Kirsch M, Koch E, Steinmeier R, Morgenstern U, Schackert G (2013) Intraoperative optical imaging of intrinsic signals: a reliable method for visualizing stimulated functional brain areas during surgery. J Neurosurg 119(4):853–863. https://doi.org/10.3171/ 2013.5.JNS122155 45. Roe AW, Winberry JE, Friedman RM (2017) Study of single and multidigit activation in monkey somatosensory cortex using voltage-sensitive dye imaging. Neurophotonics 4(3):031219. https://doi.org/10.1117/1.NPh.4.3.031219 46. Zhou Q, Wang Y, Yi L, Tan Z, Jiang Y (2017) multisensory interplay within human auditory cortex: new evidence from intraoperative optical imaging of intrinsic signal. World Neurosurg 98:251–257. https://doi.org/10.1016/j.wneu.2016.10.100 47. Prakash N, Uhlemann F, Sheth SA, Bookheimer S, Martin N, Toga AW (2009) Current trends in intraoperative optical imaging for functional brain mapping and delineation of lesions of language cortex. NeuroImage 47(Suppl 2):T116–126. https://doi.org/10.1016/j.neuroimage. 2008.07.066 48. Haglund MM, Hochman DW (2004) Optical imaging of epileptiform activity in human neocortex. Epilepsia 45(Suppl 4):43–47. https://doi.org/10.1111/j.0013-9580.2004.04010.x 49. Schwartz TH (2005) The application of optical recording of intrinsic signals to simultaneously acquire functional, pathological and localizing information and its potential role in neurosurgery. Stereotact Funct Neurosurg 83(1):36–44. https://doi.org/10.1159/000085025 50. Haruta M, Sunaga Y, Yagamuchi T, Takehara H, Noda T, Sasagawa K, Tokuda T, Ohta J (2015) Intrinsic signal imaging of brain function using small implantable CMOS imaging device. Jpn J Appl Phys 54:1–6 51. Yagamuchi T, Takehara H, Sunaga Y, Haruta M, Motoyama M, Ohta H, Noda T, Sasagawa K, Tokuda T, Ohta J (2016) Implantable self-reset CMOS image sensor and its application to hemodynamic response detection in living mouse brain. Jpn J Appl Phys 55:1–6

Ron D. Frostig, Ph.D. is a Professor of Neurobiology and Behavior at the University of California, Irvine. His major research interests include the structure, function and plasticity of sensory neocortex in rodents, which he studies by employing several functional imaging techniques, electrophysiological techniques, and anatomical techniques. These studies led to the formulation of a novel way of understanding neocortical structure-function relationship and their plasticity. In recent years Dr. Frostig has also expanded to pre-clinical research following his discovery of tactile stimulation-based protection of neocortex from impending ischemic stroke. Dr. Frostig was trained in electrophysiology by Moshe Abeles (Hebrew University, Jerusalem) and Ron Harper (UCLA, Los-Angeles) and was also trained in developing and applying novel functional imaging methods by Amiram Grinvald and Torsten Wiesel (Rockefeller University, New York).

Implantable CMOS Fluorescent Imaging Devices Kiyotaka Sasagawa, Makito Haruta, Yasumi Ohta, Hironari Takehara, Takashi Tokuda, and Jun Ohta

1 Introduction Optical techniques are one of the most useful methods for elucidating how the brain works. In particular, by using a potential-sensitive dye [1] or a calcium probe [2], potential changes in cell membrane or Ca2+ ion concentration associated with neural activities can be converted to fluorescence intensity. Visualizing neural activity by combining these and microscopic techniques is widely used in recent neurosciences. In order to elucidate complex brain functions, it is also necessary to observe activities associated with behavior in a living state. Due to such a demand, an apparatus that enables in-vivo imaging of a small animal such as a mouse has been developed recently. The most basic method for optically observing the brain is to perform craniotomy and observe with a microscope or the like. Sufficient spatial resolution and fluorescence detection performance for observing cells can be realized using a commercially available microscope. In addition, if a confocal microscope or a multiphoton microscope is used, a very high contrast and a three-dimensional image can be obtained. However, since the apparatus is large and it is necessary to fix the head to be observed, it is difficult to observe in a freely moving state. As a technique for solving this problem, there is a case where a freely rotatable ball is placed under an observation target and pseudo-free action is performed [3].

K. Sasagawa (B) · M. Haruta · Y. Ohta · H. Takehara · T. Tokuda · J. Ohta Division of Materials Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan e-mail: [email protected]; [email protected] T. Tokuda Future Interdisciplinary Research of Science and Technology (FIRST), Tokyo Institute of Technology, Tokyo, Japan © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_6

129

130

K. Sasagawa et al.

Control wires Image sensor

Mouse Brain

Fig. 1 Configuration of a biological implant image sensor. A specially designed ultra-compact image sensor is placed very close to the observation area

On the other hand, a microscopic microscope that can be mounted on the head of a mouse has been developed. It consists of a relatively small image sensor that is installed in recent smartphones and a miniaturized optical system of a general fluorescence microscope, and realizes one-photon fluorescence imaging with a weight of about 2 g [4–6]. As a technique using a large microscope, a method using an optical fiber bundle has been developed [7]. Since restrictions on the dimensions of the microscope optical system are greatly relaxed, it is easy to handle multicolor fluorescence observation using a plurality of types of filters. On the other hand, the spatial resolution is basically determined by the core pitch of the optical fiber. Further, when the imaging area is increased, the rigidity of the optical fiber bundle that is a bundle of glass becomes a problem. The authors have designed a living body implant image sensor, which is a method different from these, and has developed a brain activity imaging system using it [8– 12]. Figure 1 shows a schematic configuration of the system. An ultra-small sensor equipped with the minimum necessary functions as an image sensor can be arranged in the very vicinity of an observation target. Thereby, since an optical system can be simplified, the dimension and weight of the insertion part can also be suppressed very low. Because of this feature, it is particularly suitable for deep observation of the living body [10] and simultaneous observation of a plurality of regions as compared with other methods. In addition, in in-vivo cross-sectional observation using a microscope system, a small prism or the like is often used, but the invasiveness is considerably increased. On the other hand, it can be said that cross-sectional observation is possible with low invasiveness by inserting a very thin image sensor. In this paper, the outline of the living implant image sensor is described, and the ultra-compact LED that is a light source for fluorescence observation to improve the performance and the hybrid filter that realizes high excitation light removal performance are explained.

Implantable CMOS Fluorescent Imaging Devices

131

2 Implantable CMOS Image Sensor A CMOS image sensor is one of solid-state image sensors and is mounted on various devices such as smartphones. What is mounted on a portable device is smaller than a camera-dedicated device, but it is difficult to apply as it is when considering the use of imaging in a living body. By designing a dedicated CMOS image sensor for the purpose of being implanted in a living body, it is possible to provide unique features that are difficult to achieve when using a normal image sensor. By limiting the functions to necessary and sufficient for the specific purpose, the size of the image sensor can be reduced. By making it as small as possible, the invasiveness when implanted or inserted into a living body is reduced. In other words, the imaging surface can be arranged in the very vicinity of the observation target region. By adopting such a so-called contact imaging technique, a fine structure can be observed without mounting a lens. Not using a lens leads to further reduction in size and weight. The weight of the device with a small image sensor shown in Fig. 2 is about 0.02 g for placement on the brain surface, which is about 1/1000 of an adult mouse [12]. With this, even in small animals such as mice, restrictions on behavior due to implantation can be greatly reduced. A general image sensor has a square plate shape, and a large number of electrodes for wiring are arranged on the four sides. Many wirings facilitate various settings, stable operation, high-speed driving etc. But it is inconvenience for implantation in a living body. When inserting into a deep part of a living body, the wiring is required to be arranged on only one side. Also, in an imaging experiment under freely moving condition, it is desirable that the number of wires be small in order to reduce the restriction on the movement of the target by the control signal line. Figure 2a shows an example of an image sensor that was designed. In the imaging area, 120 × 268 pixels of 7.5 µm square are arranged. That is, the imaging area is 0.9 mm × 2.0 mm. The width of the chip is about 1.0 mm, and the thickness is reduced to about 150 µm by polishing. The number of wires required for driving is four, and the electrodes for connection are arranged at one end of the chip. The Fig. 2 a Prototype biological implant image sensor. b Imaging example of blood vessel image on brain surface. Copyright (2014) The Japan Society of Applied Physics. Modified from Ref. [12] 5 mm

132

K. Sasagawa et al.

breakdown of the wiring is two for supplying power, one drive control line, and one output signal line. In the image sensor chip, a circuit for performing initialization at start-up and a scanner circuit for sequentially selecting pixels are mounted, and an operation with fewer wirings is possible. The luminance information read by each pixel is an analog signal but is finally converted into a digital signal and taken into the computer. The conversion circuit for digital signals is arranged outside the chip, and the power consumption and circuit dimensions of the embedded image sensor are kept low. As an example, in Fig. 2a, the image sensor is arranged on a flexible substrate using polyimide. The chip of the image sensor circuit is formed on silicon and has high rigidity, but by using a flexible substrate. The stress applied by the living body can be reduced to some extent. In addition, when the light source is arranged, it can be mounted relatively easily on a flexible substrate. In order to be implanted and driven in a living body, waterproofing is essential. In order to perform observation in a state of being substantially in contact with an observation object without using a lens, it is necessary to cover the whole with a transparent and thin film. One resin film that realizes this is a parylene-C film. This material is also highly biocompatible and transparent to visible light. Moreover, it can form into a film so that the whole may be wrapped by vapor deposition. The device shown in Fig. 2a is coated with a parylene-C film of about 1.5 µm and can be used without problems for imaging experiments in a state where it is implanted in a living body. Figure 2b is an example of imaging by a living implant image sensor arranged on the rat brain surface. Although the lens is not mounted, the blood vessel structure is clearly observed. A feature of the CMOS image sensor is that it is relatively easy to integrate various circuits. It is expected that more multifaceted information can be obtained not only by acquiring images but also by integrating other measurement functions. However, in a biological implant device, an increase in size leads to an increase in invasiveness to a manipulative body, and therefore, it is necessary to be a highly important circuit arranged near the observation region. Circuits that satisfy such conditions include electrodes and amplifiers for potential measurement. Furthermore, weak signals can be measured by integrating the amplifier on a chip near the measurement electrode. It is expected to lead to multifaceted analysis by enabling measurement of not only the light but also the potential associated with cell activity. The implantable image sensor introduced in this paper is a system connected to an external device by wire. Although the number of wirings is small and does not give a large burden, the observation target cannot be completely free. In the future, the introduction of wireless communication technologies such as Bluetooth will also require the development of wireless systems. In recent years, application to small animals, such as mice, has also come into view due to the miniaturization of wireless modules. Although further improvement is desired in communication speed, battery capacity, etc., it is considered that it can be sufficiently realized under limited conditions.

Implantable CMOS Fluorescent Imaging Devices

133

3 Ultra-Small Fluorescent Excitation Light Source A small light-emitting diode (LED) is used as a light source to be combined with an implantable CMOS image sensor. The light emitting diode emits light in a specific wavelength band and has high energy efficiency. In addition, since it is manufactured by a semiconductor manufacturing process, it is relatively easy to downsize, and a small one with a side of about 0.3 mm is also commercially available. By utilizing this, we have developed an ultra-compact fluorescent excitation light source that has high affinity with biological implantable devices.

3.1 Ultra-Thinning by Laser Lift-Off Process When inserting into the brain and observing the depth, it is desirable that the device dimensions be as small as possible to reduce invasiveness. In order to reduce the thickness of the light source portion, a laser lift off (LLO) method [13] is used. When observing green fluorescence, a blue LED is used as the light source. Most blue LEDs have a gallium nitride layer as a light emitting layer on a sapphire substrate, but sapphire has a wide light transmission spectrum and transmits ultraviolet light. On the other hand, gallium nitride which is a light emitting layer transmits visible light but does not transmit ultraviolet light. Utilizing this characteristic, when high-intensity ultraviolet laser light is irradiated from the surface on the sapphire side as shown in Fig. 3a, it is absorbed by gallium nitride near the interface and heated locally. As a result, the gallium nitride near the interface is decomposed and can be peeled off from the sapphire substrate. Figure 3b shows an example of a peeled blue LED. The thickness is about 8 µm, that is very thin as compared to the thickness of about 90 µm including the substrate before peeling. In addition, it can be confirmed that light is emitted without any problem even after peeling [14]. Fig. 3 a Stripping of LED light-emitting layer by laser lift-off method. b Micrograph of ultra-thin LED. The width and height are approximately 0.3 mm

(a)

UV laser

(b)

134

K. Sasagawa et al.

3.2 Wavelength Limitation by Integration of Excitation Filters An LED emits light in a specific color, but when considered as a light source for fluorescence observation, its emission spectrum is not necessarily narrow. Figure 4 shows an example of the emission spectrum of a blue LED. When the filter is not mounted, although there is a light emission peak in the blue wavelength band, the green wavelength band also has an intensity of several percent with respect to the peak wavelength. In fact, when the blue component is cut by a filter and observed, the blue LED looks green. This green emission intensity becomes a large background light component when observing fluorescence of green fluorescent protein (GFP) or the like. In this study, we tried to improve it by installing an excitation filter that can be mounted on a thin LED. When GFP is an observation object, the excitation filter is required to have high wavelength selectivity. This is because the interval between the absorption spectrum and the emission spectrum of GFP is narrow. Changing the characteristics from the reflection band to the transmission band with a slight difference in wavelength is difficult to achieve with an absorption filter, and an interference filter must be used. By applying the LLO method used for LED peeling, we developed a technology for mounting an interference filter by transfer. An interference filter is formed on quartz glass exhibiting high transparency to ultraviolet light, and the filter is peeled off from the interface by ultraviolet irradiation from the back surface. The solid line in Fig. 4 shows the emission spectrum of the LED light source equipped with the interference filter. A short pass filter with a cutoff wavelength of 475 nm was used as the filter. From this result, it can be seen that the intensity of the green wavelength band is greatly reduced. The overlapping portion with the transmission spectrum (dotted line) of the excitation light removal filter shown as a reference is greatly reduced, suggesting that the filter can efficiently remove light from the excitation light source. GFP emission band 1

Emission/Transmission (Normalized)

Fig. 4 Comparison of excitation light source spectrum and excitation light removal filter transmission spectrum with and without excitation filter

0.8 0.6 0.4 0.2 0 400

450

500

Wavelength (nm)

550

600

Implantable CMOS Fluorescent Imaging Devices (a) Fluorescent microscope

(b) Lens-free device w/o excitation filter

135 (c) Lens-free device with excitation filter

Fig. 5 Observation results of brain slices expressing green fluorescence. a Fluorescence image by fluorescence microscope. Imaging results using b light source without excitation filter and c small light source with excitation filter

The result of Fig. 4 is obtained by observing light emitted in a direction perpendicular to the LED. The transmission spectrum of the interference filter has an angle dependency, and the inclination increases and shifts to the short wavelength side. As described above, the emission spectrum of the LED has a broadening, but as it tilts, the wavelength component that can be transmitted decreases. As a result, the radiation angle is limited. This effect can be applied to the use of limiting the range of light irradiation by providing directivity, although the energy use efficiency is low. Figure 5 shows an example of fluorescence observation in which comparison is made between a prototype LED light source with a filter and an LED light source without a filter. The observation object is a mouse brain slice (GAD67-GFP) with a thickness of 100 µm expressing GFP. Figure 5a shows the result of observation with a fluorescence microscope. Figure 5b shows the result of transmission contact fluorescence observation using an LED light source without a filter and an image sensor with a filter. Compared with the fluorescence microscope, the contrast is completely different. This is because the green component of the transmitted LED is stronger than the GFP fluorescence and is in bright field observation. Figure 5c shows the result of a similar observation using a filter-mounted LED light source. The image shows almost the same pattern as in Fig. 5a, and fluorescence from GFP is detected. By peeling off and mounting the interference filter in this way, a thin LED light source having a wavelength characteristic suitable as an excitation light source can be produced.

3.3 Image Sensor with Small Light Source Since the biological implant image sensor is uniquely designed, various functions can be added. Taking advantage of this feature, we prototyped an image sensor that

136 Fig. 6 Ultra-thin LED-equipped biological implant image sensor. Copyright (2016) The Japan Society of Applied Physics. Quoted from Ref. [15]

K. Sasagawa et al.

0.16 mm

Pixel array (42 x 160 pixels)

0.5 mm

can be mounted with a thin LED [15]. Figure 6 shows a photograph of the prototype image sensor chip. LED mounting locations are arranged above and below the pixel array, which is the imaging unit. Further, the wiring for supplying power to the LED uses a metal wiring layer inside the chip. Furthermore, the metal wiring layer is also used as a light shielding layer around the LED. This method eliminates the need for a substrate by mounting LEDs directly on the chip. Therefore, it is easy to reduce the thickness as compared with the case where the LEDs are separately mounted on the substrate. The elements of the integrated circuit are formed to a depth of several µm near the surface of the Si substrate, and it is possible to further reduce the thickness. By using this technology, it is expected that the invasiveness to living tissue can be further reduced.

4 Excitation Light Removal Filter for Lensless System With the development of filter technology, it is possible to easily observe individual cells that emit fluorescence using a general fluorescence observation microscope. However, when fluorescence observation is performed using a device that does not use a lens, a problem different from that of a general microscope using a lens occurs.

4.1 Interference Filter The excitation light removal filter most often used in the current fluorescence microscope is an interference filter in which thin film transparent materials having different refractive indexes are laminated. By precisely adjusting the thickness of each layer and controlling the phase difference of light reflected at each interface, only a specific wavelength band can be transmitted or reflected. It is also possible to achieve very high reflectance (low transmittance) and high wavelength selectivity (difference in transmittance due to wavelength difference). With the maturation of manufacturing technology, it is widely used because of its design flexibility and high performance. One of the important characteristics of the interference filter is that the transmission characteristic has an incident angle dependency. The effective distance when light passes through the film constituting the interference filter varies depending on

Implantable CMOS Fluorescent Imaging Devices

137

Fig. 7 Change of transmission spectrum of a interference filter and b absorption filter with incident angle

the incident angle. As a result, the characteristics change because the phase relationship of the light reflected at each interface changes. As shown in Fig. 7a, the transmission spectrum has an effect of gradually shifting to the short wavelength side as the incident angle is inclined. This effect is not a big problem in a microscope optical system using a lens. When excitation light is irradiated onto the observation target, the light is scattered and spreads in various directions, but the light from the observation target is converted into a nearly parallel light beam by the objective lens and then enters the excitation light removal filter. To do. Although the incident angle to the filter differs depending on the position of the observation object, the difference is slight and it can be considered that the incident light is almost perpendicular. On the other hand, in an optical system that does not use a lens, light scattered by an observation target enters the filter at various angles. In normal fluorescence observation by one-photon excitation, the excitation light has a shorter wavelength than the fluorescence, and the excitation light removal filter is designed to reflect the excitation light and transmit the fluorescence having a longer wavelength. As described above, the transmission spectrum shifts to the short wavelength side with the inclination with respect to the filter. Therefore, when the inclination becomes a certain degree or more, the transmission wavelength band includes the excitation light wavelength band, and the transmission spectrum is transmitted. As a result, in the lensless system, even if an interference filter is used, the excitation light removal performance is almost determined by the scattering efficiency of the observation target. Scattering efficiency is determined by the difference in refractive index between the object to be observed and the surrounding medium. Scattering from cells in the liquid is not so large, but the fluorescence efficiency of fluorescent proteins is about 0.1% or less, which is problematic. There are many cases. Further, the reflectance and wavelength selectivity of the interference filter can be improved by increasing the number of layers of the multilayer film constituting the filter, but it is impossible to avoid a transmission spectrum shift due to the incident angle.

138

K. Sasagawa et al.

4.2 Absorption Filter In the absorption filter, a transmission characteristic is controlled by adding a pigment or the like to glass or resin and absorbing light having a specific wavelength. The transmission spectrum of the absorption filter is largely determined by the dye used. Therefore, in general, the wavelength selectivity is not high compared to an interference filter composed of a sufficiently large number of layers. One of the important features of the absorption filter is that it is different from the interference filter and has no angular dependence of the transmission spectrum. Figure 7b shows the transmission spectrum of the yellow absorption filter. In this way, even if the incident angle changes, the transmission characteristics do not substantially change. In this respect, it can be said that it is also suitable for a lensless fluorescent optical system. The transmittance of the absorption filter basically decreases exponentially with the thickness. Also, the absorption coefficient increases with the dye concentration. However, if the concentration is made as high as possible and a filter having the required thickness is prepared based on the absorption coefficient, it will not be possible to observe weak fluorescence. This is because, in the absorption filter, the filter itself that has absorbed the excitation light emits fluorescence. Therefore, there is a limit point in the effective excitation light removal performance that can be achieved by increasing the filter thickness. The fluorescence intensity varies depending on the dye, but in many cases, it is not sufficiently low as compared with fluorescent proteins and the like. This problem can be reduced in a fluorescence observation system using a lens by increasing the distance between the observation target and the filter. Since the fluorescence basically has no directivity and is radiated in various directions, the fluorescence of the filter at a position off the focal point of the lens is not condensed on the imaging surface. However, when the observation target is close to the imaging surface as in the proposed lensless fluorescence observation system, it is almost impossible to separate them.

4.3 Hybrid Filter As described in the previous section, the lensless fluorescence observation system has a problem that the interference filter and the absorption filter are different from the fluorescence observation system using a lens. The interference filter transmits scattered light components having a large inclination, but these respective drawbacks are complementary. Since the interference filter reflects the energy of incident light, the filter itself does not emit fluorescence. A part of the scattered light is transmitted but is small compared to the energy of the entire incident light. If the incident light intensity is low, the fluorescence of the absorption filter also decreases accordingly.

Implantable CMOS Fluorescent Imaging Devices

139

(b)

(a) Interference filter (>510 nm pass) Fiber optic plate Absorption filter Image sensor

Fig. 8 a Configuration of image sensor with hybrid filter. b Excitation light transmittance measurement result of hybrid filter and absorption filter (Modified from Ref. [16])

Based on this idea, a hybrid filter composed of an interference filter and an absorption filter was proposed and fabricated [16]. Usually, an interference filter is formed on a flat glass substrate because it is necessary to stack a high-quality film with a controlled refractive index and film thickness. However, in the lensless system, the spatial resolution is reduced due to the spread of light rays that pass through the substrate. Therefore, a fiber optic plate (FOP), which is a plate composed of a bundle of optical fibers, was used as the substrate (Fig. 8a). In FOP, since light incident on each point on the surface propagates through the core of the optical fiber and is emitted from the opposite surface, the spatial resolution is determined by the core pitch. The FOP used this time has a core pitch of 3 µm, which is finer than the pixel pitch of the image sensor used. The absorption filter was produced by using a yellow dye that is relatively less fluorescent and highly soluble, mixed with a resin solution as a base material, and solidified. In order to mount a flat film on a plate having a small area, a method of transferring the produced film using a spin coating method was used. Figure 8b shows the result of measuring the excitation light transmittance of each filter. The horizontal axis shows the thickness of the absorption filter. In a sufficiently thin region, the transmittance decreases exponentially with the thickness, but beyond a certain point, the transmittance is increased even if the filter is thickened. Has not decreased. Correctly, the excitation light continues to attenuate, but the effective transmittance is at the bottom because the filter itself emits fluorescence. On the other hand, the hybrid filter has an excitation light transmittance of about 10–4 with no absorption filter mounted. From this point, as the thickness of the absorption filter increases, a curve similar to that of the absorption filter alone is drawn. Yes. As a result, the excitation light transmittance has reached about 10–8 , and performance close to that of an interference filter for a fluorescence microscope is realized. Figure 9 is an example of fluorescent bead observation in a device in which half of one sensor is a hybrid filter and the other half is only an absorption filter in order to compare the hybrid filter and the absorption filter. By using the hybrid filter, the background component is suppressed and the fluorescent beads can be clearly

140 Fig. 9 Imaging results of fluorescent beads with a comparative image sensor equipped with an absorption filter and a hybrid filter (Modified from Ref. [16])

K. Sasagawa et al. Absorption filter

FOP

Hybrid filter

300 µm observed. Further, the fluorescence from the fluorescent beads is also brightened, because most of the excitation light is reflected by the interference filter layer on the surface, and the intensity of the light applied to the fluorescent beads is increased. A hybrid filter using FOP as a substrate functions as a single filter and is easy to handle, so it can be combined with a commercially available image sensor [17, 18]. This makes it possible to produce large-format lensless fluorescence imaging. In order to increase the field of view, it is necessary to increase not only the image sensor but also the lens in the lens optical system. In the lensless system, dimensions other than the sensor do not change significantly. For this reason, even an imaging area of several tens of mm2 can be realized with a palm size.

4.4 Installation of Hybrid Filter on Implantable Image Sensor The hybrid filter introduced in the previous section avoids a decrease in spatial resolution due to the thickness of the device by forming an interference filter on the FOP. However, since the volume of the device becomes large, such a method is not suitable for a living body implant image sensor, particularly a penetration type sensor. However, an absorption filter composed of a resin added with a dye at a high concentration has low mechanical strength and heat resistance, and it is difficult to directly form an interference filter. To solve this problem, we applied the LLO method, which was also used for mounting the filter on the light source and transferred the interference filter [19–21]. By adopting the transfer method, it becomes possible to mount an interference filter produced by a general mature film forming method on a resin filter. Figure 10a shows an example of a hybrid filter living body implant image sensor manufactured by the proposed method. In this sensor, the interference filter has a band-pass configuration in which a long-pass filter and a short-pass filter are combined. Thereby, the autofluorescence of the red band emitted from the living

Implantable CMOS Fluorescent Imaging Devices

(a)

141 GFP emission band

(b) 1 Normalized sensitivity

Image sensor with filter

5 mm

0.8 0.6 0.4 0.2 0

400

500 600 Wavelength (nm)

700

Fig. 10 a Biological implant image sensor with hybrid filter. b Sensitivity spectrum

tissue is reduced, and an improvement in the green fluorescence detection performance is expected. Figure 10b shows spectral sensitivity characteristics of a sensor equipped with the proposed filter. In this measurement, excitation light was irradiated perpendicularly to the sensor. The region where the wavelength is 520 nm or less is almost impermeable. The boundary on the short wavelength side is designed to be almost the same between the interference filter and the absorption filter. On the other hand, the boundary on the long wavelength side (wavelength 550 nm) is determined by a short path type interference filter. The transmission spectrum of the absorption filter does not depend on the incident angle, but in the interference filter, the incident angle is inclined and shifted to the short wavelength side. As a result, the transmission wavelength band becomes narrow according to the inclination, and the incident angle at which the pixel has sensitivity is limited.

5 Toward Improvement of Spatial Resolution of Implantable Image Sensors The implantable image sensors do not have a lens. This feature enables low invasiveness and light weight. However, it is difficult to achieve sufficiently high spatial resolution for single cell observation due to the “blur” with distance. The hybrid filter structure described above requires a thickness of about 10 µm or more. Further, when the puncture is inserted into the brain tissue, the area very close to the punctured portion is damaged and not suitable for observation. Thus, it is necessary to observe cells several tens of µm ahead. The “blur” due to this distance cannot be ignored when observing at the cellular level. One of the methods for increasing the effective spatial resolution in a lensless manner is to perform light field imaging. From position and angle information, an original image can be restored by image processing. By using a metal pattern, an

142

K. Sasagawa et al.

image sensor for such a purpose can be fabricated [22–24]. For example, a group at Cormell University has reported angle-sensitive pixels using a two-layer grating [22]. The purpose of these sensors is to realize the functions of a general camera with a single image sensor without using a lens. It is assumed that an image of object located very far from the sensor is acquired. In this purpose, a high angular resolution is required. On the other hand, the sensor inserted into the brain targets neural cells located several tens of µm away from the sensor. Therefore, the angular resolution does not need to be high. However, it is necessary to detect even incident light having a large inclination. As a method satisfying such a requirement, we have proposed an angle selection pixel (Fig. 11a) [24]. One aperture is shared by the four pixels. In this structure, the vertically incident light is blocked by the lower metal wiring layer, and only the obliquely incident light is detected by each pixel. One of the advantages of this structure is that high rejection of vertically incident light can be achieved. In addition, there is almost no polarization dependence since no grating structure is used. On the other hand, it is assumed that one pixel detects light incident in one direction. In order to acquire data having a high angular resolution, a pixel array including various types of pixels is required. However, when an observation target near the sensor pole is assumed as described above, a high angular resolution is not required. Figure 11b shows an example of a prototype chip. In this chip, a set of pixels is composed of five pixels, which is one normal pixel and four angle selection pixels sharing one aperture. The size of one image pixel is 7.5 µm, and the aperture pitch is 16.8 µm. Figure 12 shows the angle dependence of the sensitivity. The normal pixel has a peak for the normal incident light, whereas the angle selection pixel has a peak at 40° with half-width of 23°. The selectivity between the peak and the normal incidence component was approximately 25.

Fig. 11 a Schematic representation of angle-selective pixel structure. b Micrograph of proposed CMOS image sensor and pixel array configuration (Modified from Ref. [24])

Implantable CMOS Fluorescent Imaging Devices

143

Fig. 12 Pixel output of normal and angle-selective pixels as functions of incident angle (Modified from Ref. [24])

Using the proposed sensor, image recovering of the fluorescent beads was performed. A filter containing a yellow absorbing dye was coated on the pixel array as an emission filter. And, a gap was formed with a cover glass. 15-µm fluorescent beads (F8844, Thermo Fisher) were dispersed on the chip and irradiated with excitation light for observation. Using a pattern obtained from a single bead as a point spread function, deconvolution by the Lucy-Richardson method was performed from each angle image, and a restored image was obtained as a product of the deconvoluted images. Figure 13 shows the result. In a deconvolution process, it is usual that artifacts may be emphasized along with the number of repetitions. However, the results of using the proposed sensor show that the image of each bead can be restored and suggest that the combination of angle-resolved pixel array and image processing can improve the spatial resolution.

(a)

(b)

Fig. 13 Comparison of three fluorescent beads observation results by fluorescence microscope and combination of proposed image sensor and image processing a Reference image with fluorescent microscope. b Reconstructed image from five different images obtained by proposed image sensor (Modified from Ref. [24])

144

K. Sasagawa et al.

6 Summary It was shown that by using a specially designed image sensor, an ultra-compact imaging system that cannot be realized with a normal microscope optical system can be constructed. Compared with other technologies, the feature is particularly suitable for measurement under free action, deep brain observation, simultaneous observation of multiple points, and the like. In this paper, we showed that the use of ultra-compact light source and filter mounting technology that applied exfoliation and transfer methods can further reduce the invasion and improve the fluorescence observation performance of living implantable image sensors. Future developments include multi-functionalization and wireless integration that integrates measurement technologies other than image sensors in addition to higher performance. This is expected to contribute to the progress of brain function elucidation.

References 1. 2. 3. 4. 5. 6. 7.

8. 9. 10. 11. 12.

13. 14. 15. 16.

Blasdel GG, Salama G (1986) Nature 321:579–585 Nakai J, Ohkura M, Imoto K (2001) Nat Biotechnol 19:137–141 Dombeck DA, Graziano MS, Tank DW (2009) J Neurosci 29:13751–13760 Ferezou I, Bolea S, Petersen CCH (2006) Neuron 50:617–629. Sawinski J, Wallace DJ, Greenberg DS, Grossmann S, Denk W, Kerr JN (2009) Proc Natl Acad Sci USA 106:19557–19562 Ghosh KK, Burns LD, Cocker ED, Nimmerjahn A, Ziv Y, El Gamal A, Schnitzer MJ (2011) Nat Methods 8:871–878 UCLA miniscope project. https://miniscope.org Ohta J, Higuchi A, Tagawa A, Sasagawa K, Tokuda T, Hatanaka Y, Ishikawa Y Shiosaka S (2008) An implantable CMOS image sensor for monitoring deep brain activities of a freely moving mouse. In: 2008 IEEE biomedical circuits and systems conference. IEEE, pp 269–272 Ohta J, Tokuda T, Sasagawa K, Noda T (2009) Sensors 9, 9073–9093 Ohta J, Ohta Y, Takehara H, Noda T, Sasagawa K, Tokuda T, Haruta M, Kobayashi T, Akay YM, Akay M (2017) Proc IEEE 105:158–166 Takehara H, Katsuragi Y, Ohta Y, Motoyama M, Takehara H, Noda T, Sasagawa K, Tokuda T, Ohta J (2016) Appl Phys Express 9:047001 Kobayashi T, Motoyama M, Masuda H, Ohta Y, Haruta M, Noda T, Sasagawa K, Tokuda T, Tamura H, Ishikawa Y, Shiosaka S, Ohta J (2012) Biosens Bioelectron 38:321–330 Haruta M, Sunaga Y, Yamaguchi T, Takehara H, Noda T, Sasagawa K, Tokuda T, Ohta J (2015) Jpn J Appl Phys 54:04DL10. Sunaga Y, Yamaura H, Haruta M, Yamaguchi T, Motoyama M, Ohta Y, Takehara H, Noda T, Sasagawa K, Tokuda T, Yoshimura Y (2016) Implantable imaging device for brain functional imaging system using flavoprotein fluorescence. Jpn J Appl Phys 55(3S2):03DF02 Wong WS, Sands T, Cheung NW, Kneissl M, Bour DP, Mei P, Romano LT, Johnson NM (1999) Appl Phys Lett 75:1360–1362 Sasagawa K, Haruta M, Fujimoto K, Ohta Y, Noda T, Tokuda T, Ohta J (2017) The 13th IEEE BioCAS (BioCAS2017) Sasagawa K, Yamaguchi T, Haruta M, Sunaga Y, Ohta Y, Takehara H, Takehara H, Noda T, Tokuda T, Ohta J (2016) Ext. Abstr. Solid state devices and materials, H-4-05 Sasagawa K, Kimura A, Haruta M, Noda T, Tokuda T, Ohta J (2018) Biomed Opt Express 9:4329–4344

Implantable CMOS Fluorescent Imaging Devices

145

17. Sasagawa K, Ohta Y, Kawahara M, Haruta M, Noda T, Tokuda T, Ohta J (2019) AIP Adv 9:035108 18. Hee WS, Sasagawa K, Kameyama A, Kimura A, Haruta M, Tokuda T, Ohta J (2019) Lens-free dual-color fluorescent CMOS image sensor for Förster resonance energy transfer imaging. Sens Mater 31(8):2579–2594 19. Sasagawa K, Ohta Y, Haruta M, Noda T, Tokuda T, Ohta J (2018) IEEE BioCAS 2018. Cleveland, OH, USA 20. Rustami E, Sasagawa K, Ohta Y, Haruta M, Noda T, Tokuda T, Ohta J (2019) A thin composite emission filter and fiber coupled laser excitation for implantable fluorescence imager application. In: 2019 IEEE international symposium on circuits and systems (ISCAS) A2L-I-2, Sapporo, Japan 21. Rustami E, Sasagawa K, Sugie K, Ohta Y, Haruta M, Noda T, Tokuda T, Ohta J, Needletype image sensor with band-pass composite emission filter and parallel fiber-coupled laser excitation. IEEE Trans Circ Syst I (to be published) 22. Wang A, Molnar A (2012) A light-field image sensor in 180 nm CMOS. J Solid-State Circ 47(1):257–271. https://doi.org/10.1109/JSSC.2011.2164669 23. Adams JK, Boominathan V, Avants BW, Vercosa DG, Ye F, Baraniuk RG, Robinson JT, Veeraraghavan A (2017) Sci Adv 3:e1701548 24. Sugie K, Sasagawa K, Guinto MC, Haruta M, Tokuda T, Ohta J (2019) Electron Lett 55(13):729–731

Kiyotaka Sasagawa received his B.S. from Kyoto University, Kyoto, Japan, in 1999. He then moved to Nara Institute of Science and Technology (NAIST), Nara, Japan and finished M.E. and Ph.D. degrees in 2001 and 2004, respectively. From 2004 to 2008, he was a researcher with the National Institute of Information and Communications Technology (NICT), Koganei, Tokyo, Japan. His research projects were electromagnetic field measurement based on microwave photonics techniques and nonlinear micro-optic devices using whispering gallery modes. Afterwards, he had been an assistant professor since 2008 and has been working as an associate professor since 2019 at NAIST. His current research interests involve CMOS image sensors for invivo bioimaging, lensless fluorescence imaging devices, retinal prosthetic devices and integrated circuit design for such biomedical applications.

Neural Computation and Data Analysis

Data Analysis Method for Neuroimaging Data: Task-Related Component Analysis and Its Applications to fNIRS Data Hirokazu Tanaka, Takusige Katura, and Hiroki Sato

1 Introduction Data analysis methods in neuroimaging play an essential role in removing artifactual noises and extracting neural activations of interest because raw data contain various types of noises leading to spurious, false-positive activations. There is no one-sizefits-all method for analyzing neuroimaging data, and a number of analysis methods for various purposes have been and are being developed. The choice of data analysis method depends critically on the nature of a question of investigator and the design of an experiment, and an analysis method that is not appropriately selected could lead to an incorrect conclusion. Therefore, one should inspect whether an analysis method is suited for the particular experiment and hypothesis and should also be open-minded to new methods. This chapter provides an overview of neuroimaging data methods with a focus on functional near-infrared spectroscopy (fNIRS) and proposes our new analysis method. The philosopher Karl Popper once wrote that “no reproducible single occurrences are of no significance to science” [1]. Our method called task-relatedcomponent analysis (TRCA) is proposed based on the belief that a signal that appears repeatedly in a same task or condition should be regarded as task related, in much the H. Tanaka (B) Faculty of Information Technology, Department of Intelligent Systems, Tokyo City University, Setagaya, Tokyo, Japan e-mail: [email protected] T. Katura NeU Corporation, Creation Work Unit, Japan H. Sato Department of Bioscience and Engineering, College of Systems Engineering and Science, Shibaura Institute of Technology, Koto City, Tokyo, Japan e-mail: [email protected] © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_7

149

150

H. Tanaka et al.

same spirit of Karl Popper. This chapter provides a review of previous methods and their advantages and disadvantages, and introduces a formulation and applications of TRCA in comparison with previous methods [2, 3]. Although we here focus on the applications to fNIRS data, the concept of reproducibility is not limited only to fNIRS but applicable to biophysical data analysis, so we expect that TRCA attracts attention from a range of applications beyond fNIRS.

2 Overview of Functional Near-Infrared Spectroscopy (fNIRS) Functional NIRS for brain functions relies on the two biophysical principles about the interaction among light, biological tissues and hemoglobin. First, biological tissue is optically transparent to light with wavelengths ranging from visible to near-infrared region, which often is called the “biological optical window.” Therefore, the light within this window allows a non-invasive investigation of biological tissue from externally located probes. Second, there is a considerable difference between the absorption coefficient of oxygenated hemoglobin (oxy-Hb) and that of deoxygenated hemoglobin (deoxy-Hb), the measurement of light absorption in the biological optical window quantifies the oxygen consumption in the local tissue. By utilizing these principles, Millikan’s development of a spectroscopic method for estimating the oxygen saturation of blood in human tissue [4] was followed by the development of the NIRS technique for noninvasive measurement of tissue oxygenation in the brain [5, 6]. Of particular importance was the application of NIRS to the measurement of hemodynamic signals related to functional activation in the brain [7–10], often referred to as fNIRS. The fNIRS technique measures relative changes in the concentration of oxyand deoxy-Hb signal, taking advantage of the difference in absorption coefficients between oxy-Hb and deoxy-Hb depending on the light wavelength [6]. After fNIRS imaging technique with multiple measurement positions was introduced [11, 12], a number of studies using fNIRS were conducted in various fields as it offers several advantages to other modalities. The fact that fNIRS requires little constrains on body movements facilitates measurements of the brain activation of infants [13–17] and of multiple subjects communicating to each other [18, 19], which are more difficult or even impossible with functional MRI (fMRI). Moreover, with recent development of wearable instruments, fNIRS has become more user-friendly and will extend the possibility of brain measurement applications to novel fields that have not been explored yet. The unique feature of fNIRS is the simultaneous measurement of oxy-Hb and deoxy-Hb changes. There are other neuroimaging methods that measure brain functions through hemodynamic changes and metabolic consumptions, including fMRI as a prime example. The origin of the fNIRS deoxy-Hb signal is assumed to be the same as that of the blood oxygenation level-dependent (BOLD) signal in fMRI, which is sensitive to changes in the concentration of deoxy-Hb in local blood vessels [20, 21]. fNIRS-Hb signal was shown to reflect the fMRI-BOLD signal in the gray

Data Analysis Method for Neuroimaging Data: Task-Related Component …

151

matter of activated area in a working memory task [22]. Most of fNIRS studies, however, use oxy-Hb signals as the main index of brain activation and do not exploit the advantage of simultaneous oxy- and deoxy-Hb measurement. As seen below, an extension of TRCA adopts this advantage so as to extract task-related oxygenation and blood volume changes as negative and positive covariations between oxy- and deoxy-Hb, respectively.

3 Review of Analysis Methods Although fNIRS has been around for more than twenty years since its first application to the human brain, there is no unique, standard method for its data analysis, besides simple train averaging. It makes a sharp contrast with the case of fMRI, where general linear modeling and statistical parametric mapping are a gold standard. It is probably because the spatially uniform sensitivity of fMRI allows a statistical test with a single threshold in constructing a spatial map (t-map or F-map), in contrast to spatially discrete and non-uniform sampling of fNIRS. Analysis methods for neuroimaging data analysis fall broadly into two distinct (but not mutually exclusive) categories: hypothesis-driven and data-driven approaches. There are pros and cons for both approaches as reviewed below with a focus of fNIRS studies, as even among experts the opinion is divided between the hypothesis- and the datadriven approaches [23, 24]. More sophisticated methods such as multivariate pattern analyses and connectivity analyses are not reviewed here.

3.1 Hypothesis-Driven Approach Analysis methods of hypothesis-driven approach make certain assumptions of how observed data are generated from externally provided stimuli (i.e., visual or auditory stimuli) or from unobservable neural activities. These methods hence require hypotheses or generative models of observed signals, so these are most powerful when plausible hypotheses are already given. The most notable example of this approach is a general linear model (GLM) and statistical parametric mapping (SPM), widely used both in fMRI [25] and fNIRS studies [14]. A GLM establishes a linear model between experimental conditions and observed responses, and SPM evaluates statistically whether a channel in fNIRS or a voxel in fMRI is significantly correlated with GLM parameters. Despite its wide use in neuroimaging data analysis, the hypothesis-driven approach by definition depends on how a hypothesis or a GLM is constructed, so inappropriate modeling of generative model in an experiment might well lead to an incorrect conclusion. Also, a statistical test based on SPM for fNIRS with a single thresholding value is not fully justified because fNIRS channel sensitivity is location dependent due to variable thickness of the skull that leads to variable optical path lengths. In general, the hypothesis-driven approach allows us to ask only whether a particular hypothesis

152

H. Tanaka et al.

is supported by data or not, therefore an unsupposed finding cannot be expected. In sum, the hypothesis-driven approach is recommended for fMRI when a well-defined hypothesis is already at hand.

3.2 Data-Driven Approach Analysis methods of data-driven approach assume general statistical assumptions about observed data and extract statistical regularities without requiring a generative model underlying the observed data. Therefore, this approach is effective when there is no specific working hypothesis of how the observed data are generated from unobservable, neural activities, most suitable for an exploratory analysis. The most notable examples of this approach include principal component analysis (PCA) and independent component analysis (ICA). PCA is a method with a covariance matrix (secondorder statistics) and extracts principal components that predominantly contributes the variance of observed data. PCA is formulated and solved as a matrix eigenvalue problem. ICA is a method based on probability density functions and extracts independent components that are mutually independent to each other. ICA is formulated as maximization of the Shannon entropy or minimization of the Kullback–Leibler divergence and is solved by an iterative learning algorithm. It is particularly suitable for a resting-state experiment in which subjects perform no task, where it is not apparent how a GLM is constructed for. However, the hypothesis-driven approach suffers from an interpretation issue, i.e., how the researcher determines what components have neural origins and what other components are artifactual. This interpretation process leaves a room of arbitrary selection that may depend on the expertise of researchers. In sum, this data-driven approach is recommended for an exploratory analysis when no specific hypothesis about data is available, but a care must be taken in interpreting the results.

4 Task-Related Component Analysis: Formulation Our method, task-related component analysis (TRCA), is proposed based on a concept that reproducibility of experimental results lies at the heart of scientific disciplines. Usually, an experiment session holds multiple blocks or events that have a same experimental condition, and the signals from the same condition are averaged to improve a task-related signal and to eliminate artifactual components. We therefore define a task-related component as a component that robustly and consistently appears in every block or every event of same condition. TRCA constructs a task-related component as a weighted sum of observed data and determines the weight coefficients so as to maximize its block-by-block reproducibility. Therefore, TRCA requires only the timing of blocks of a same experimental condition, unlike hypothesis-driven methods that require detailed assumptions about generative models. Moreover, TRCA provides a systematic and objective method of statistical

Data Analysis Method for Neuroimaging Data: Task-Related Component …

153

Fig. 1 Schematics of TRCA. a Observed multiple time series are decomposed into b task-related components using the generalized eigenvalue algorithm c Statistical significance of the eigenvalues (red crosses on the horizontal axis) is evaluated using the resampling procedure. The dotted vertical line indicates a 99% confidence level. d Two eigenvalues outside the confidence level are selected as being significantly task-related in this example. The insets on right depict block averages of the two components, respectively. These figures were created with fNIRS finger tapping data. Adopted from [2]

test for the reproducibility of task-related components, so the subjective, researcherdependent process of choice and interpretation of components that is necessary in most data-driven methods is no longer needed. A conceptually similar (but technically different) method based on the reproducibility principle was proposed for MEG/EEG data analysis [26]. We here provide a basic formulation of TRCA and several extensions, as schematically illustrated in Fig. 1.

4.1 Signal Reconstruction from Weighted Linear Summation TRCA assumes a linear weighted sum of input time series {x i (t)}, where x i (t) stands for time series observed of oxy-Hb change from i-th channel. A task-related component is constructed as

154

H. Tanaka et al.

y(t) =

N 

wi xi (t) = wT x(t)

(1)

i=1

where {wi } are coefficients or weights that are optimized so as to maximize the block-by-block reproducibility of y(t). This assumption of linear weighted sum is widely used for other typical analysis methods including PCA, ICA, common spatial filters, linear discriminant analysis and support vector machine. The weights are to be determined according to the reproducibility requirement described below.

4.2 Task-Related Component Analysis as Generalized Eigenvalue Problem TRCA defines task-related components by maximizing the reproducibility among task blocks while maintaining the variance of task-related components constant. One measure of the reproducibility is a covariance between a pair of blocks, i.e., k-th and l-th blocks as   Cˆ kl = Cov y (k) (t), y (l) (t) N    = wi w j Cov xi(k) (t), xk(l) (t)

(2)

i, j=1

The reproducibility of a task-related component should be not only for a particular block pair but also for all possible pairs, so the proposed objective function to be maximized is K  k,l=1 k=l

=

K 

Cˆ kl =

  Cov y (k) (t), y (l) (t)

k,l=1 k=l

N K  



 wi w j Cov xi(k) (t), x (l) (t) = wT Sw j

(3)

k,l=1 i, j=1 k=l

where the matrix S is defined as (S)i j ≡

K 

  Cov xi(k) (t), x (l) j (t)

(4)

k,l=1 k=l

The objective function is quadratic in terms of the weights, so a proper normalization constraint has to be imposed. We require a unit variance of task-related

Data Analysis Method for Neuroimaging Data: Task-Related Component …

155

component as Var(y(t)) =

N 

  wi w j Cov xi (t), x j (t)

i, j=1

= wT Qw = 1

(5)

Since the signal y(t) in Eq. (1) is normalized to zero mean and unit variance, it is unitless and “arbitrary unit (a. u.)” is used for labels of y-axes. TRCA is formulated as a maximization of the objective function (3) under the normalization constraint (5). It is well known that the constrained maximization problem is equivalent to maximizing a quotient as ˆ = arg max w w

wT Sw wT Qw

(6)

This quotient is called the Rayleigh–Ritz quotient, and this maximization problem is formulated and solved as a generalized eigenvalue problem. The solution is   obtained as N eigenvectors W = w1 w2 · · · w N of the matrix Q−1 S, with corresponding eigenvalues λ1 ≥ λ2 ≥ · · · ≥ λ N in a decreasing order, i.e., w1 and wN having the largest and the smallest eigenvalues, respectively. The eigenvector with the largest eigenvalue is often called the primal eigenvector. Note that the value of the eigenvalue suggests how reproducible in a block-by-block basis the corresponding signal is, and can hence be used for statistical test of its reproducibility. A Matlab function for TRCA is found in Appendix A.

4.3 Statistical Test of Signal Reproducibility Every eigenvector w comes with a corresponding eigenvalue λ. The eigenvalue has a natural interpretation: ˆ = ˆ T Sw w

  ˆ T Qw ˆ ˆ ˆ T Sw λw w ˆ ˆ = = λ ∵ S w = λQ w ˆ T Qw ˆ T Qw ˆ ˆ w w

(7)

Thereby, the value of eigenvalue equals to the block-by-block covariance of corresponding eigenvector. We therefore propose a statistical test of task-related component based on the associated eigenvalue. If the eigenvalue is above a threshold value that is determined by a null hypothesis, the corresponding task-related component has a statistical significance; otherwise, the component is not statistically significant and is excluded for a further analysis. Our null hypothesis is that there are no reproducible components in the experimentally assigned task blocks. The null hypothesis predicts an eigenvalue distribution (null distribution) by uniformly randomizing the timing of task blocks instead of the timing that is actually used in an experiment.

156

H. Tanaka et al.

Practically, 200 times of the resampling procedure are iterated to obtain the null distribution. The eigenvalues obtained with the timings of experimental blocks are then statistically tested by comparing with the null distribution. We set, for convenience, a 99% confidence interval reflecting a significance level of 0.01. A Matlab script for the resampling statistical test is found in Appendix B.

4.4 Mapping of a Task-Related Component Once a statistically significant task-related component is identified, the next task is to compute how each NIRS channel contains that task-related component. We take a Pearson correlation coefficient between time series of i-th channel and task-related component as a measure for this purpose, simply E[xi (t) · y(t)] where E[·] means time average. Then these correlation coefficients were mapped on a cortical surface for visualization. Because the spatial maps were defined by correlation coefficients, their values range from −1 to 1 and were unitless. When there are multiple subjects, the individual spatial maps were averaged for a mean spatial map.

5 Extensions of TRCA 5.1 Task-Discriminative Components It is often desirable to make a comparison or a contrast between neural signals from multiple task conditions, and we here show that an extension of TRCA can construct a component that are reproducible in each condition and are maximally contrasting between conditions. If there are two conditions in an experiment (e.g., left or right finger tapping), it would be of interest to find a component that appears consistently within the same condition and differs maximally across different conditions. TRCA is extended so as to extract such a component that contrasts two conditions while keeping the reproducibility in each condition, referred to as discriminative TRCA. Let us consider two types of tasks,   say A and B, which haveK A and K B blocks with task periods t ∈ tkA , tkA + T (k = 1, · · · , K A ) and t ∈ tkB , tkB + T (k = 1, · · · , K B ), respectively. We first quantify the reproducibility of task-A blocks as a sum of covariances as   CklAA = Cov y (kA ) (t), y (lA ) (t)    = wi w j Cov xi(kA ) (t), x (lj A ) (t) , i, j

and the reproducibility of task-B blocks as a sum of covariances as

(8)

Data Analysis Method for Neuroimaging Data: Task-Related Component …

  CklBB = Cov y (kB ) (t), y (lB ) (t)    = wi w j Cov xi(kB ) (t), x (lj B ) (t)

157

(9)

i, j

To make a contrast between the conditions, we compute a sum of covariances between A and B,   Cˆ klAB = Cov y (kA ) (t), y (lB ) (t)    = wi w j Cov xi(kA ) (t), x (lj B ) (t)

(10)

i, j

and a sum of covariances between B and A,   Cˆ klBA = Cov y (kB ) (t), y (lA ) (t)    = wi w j Cov xi(kB ) (t), x (lj A ) (t) .

(11)

i, j

Finally, the proposed objective function is a difference between the withincondition covariances and the across-condition covariances, K A ,K B

A ,K B    K  AA BB ˆ ˆ Ckl + Ckl − Cˆ klAB + Cˆ klBA = wT Sw,

k,l=1 k=l

(12)

k,l=1

where the components of the matrix S is defined as (S)i j ≡

   

 Cov xi(kA ) (t), x (lj A ) (t) + Cov xi(kB ) (t), x (lj B ) (t) k=l

   

 − Cov xi(kA ) (t), x (lj B ) (t) + Cov xi(kB ) (t), x (lj A ) (t) .

(13)

k=l

The first sum in the left-hand side in Eq. (12) maximizes the reproducibility within one condition, and the second sum make a contrast between two conditions. With the matrix S defined above, this extension of TRCA can be solved as a generalized eigenvalue problem as described in Sect. 2.

5.2 Task-Related Oxygenation and Blood Volume Changes As reviewed in Sect. 2, fNIRS has a unique feature of simultaneous recording of oxyand deoxy-Hb, unlike fMRI that is sensitive only to paramagnetic deoxy-Hb. Therefore, by taking a positive or negative correlation between oxy- and deoxy-Hb, fNIRS

158

H. Tanaka et al.

can detect informative hemodynamics that is not measureable with fMRI. A negative correlation between oxy- and deoxy-Hb suggests oxygenation change whereas a positive correlation suggests cerebral blood volume (CBV) change [27–30]. The basic formulation of TRCA uses only oxy-Hb changes, and it is straightforward to take covariation of oxy- and deoxy-Hb changes into consideration within the TRCA framework. There are two versions of TRCA extension in this direction: one that maximizes positive covariation of oxy- and deoxy-Hb (hereafter referred to as TRCA+ ) and the formulaother that maximizes negative covariation (TRCA− ). Following  the original tion of TRCA, one weighted linear sum of oxy-Hb changes ( xoxy,i , i = 1, . . . , N , N: the number of fNIRS channels), yoxy (t) =

N 

wi xoxy,i (t) = wT Xoxy (t),

(14)

i=1

and another of deoxy-Hb changes ({xdeoxy,i }, i = 1, . . . , N ), ydeoxy (t) =

N 

wi xdeoxy,i (t) = wT Xdeoxy (t).

(15)

i=1

 xoxy,i and xdeoxy,i are normalized to zero mean and unit variance before taking the weighted sums. Note that the same weights {wi } are used in constructing yoxy and ydeoxy . We now formulate the reproducibility requirements, 

  (k) (l) CklOO = cov yoxy (t), yoxy (t) and   (k) (l) CklDD = cov ydeoxy (t), ydeoxy (t) ,

(16)

and the covariation requirements,   (k) (k) (t), ydeoxy (t) and CkOD = cov yoxy   (k) (k) CkDO = cov ydeoxy (t), yoxy (t) .

(17)

The objective function for TRCA+ is K K   OO    OD  Ckl + CklDD + Ck + CkDO = w T S+ w. k,l=1 k=l

k=1

(18)

Data Analysis Method for Neuroimaging Data: Task-Related Component …

159

and for TRCA− is K K   OO    OD  Ckl + CklDD − Ck + CkDO = w T S− w. k,l=1 k=l

(19)

k=1

In addition the constraint of normalized variance is imposed as     Var yoxy (t) + Var ydeoxy (t) = wT Qw = 1,

(20)

where the N × N matrix Q is defined as an average of covariance matrices of oxyand dexoy-Hb. With the matrices S+ (for TRCA+ ) in Eq. (18) or S− (for TRCA− ) in Eq. (19) and Q in Eq. (20), discriminative TRCA is solved in the same way described in Sect. 4. Hereafter TRCA+ and TRCA− are collectively referred to as TRCA± .

5.3 Online TRCA TRCA was originally formulated as a batch or offline algorithm but can be extended to an online algorithm, where the weight vector w(K) with K-block data is incrementally computed from the weight vector w(K−1) with (K−)-block data. The online algorithm presented here were derived from [31], which was recently extended to an application to EEG data [32]. A primal eigenvector w1 (i.e., the eigenvector with the largest generalized eigenvalue) should satisfy the generalized eigenvalue problem as Sw1 = λQw1 =

w1T Sw1 Qw1 , w1T Qw1

(21)

or, w1 =

w1T Qw1 −1 Q Sw1 . w1T Sw1

(22)

This form suggests an online, recursive update solution for the generalized eigenvalue problem for the primal eigenvector w1 (K) as w1 (K ) =

where

w1T (K − 1)Q(K )w1 (K − 1) −1 Q (K )S(K )w1 (K − 1) w1T (K − 1)S(K )w1 (K − 1)

(23)

160

H. Tanaka et al.

(S(K ))i j ≡ (S(K ))i j +

K −1 

  ) Cov xi(k) (t), x (K j (t)

(24)

k=1

and the covariance matrix Q(K) and its inverse Q−1 (K) can be recursively computed using the matrix inversion lemma. The convergence of this algorithm is guaranteed [31]. For non-primal, lower-order eigenvectors, 

T Sn−1 (K )wn−1 (K − 1)wn−1 (K − 1) Sn−1 (K ) Sn (K ) = I − T wn−1 (K − 1)Sn−1 (K )wn−1 (K − 1)

(25)

Qn (K ) = Qn−1 (K )

(26)

Here S1 (K ) = S(K ) and Q1 (K ) = Q(K ). Note that the matrix Sn is obtained from the matrix Sn−1 by projecting out wn−1 , i.e., Sn (K )wn−1 (K − 1) = 0. With these matrices Sn and Qn , the n-th order eigenvector wn (K) is computed recursively from the eigenvector wn (K−1) as wn (K ) =

wnT (K − 1)Qn (K )wn (K − 1) −1 Q (K )Sn (K )wn (K − 1). wnT (K − 1)Sn (K )wn (K − 1) n

(27)

This recursive formulation of TRCA is useful for online applications such as brain-computer interfaces.

5.4 Delayed TRCA Until now, TRCA is formulated by assuming an instantaneous weighted mixture of data as in Eq. (1). Because the rise time of hemodynamic responses may vary in channel to channel, the assumption of instantaneous mixing can be relaxed by introducing the channel-dependent delay times as in y(t) =

N 

wi xi (t + di ).

(28)

i=1

Here d i is a delay time constants for i-th channel. We provide an iterative algorithm that determines the weights {wi } and the delays {d i } in an alternative way, inspired by the well-known matching pursuit algorithm. 1. Initialize the delays {d i } to zero. 2. Compute {wi } with the largest eigenvalue following the standard procedure of the original TRCA.

Data Analysis Method for Neuroimaging Data: Task-Related Component …

161

3. Optimize {d i } with the given values of {wi } under physiologically plausible values (−1.0 < d i < +1.0 for our case) according to an adaptive correlating filter algorithm [33]. 4. Repeat 1 and 2 until convergence. 5. Subtract y from all x i , i. e., xi (t) ← xi (t) − (y(t − di ) · xi (t))y(t − di ) 6. Renormalize {x i } to unit variance. 7. Repeat Steps 1–5 to find next component.

6 Task-Related Component Analysis: Applications Here some illustrative applications of TRCA and its extensions are provided. Details and more extensive examples can be found in our previous papers [2, 3].

6.1 TRCA As a demonstration of how TRCA works, we first applied TRCA to synthetic data of block-design and event-design. A synthetic block-design data was created in two steps. First, four time series of hemodynamic responses, systemic Mayer wave, non-periodic low-frequency noise and spiky motion artifacts were generated, and then these time series were mixed with randomly chosen coefficients to emulate synthetic observation data. Five blocks of thirty seconds were assumed in constructing the hemodynamic responses and were therefore reproducible in every task block, whereas the other three time series (Mayer wave, non-periodic noise and motion artifacts) were not synchronized with task blocks and were not task-related. A synthetic event-data was created in the same way except that hemodynamic responses were generated in response to events that had some overlaps. We asked, when task onsets were given, whether TRCA could recover the hemodynamic responses both in the block- and the event-design data sets. Results of TRCA applied to block-design and event-design synthetic data are presented in Figs. 2 and 3, respectively. In both data sets, TRCA successfully extracted the hemodynamic responses as the primal components with the largest eigenvalues (therefore significantly task-related), and the others with negligible eigenvalues (therefore not significantly task-related). It is noteworthy that TRCA, originally designed for a block-design experiment, succeeded also in an event-design experiment. Encouraged by the successful application of TRCA to the synthetic data, we next applied to TRCA to fNIRS data of a finger tapping experiment with twenty-nine subjects [34]. The subjects performed a finger-tapping task either with left or right hand for five blocks of thirty seconds, interleaved with rest periods of thirty seconds.

162

H. Tanaka et al.

Fig. 2 Extraction of task-related component with variable activation amplitudes. Original signals in (a) (from top to bottom: hemodynamic response, Mayer wave, low-pass filtered Gaussian noise, and simulated motion artifacts) were randomly mixed to give observed time courses in (b). Extracted task-related components are shown with corresponding eigenvalules in (c). The dominat component (thick colored lines) was block-by-block compared with the task-related hemodynamics (thin, dashed colored lines) in (d). Adopted from [2]

Data Analysis Method for Neuroimaging Data: Task-Related Component …

a

b

c

d

163

Fig. 3 Application of TRCA to an event-related data. a Original signals, b input time courses, and c Extracted task-related components with corresponding eigenvalues. Vertical red dashed lines denote the timings of event onsets. d Dominant components aligned with event onsets (gray lines). The black thick line denotes the actual hemodynamic response used to create the synthetic data (i.e., five-second box-car function convolved with hemodynamic response function). Adopted from [2]

Twenty-four NIRS channels were placed over sensorimotor areas of both hemispheres (twelve channels over each hemisphere). The amplitudes of signals varied from channel to channel, so they were normalized to zero mean and unit variance before applying TRCA. Here only oxy-Hb data were analyzed. First, for individual

164

H. Tanaka et al.

a

b

c

d

Fig. 4 a One component of right finger tapping and b the corresponding projection map. c Another component of right finger tapping and d the corresponding projection map. Adopted from [2]

subjects, TRCA solved the generalized eigenvalue problem (Sect. 4.2) and then performed a statistical test based on resampling (Sect. 4.3). The number of statistically significant task-related components turned out to range from zero to three, and most subjects had two components. We then clustered these significant components using a k-means clustering analysis with k = 2. Figure 4 shows the results of the two components in the right finger-tapping experiment (similar results for left finger tapping, not included here). One component had a gradually changing time course that was similar to a conventional hemodynamics (Fig. 4a), and was lateralized to the contralateral hemisphere (the left hemisphere in this case) (Fig. 4b). Considering its time series and lateralized localization, this component was identified with a hemodynamic response to finger movements. The other component, on the other hand, was piece-wise linear or triangular in time (Fig. 4c), and was rather uniform in space over the both hemispheres (Fig. 4d). Its uniform distribution implied its systemic origin of movement execution. These two components were already reported in an analysis using ICA; however, to identify the two components required an additional selection criterion [35]. Note that TRCA requires no selection of components.

6.2 Discriminative TRCA Discriminative TRCA provides a succinct description to classify two tasks from fNIRS signals. We applied this method to the same data sets of finger tapping experiment used in the previous subsection. For a representative subject, Fig. 5a shows

Data Analysis Method for Neuroimaging Data: Task-Related Component …

165

a

b

c

d

Fig. 5 a The most distinctive time course computed from a representative subject. Blue and red shaded area denote the task periods of right and left finger tapping, respectively. b Task-block average of the time course in panel a. Thin blue and red lines denote individual right and left finger tapping blocks, respectively, and thick blue and red lines are their averages. c Task block average computed from all subjects. Error bars indicate standard errors. d Corresponding map averaged over all subjects. Adopted from [2]

a time course that were reproducible in each condition (blue and red for right and left tapping, respectively) and were negatively correlated across the two conditions, emphasized as a block average in Fig. 5b. We observed similar components across all subjects (Fig. 5c), with positive and negative maps in the right and left hemispheres, respectively.

166

H. Tanaka et al.

6.3 RCA± Finally we present the results of TRCA± applied to synthetic data and fNIRS data of a working memory experiment. First, a synthetic data was created by simulating hemodynamic responses using a balloon model [36, 37]. A balloon model generated time series of oxy-Hb, deoxy-Hb and local blood volume in response to a task, which was assumed to be block-designed. In addition, we included artifactual components of Mayer wave and body movements (Fig. 6a). The four time series were mixed with variable weights with an assumption that oxy- and deoxy-Hb contributed in a negatively correlated way whereas blood volume contributed in a positively correlated way

a

b

c e

d

Fig. 6 Reconstruction of CBV and oxygenation changes synthesized with a balloon model. a Synthetic time courses of oxygenation (oxy- and deoxy-Hb), CBV, and artifacts (the Mayer wave and body motion). b Simulated observed data (xoxy,i and xdeoxy,i ). c Reconstructed CBV change + and by TRCA+ and d oxygenation change by TRCA− . The red and blue solid lines depict yoxy + − − ydeoxy , respectively, in c, and yoxy and ydeoxy , respectively, in d. The black dashed lines depict the corresponding model CBV and oxygenation changes, respectively, in c and d (normalized to zero − + mean and unit variance for a comparison). e q−v phase plot of ydeoxy and ydeoxy . Adopted from [3]

Data Analysis Method for Neuroimaging Data: Task-Related Component …

167

to synthetic oxy- and deoxy-Hb (Fig. 6b). Note that from Fig. 6b a visual inspection could not identify the hemoglobin and blood-volume components. TRCA± successfully extracted oxy-Hb, deoxy-Hb and blood volume components (Fig. 6c). With these components, a so-called q–v plot could be depicted (Fig. 6d) [38]. We extend our application of TRCA± to a cognitive task of working memory (WM). Seventeen subjects participated in three sessions of a spatial and a verbal WM experiment [39]. Each session consisted of sixteen task blocks for a 10.5 s duration composed of the stimulus presentation period (1.5 s), the maintenance period (7.0 s), and the retrieval period (2.0 s). During the maintenance period the subjects were asked to memorize locations of white squares or types of Japanese characters that were presented in the presentation period for a spatial and a verbal task, respectively. Fifteen laser sources and fifteen detectors were located over the prefrontal cortex in a 3 × 10 lattice pattern, thereby providing forty-seven channels. We here present the results for the verbal WM experiment (similar results for the spatial WM experiment can be found in [3]. The oxygenation components exhibited negatively correlated time courses (Fig. 7a) and were localized around the dorsolateral prefrontal areas in both hemispheres (Fig. 7c), in consistent with previous fMRI studies. In contrast, the CBV components showed positively correlated time courses (Fig. 7b) and were localized in the ventral prefrontal cortex (Fig. 7d).

7 Summary This chapter provides a brief introduction to fNIRS and its analysis methods and introduces the formulation of TRCA and the applications to synthetic and fNIRS data. TRCA has a few advantages in comparison with extant methods; unlike the hypothesis-driven approach, there is no need for a hypothesis or a generative model, and unlike the data-driven approach, no subjective interpretation of components is necessary. Although motivated by fNIRS data analysis, the concept of reproducibility of experimental results is universal and the generic formulation of TRCA is readily applicable to data analysis of other neuroimaging modalities and biophysical signals in general.

168

H. Tanaka et al.

a

b

c

d

Fig. 7 a Task-related oxygenation and b task-related CBV changes found in the spatial WM experiment. The red and black solid lines represent yoxy and ydeoxy, respectively, along with standard errors indicated by the shaded areas. The blue shaded areas are the task periods of 10 s stating with the onset of stimulus presentation. Spatial maps for c oxygenation and d CBV changes. Adopted from [3]

Data Analysis Method for Neuroimaging Data: Task-Related Component …

Appendix A: MATLAB function of TRCA

function [Y, V, D, S, Q] = TRCA(X, t1, Nexp) % X : data matrix (N channels * T time points) % t1 : task onsets (vector in sampling unit) % Nexp : task duration (scalar in sampling unit)

Nchannels = size(X, 1); Nblocks = length(t1);

% data of task duration X = X -repmat(mean(X,2),1,size(X,2)); Xb = zeros(Nchannels, Nexp+1, Nblocks); for i=1:Nblocks Xb(:, :, i) = X(:, t1(i):t1(i)+Nexp) - repmat(mean(X(:, t1(i):t1(i)+Nexp),2),1,Nexp+1); end

169

170

H. Tanaka et al. % computation of correlation matrices: S = zeros(Nchannels); for i=1:Nblocks-1 for j=i+1:Nblocks S = S + Xb(:, :, i)*Xb(:, :, j)'; end end S = S+S'; Q = X*X';

% TRCA eigenvalue algorithm [V, D] = eig(inv(Q)*S); Y = V'*X; D = diag(D); [D, index] = sort(D, 'descend'); V = V(:,index); Y = Y(index,:);

end

Appendix B: MATLAB script for resampling statistical test

Nresample = 200; Dstats = [];

% number of resampling

% eigenvalues derived from the null hypothesis

for m=1:Nresample % compute randomized block onsets:

Data Analysis Method for Neuroimaging Data: Task-Related Component …

171

stimrnd = sort(fix( rand(size(stim))*(t(end)* -Nexp)+Nexp)); % compute eigenvalues with randomized block onsets: [Ydummy, Vdummy, Ddummy, Sdummy, Qdummy, ddummy] = TRCA(X, stimrnd, Nexp); Dstats = [Dstats; Ddummy]; end

% compute the thresholding value: q = quantile(Dstats, 0.99);

References 1. Popper KR (1959) The logic of scientific discovery. Hutchinson, London, United Kingdom 2. Tanaka H, Katura T, Sato H (2013) Task-related component analysis for functional neuroimaging and application to near-infrared spectroscopy data. Neuroimage 64:308–327 3. Tanaka H, Katura T, Sato H (2014) Task-related oxygenation and cerebral blood volume changes estimated from NIRS signals in motor and cognitive tasks. Neuroimage 94:107–119 4. Millikan GA (1942) The oximeter, an instrument for measuring continuously the oxygen saturation of arterial blood in man. Rev Sci Instrum 13:434–444 5. Jöbsis-VanderVliet F, Piantadosi C, Sylvia A, Lucas S, Keizer H (1988) Near-infrared monitoring of cerebral oxygen sufficiency. I. Spectra of cytochrome c oxidase. Neurol Res 10:7–17 6. Wray S, Cope M, Delpy DT, Wyatt JS, Reynolds EOR (1988) Characterization of the near infrared absorption spectra of cytochrome aa3 and haemoglobin for the non-invasive monitoring of cerebral oxygenation. Biochimica et Biophysica Acta (BBA)-Bioenergetics 933:184–192 7. Chance B, Zhuang Z, UnAh C, Alter C, Lipton L (1993) Cognition-activated low-frequency modulation of light absorption in human brain. Proc Natl Acad Sci USA 90:3770–3774 8. Hoshi Y, Tamura M (1993) Dynamic multichannel near-infrared optical imaging of human brain activity. J Appl Physiol 75:1842–1846 9. Kato T, Kamei A, Takashima S, Ozaki T (1993) Human visual cortical function during photic stimulation monitoring by means of near-infrared spectroscopy. J Cereb Blood Flow Metab 13:516–520 10. Villringer A, Planck J, Hock C, Schleinkofer L, Dirnagl U (1993) Near infrared spectroscopy (NIRS): a new tool to study hemodynamic changes during activation of brain function in human adults. Neurosci Lett 154:101–104 11. Maki A, Yamashita Y, Ito Y, Watanabe E, Mayanagi Y, Koizumi H (1995) Spatial and temporal analysis of human motor activity using noninvasive NIR topography. Med Phys 22:1997–2005 12. Yamashita Y, Maki A, Koizumi H (1996) Near-infrared topographic measurement system: Imaging of absorbers localized in a scattering medium. Rev Sci Instrum 67:730–732 13. Homae F, Watanabe H, Otobe T, Nakano T, Go T, Konishi Y, Taga G (2010) Development of global cortical networks in early infancy. J Neurosci 30:4877–4882 14. Minagawa-Kawai Y, van der Lely H, Ramus F, Sato Y, Mazuka R, Dupoux E (2011) Optical brain imaging reveals general auditory and language-specific processing in early infant development. Cereb Cortex 21:254–261

172

H. Tanaka et al.

15. Pena M, Maki A, Kovacic D, Dehaene-Lambertz G, Koizumi H, Bouquet F, Mehler J (2003) Sounds and silence: an optical topography study of language recognition at birth. Proc Natl Acad Sci USA 100:11702–11705 16. Sato H, Hirabayashi Y, Tsubokura H, Kanai M, Ashida T, Konishi I, Uchida-Ota M, Konishi Y, Maki A (2012) Cerebral hemodynamics in newborn infants exposed to speech sounds: a whole-head optical topography study. Hum Brain Mapp 33:2092–2103 17. Taga G, Asakawa K, Maki A, Konishi Y, Koizumi H (2003) Brain imaging in awake infants by near-infrared optical topography. Proc Natl Acad Sci USA 100:10722–10727 18. Cui X, Bryant DM, Reiss AL (2012) NIRS-based hyperscanning reveals increased interpersonal coherence in superior frontal cortex during cooperation. Neuroimage 59:2430–2437 19. Funane T, Kiguchi M, Atsumori H, Sato H, Kubota K, Koizumi H (2011) Synchronous activity of two people’s prefrontal cortices during a cooperative task measured by simultaneous nearinfrared spectroscopy. J Biomed Opt 16:077011 20. Ogawa S, Lee TM, Kay AR, Tank DW (1990) Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc Natl Acad Sci USA 87:9868–9872 21. Ogawa S, Lee TM, Nayak AS, Glynn P (1990) Oxygenation-sensitive contrast in magnetic resonance image of rodent brain at high magnetic fields. Magn Reson Med 14:68–78 22. Sato H, Yahata N, Funane T, Takizawa R, Katura T, Atsumori H, Nishimura Y, Kinoshita A, Kiguchi M, Koizumi H, Fukuda M, Kasai K (2013) A NIRS-fMRI investigation of prefrontal cortex activity during a working memory task. Neuroimage 83C:158–173 23. Friston KJ (1998) Modes or models: a critique on independent component analysis for fMRI. Trends Cogn Sci 2:373–375 24. McKeown MJ, Sejnowski TJ (1998) Independent component analysis of fMRI data: examining the assumptions. Hum Brain Map 6:368–372 25. Friston KJ, Holmes AP, Worsley KJ, Poline JP, Frith CD, Frackowiak RSJ (1994) Statistical parametric maps in functional imaging: a general linear approach. Hum Brain Map 2:189–210 26. de Cheveigne A (2010) Time-shift denoising source separation. J Neurosci Methods 189:113– 120 27. Sirotin YB, Das A (2009) Anticipatory haemodynamic signals in sensory cortex not predicted by local neuronal activity. Nature 457:475–479 28. Wyatt JS, Cope M, Delpy DT, Wray S, Reynolds EO (1986) Quantification of cerebral oxygenation and haemodynamics in sick newborn infants by near infrared spectrophotometry. Lancet 2:1063–1066 29. Wyatt JS, Cope M, Delpy DT, Richardson CE, Edwards AD, Wray S, Reynolds EO (1990) Quantitation of cerebral blood volume in human infants by near-infrared spectroscopy. J Appl Physiol 68:1086–1091 30. Yamada T, Umeyama S, Matsuda K (2012) Separation of fNIRS signals into functional and systemic components based on differences in hemodynamic modalities. PLoS ONE 7:e50271 31. Rao YN, Principe JC (2001) An RLS type algorithm for generalized eigendecomposition. In: Neural networks for signal processing XI, 2001. Proceedings of the 2001 IEEE signal processing society workshop. IEEE, pp 263–272 32. Woehrle H, Krell M, Straube S, Kim S, Kirchner E, Kirchner F (2015) an adaptive spatial filter for user-independent single trial detection of event-related potentials 33. Woody CD (1967) Characterization of an adaptive filter for the analysis of variable latency neuroelectric signals. Med Biol Eng 5:539–554 34. Sato H, Fuchino Y, Kiguchi M, Katura T, Maki A, Yoro T, Koizumi H (2005) Intersubject variability of near-infrared spectroscopy signals during sensorimotor cortex activation. J Biomed Opt 10:44001 35. Katura T, Sato H, Fuchino Y, Yoshida T, Atsumori H, Kiguchi M, Maki A, Abe M, Tanaka N (2008) Extracting task-related activation components from optical topography measurement using independent components analysis. J Biomed Optics 13:054008 36. Buxton RB, Wong EC, Frank LR (1998) Dynamics of blood flow and oxygenation changes during brain activation: the balloon model. Magn Reson Med 39:855–864

Data Analysis Method for Neuroimaging Data: Task-Related Component …

173

37. Buxton RB, Uludag K, Dubowitz DJ, Liu TT (2004) Modeling the hemodynamic response to brain activation. Neuroimage 23(Suppl 1):S220-233 38. Toyoda H, Kashikura K, Okada T, Nakashita S, Honda M, Yonekura Y, Kawaguchi H, Maki A, Sadato N (2008) Source of nonlinearity of the BOLD response revealed by simultaneous fMRI and NIRS. Neuroimage 39:997–1013 39. Sato H, Aoki R, Katura T, Matsuda R, Koizumi H (2011) Correlation of within-individual fluctuation of depressed mood with prefrontal cortex activity during verbal working memory task: optical topography study. J Biomed Opt 16:126007

Hirokazu Tanaka majored in theoretical physics and obtained a Ph.D. from Kyoto University, Japan, in 2000. After finishing the Ph.D. training, he switched to neuroscience with an emphasis on computational modeling, neuroimaging signal processing, and human psychophysics. He studied as a researcher at RIKEN Brain Science Institute, a postdoctoral fellow at Columbia University, a research specialist at the Salk Institute of Biological Studies, a senior researcher at Hitachi Advanced Research Laboratory, and an associate professor at Japan Advanced Institute of Science and Technology, before joining Tokyo City University, where he is currently a professor. His interests include computational modeling of sensory processing, motor control and motor learning, statistical signal processing of neuroimaging data, and mobile brain/body imaging (MoBI).

Towards Automated Processing and Analysis of Neuronal Big Data Acquired Using High-Resolution Brain-Chip Interfaces Mufti Mahmud, Claudia Cecchetto, Marta Maschietto, Roland Thewes, and Stefano Vassanelli

1 Introduction Technological advancements in the recent years have allowed scientists to have unprecedented access to brain data towards effective screening and diagnosis of diseases and devising their treatment [1]. Being the most complex organ in mammals, the brain’s decision-making and pattern recognition capability outperforms any existing computing machines developed to date [2]. To understand brain’s functionality, diagnose its disorders and devise relevant treatments, scientists have adopted different approaches [3]. One of the most popular approaches is to use suitable probes to record brain signals and monitor information transmission among neurons in the form of electrochemical signalling within large neuronal networks. Over the last two decades, micro- and nano-technologies have exponentially grown, especially in terms of novel miniaturized device development, allowing neuroscientists to monitor large neuronal populations and record their activities to decode brain functionality [2, 4–9]. When these novel neuronal probes with high spatio-temporal resolution are used for acquiring neuronal signals, they generate a large amount of data and analysing them to mine relevant information using automated, efficient, and intelligent processing methods is a big challenge [10]. To address this challenge M. Mahmud (B) Department of Computing and Technology, Nottingham Trent University, Clifton, Nottingham NG11 8NS, UK e-mail: [email protected]; [email protected] C. Cecchetto Okinawa Institute of Science and Technology, Okinawa 904-0495, Japan R. Thewes Sensor and Actuator Systems, Technische Universität Berlin, 10587 Berlin, Germany M. Maschietto · S. Vassanelli Department of Biomedical Sciences, University of Padova, Via F. Marzolo 3, 35131 Padova, Italy © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_8

175

176

M. Mahmud et al.

Fig. 1 Overview of modern neuroscience research. The data science domain bridges the computational and experimental neuroscience disciplines

the interdisciplinary ‘Neuroengineering’ community [11] has made available several automated software tools to extract different information from the data [3, 12–14]. Nonetheless, the majority of these methods were developed for systems with single or a few acquisition channels, which do not scale up very well for large multichannel recordings [15]. This demands the community to develop novel, automated, efficient and scalable analysis tools targeting multichannel neuronal data. In recent years, many of the experimental research paradigms have adopted a data driven approach where the computational and experimental neuroscience paradigms are bridged by data science (see Fig. 1) [16, 17]. This approach allows application of sophisticated methods to analyse the acquired data to derive parameters of synthetic experimentation which in turn dictates the design of novel neuroscience experiments. This has mainly been facilitated by machine learning, which has attracted a lot of attention in the recent years and has been successfully applied to tasks such as biological data mining [1], image analysis [18], financial forecasting [19], anomaly detection [20], disease detection [21], natural language processing [22] and strategic game playing [23]. Towards these goals, this chapter sheds lights firstly on the use of high spatial resolution Capacitively Coupled CMOS Multi-Electrode Arrays (CC-CMOS-MEAs) as neuronal probes in recording neurophysiological signals to study brain functions [2, 5, 7, 8, 24–29], and secondly—on the discussion about a set of ‘in house’ developed methods capable of efficiently processing and analysing the recorded neural signals [12, 30–43]. Towards the end, we indicate few major challenges in analysing large-scale neuronal data and provide some speculative research direction.

Towards Automated Processing and Analysis of Neuronal Big Data …

177

2 High-Resolution Neuronal Recording Probes To record neuronal activity, we have been using four generations of ElectrolyteOxide–Semiconductor Field Effect Transistors (EOSFET) based chips or CCCMOS-MEAs with varied number of recording electrodes and sampling frequencies. Out of the four generations, two are planar (for ex-vivo applications, Fig. 2a, c) and two are implantable (for in-vivo applications, Fig. 2d, e) chips. The ex-vivo chips have been used to record signals mainly from the rat cortical surface and superficial

a

d

b

e

c

Fig. 2 Four generations of high spatial resolution neuronal probes used for signal acquisition. Left column shows two generations of ex-vivo chips (a [24], c [29]) and a layout (b) of the chip shown in (a). Right column shows two generations of implantable needle chips (d [26], e [28])

178

M. Mahmud et al.

layers, whereas the implantable chips have been used to record signals from all the cortical layers.

2.1 EOSFET Probes with 62 Recording Sites The chips have two linear arrays, each consisting of 31 insulated EOSFETs, spaced 30–40 µm (see Fig. 2a), and 32 Electrolyte-Oxide–Semiconductor Capacitors (EOSCs) for stimulation purposes integrated in between two adjacent recording transistors (see Fig. 2b). The area covered by each recording site is either 3.1 µm × 7.2 µm or 3.5 µm × 9.0 µm [24, 25].

2.2 CC-CMOS-MEAs with 128 × 128 Recording Sites These devices have a size of 5.4 mm × 6.5 mm. They are wire bonded to a standard ceramic package with a Perspex chamber attached on top. The active area is 1 mm × 1 mm containing 128 × 128 recording sites. The pitch between two adjacent recording sites is 7.8 µm and the sampling rate is 2 kS/s (see Fig. 2c) [5, 27, 29].

2.3 Implantable EOSFET Probes with 4 × 1 Recording Sites Each of these first generation devices, consists of two parts: a needle of dimension 2 mm × 360 µm × 100 µm (length × width × thickness) with an array of four transistors (gate area 10 µm × 10 µm, pitch 80 µm) and a contact plate (10 mm long, 5 mm wide, 500 µm thick) with the bond pads [26].

2.4 Implantable CC-CMOS-MEAs with 16 × 16 Recording Sites This generation of devices features an array of 256 recording sites, arranged in a 16 × 16 matrix, placed on the tip of a 10 mm long needle which is 300 µm wide and 150 µm thick. These capacitively coupled recording sites are separated by 15 µm [7, 8, 28].

Towards Automated Processing and Analysis of Neuronal Big Data …

179

2.5 Acquired Extracellular Neuronal Signals To understand sensory information processing in the brain, neuronal signals were acquired from the ‘barrel cortex’—a part of the brain’s primary somatosensory cortex (S1) responsible for encoding and processing the dynamic information coming from the whiskers, and has been well studied and characterised—of anesthetized rats upon mechanical deflection of the whiskers (see [35] for details on the experimental procedure). Figure 3 shows representative local field potential (LFP) signals recorded by the different types of chips described in the previous subsections. Figure 3a shows a 3D plot of simultaneously recorded signals from 13 EOSFETs during an experiment with the planar chip shown in Fig. 2a. These signals were recorded by deflecting with air-puffs the whiskers of an anesthetized rat while it was placed upside down on the chip, such that the barrel cortex was in contact with the recording sites. Signal propagation along the different EOSFETs provide an electrical image of the cortical region in contact with the recording sites. The chip shown in Fig. 2c was also used in recording evoked signals with a similar strategy (see [5] for experimental procedure). Figure 3b shows single trial signals (top panel) and averaged signals of 50 trials each (bottom panel) sensed by the four recording sites of the probe (see Fig. 2d). Controlled mechanical deflection of one whisker by means of a piezoelectric bender evoked a neuronal response which was sensed differently at different locations. Figures 3c, d show whisker-evoked neuronal responses recorded by the chip shown in Fig. 2e. The ‘C’ panel shows the 3D plot of signals from one single column of the recording matrix, whereas, the ‘D’ panel shows time-lapse frames which capture the flow of the propagating LFP wave at very high spatial resolution.

3 Automated Analysis of Neuronal Signals The recorded signals described in the previous section require rigorous preprocessing and analysis for drawing meaningful conclusions. To facilitate this decoding and decision-making process, we developed several ‘in house’ methods (discussed below) as part of the ‘SigMate’ package [12, 30, 33].

3.1 Artifact Removal The stimulus induced signals are often contaminated with artifacts which partly or fully obscure real brain responses. In our data, we noticed two types of artifacts, depending on the stimulation protocol used: (1) slow, and (2) fast artifacts.

180

M. Mahmud et al.

Fig. 3 Neural signals acquired by the different types of chips. a Pseudo colour plot of a representative signal acquired by both the planar chips. b Time series signals acquired by the probe with 4 recording sites (top: single trials, bottom: averaged over 50 trials). c 3D pseudo colour plot of column 1 signals from the chip having a 16 × 16 array. d Six pseudo coloured frames at specific times showing the propagation of the LFP wave

Towards Automated Processing and Analysis of Neuronal Big Data …

181

Fig. 4 Slow stimulus artifact removal using peak-valley based method

3.1.1

Slow Artifact Removal

This type of artifacts was largely caused by the air-puff stimulation. They are difficult to remove as their frequency content is very similar to that one of the evoked responses. We developed a custom peak-valley detection-based method to estimate the artifact and then subtract it from the evoked response [37]. Figure 4 shows the outcome of the method.

3.1.2

Fast Artifact Removal

When intracortical microstimulations are applied to evoke responses, fast stimulation artifacts are noticed in the signals. Due to the complex shape and varied frequency of stimulations, removing this type of artifacts is a challenging job. Our in house developed method detects these artifacts by following signal trajectories through calculation of higher order signal derivatives and replaces them with interpolated signal segments. This robust method works well with different morphologies and frequencies of stimulation artifacts elicited by intracortical microstimulations. Figure 5 shows the artifacted and artifact removed signals using our in-house method [38].

3.2 Signal Quality Assessment Through Noise Characterisation Brain-chip interfacing setups use many electronic devices which introduce an in— separable electrical or thermal contamination to the recorded signals. We developed an automatic method, based on steady-state detection, to assess the quality of these multichannel recordings in terms of signal-to-noise ratio (SNR) and noise distribution as shown in Fig. 6. The steady-state detection method is based on higher order derivative calculation as reported in [36]. This method first detects the steady-state

182

M. Mahmud et al.

Fig. 5 Fast stimulus artifact removal using higher order derivative

Fig. 6 Signal quality assessment through steady-state detection. a The recorded raw signal with its detected prestimulus steady-state (FS), poststimulus steady-state (SS), their respective model fitting (FS-Fit and SS-Fit), and estimated measurement error of the SS (SS-ME). b Distribution and estimation of SS-ME as a signal quality metric

portion of the signal which occurs at the pre-stimulus region of the signal and uses that information to detect steady-state signal portions in the remaining of the signal from the end. Detection of steady-state portions of the signal, in turn, provides information about the signal portion representing the evoked response. Having separated these two portions of the signal, the steady-state’s SNR can be easily calculated to estimate the quality of the recorded signals.

3.3 Analysis of Evoked Potentials Cortical LFPs are considered as fingerprints of sensory information processing [42]. During this processing different cortical layers get activated at specific times as a

Towards Automated Processing and Analysis of Neuronal Big Data …

183

Fig. 7 Feature detection in LFPs. a LFP’s individual features which carry important information about the neuronal network. b Laminar profile of evoked responses with detected events

result of information segregation and integration [32]. The intra- and inter-columnar microcircuits in the barrel cortex encode and decode the ‘what’, ‘where’ and ‘when’ aspects of the sensory information triggered by the deflection of the whiskers [44].

3.3.1

Event Detection and Latency Calculation

To better understand the cortical circuitry we have developed a method which detects predefined events in the intra-cortically recorded laminar profile of LFPs [31] and estimates signal propagation delays or latencies across different cortical layers based on the detected events [34]. These latencies provide important insights on the activation of different cortical layers during sensory information processing [32] (see Fig. 7).

3.3.2

Network Decoding Through Current Estimation

We reconstructed the corresponding currents from the LFPs using the current source density (CSD) method. The CSD is a method that converts extracellular electric potentials recorded at multiple sites to estimates of current sources generating the measured potentials. From these estimated current densities, current sinks and sources were identified in order to reveal the underlying neuronal network generating those signals [32]. From each CSD, latencies of the sources and sinks were calculated. From these latencies neuronal connections and signal propagation pathways were derived and a simplified neuronal network model was obtained [45] (see Fig. 8).

184

M. Mahmud et al.

Fig. 8 Current source density estimate and network decoding. a Current source density profile corresponding to the laminar LFP profile shown in Fig. 7. In this profile the sinks have been shaded and there are two clear sink-source complexes (indicated by the two asterisks). b Simplified microcircuit network deduced from the analysis of sink-source propagation

3.3.3

Single Trial LFP Sorting

The brain is very stochastic. Despite the presence of highly specific patterns, brain activity also shows a high level of variability [42]. In traditional neurophysiology, to minimise this variability signals recorded with repeated stimulations are stimuluslocked averaged before analyses are performed [43]. Recently, it has been shown that this variability among single trial LFPs is a reflection of the background spontaneous activity and different brain states, which is lost during the averaging process [35]. This implies that the shape of individual LFPs carries useful information about the underlying neuronal network. To harvest this information, we developed a shape based single trial LFP sorting method [42]. This method uses template matching for single trial recognition and the iKMeans clustering technique to classify the recognized LFPs. Peak latencies of evoked responses were calculated from the clustered LFPs through event detection and are reported in Fig. 9.

4 Challenges and Outlook Despite the advancement in both interfacing technology and analysis methods, there are still several challenges requiring attention for future developments [46].

Towards Automated Processing and Analysis of Neuronal Big Data …

185

Fig. 9 Variabilities in single trial LFPs are represented in terms of the response peak latency

4.1 Interfacing Technology With the increasing number of recording sites in neuronal probes, obtaining stable electrophysiological signals with high Signal-to-Noise Ratio (SNR) is a big challenge [8]. Extracting the acquired data simultaneously from a large number of recording sites in a high spatio-temporal resolution neuronal probe still requires improvements [7]. Biocompatibility of the implantable probes for chronic implants is also an open question [47, 48]. Recent technologies (such as memristors [49]) facilitating the interfacing of biological and artificial neurons are yet to be fully explored [50]. The use of Brain-Machine Interfacing (BMI) in medicine and rehabilitation is still dominated by conventional EEG based BMI [51, 52], which could still be improved by applying wireless communication technologies as in-patient monitoring [53]. Moreover, only very few of the existing implantable neuronal probes eventually have been translated to clinical applications [54, 55].

4.2 Advanced Analysis On the other side, a conventional 128 channel signal acquisition system—acquiring data at 20 kHz sampling frequency with 16 bits analogue-to-digital conversion hardware—will fill up a 1 TB storage device with just ∼50 h of experimental data [13]. Making reliable conclusions from this staggering amount of data largely depends on development of new methods to analyse the acquired data in high-dimensional spaces using advanced statistical and machine learning pattern recognition algorithms [56, 57]. In addition to parallel processing, choosing appropriate analysis methods with specific techniques for data interpretation is very important. Ultimately, mapping behavioural activity to the acquired data representing neuronal connectivity,

186

M. Mahmud et al.

and seeking to understand their relationship is possible only by coupling data intensive theoretical frameworks capable of modelling the network with well-planned experimental design [58, 59]. Single workstation is lagging behind in processing such amount of data; thus, scientists have been exploring the possibility of using parallel computing for the analysis [13]. In addition, distributed and cloud-based analysis systems would greatly facilitate the time complexity [1, 14, 60].

5 Conclusion High resolution brain-chip interfaces allow to investigate brain functions at cellular levels. On one side, rapid evolution of the related technologies have made it possible to record simultaneously from multiple brain regions and recording sites. On the other side, the huge amount of generated data requires automated methods for interpretation and decoding of complex neuronal signals. To give a flavour of the state-of-the-art on brain-chip interfacing, we have reviewed four generations of high spatial resolution neuronal probes and their use in recording extracellular neuronal signals, followed by a discussion on some open-source automated methods developed by us/our lab to process and analyse the data acquired by those probes. Finally, we laid out several challenges and future perspectives. With the tremendous growth of neurotechnologies, scientists can acquire data from multiple levels and multiple sources. This poses a great challenge to the neuroscientific community to automatically process and analyse those data in order to find meaningful conclusions towards understanding brain’s functioning and to devise translatable technologies towards autonomous diagnosis and treatment strategies for treating brain diseases. This chapter introduced the reader to a set of high spatiotemporal resolution neuronal probes and automated methods for processing and analysis of extracellularly recorded neuronal signals from them. Towards the end, some perspective research lines—where future developments are expected—have also been outlined.

References 1. Mahmud M, Kaiser MS, Hussain A, Vassanelli S (2018) Applications of deep learning and reinforcement learning to biological data. IEEE Trans Neural Netw Learn Syst 29(6):2063– 2079. https://doi.org/10.1109/TNNLS.2018.2790388 2. Mahmud M, Cecchetto C, Maschietto M, Thewes R, Vassanelli S (2017) Towards highresolution brain-chip interface and automated analysis of multichannel neuronal signals. In: Proceedings R10- HTC, pp 868–872. https://doi.org/10.1109/R10-HTC.2017.8289091 3. Mahmud M, Vassanelli S (2016) Processing and analysis of multichannel extracellular neuronal signals: state-of-the-art and challenges. Front Neurosci 10:248. https://doi.org/10.3389/fnins. 2016.00248 4. Vassanelli S (2011) Brain-chip interfaces: the present and the future. Procedia Comput Sci 7:61–64. https://doi.org/10.1016/j.procs.2011.12.020

Towards Automated Processing and Analysis of Neuronal Big Data …

187

5. Vassanelli S, Mahmud M, Girardi S, Maschietto M (2012) On the way to large-scale and highresolution brain-chip interfacing. Cogn Comput 4(1):71–81. https://doi.org/10.1007/s12559011-9121-4 6. Vassanelli S (2014) Multielectrode and multitransistor arrays for in vivo recording. In: De Vittorio M, Martiradonna L, Assad J (eds) Nanotechnology and neuroscience: nano-electronic, photonic and mechanical neuronal interfacing. New York, Springer, pp 239–267 7. Schroder S, Cecchetto C, Keil S, Mahmud M, Brose E, Dogan O et al (2015) CMOScompatible purely capacitive interfaces for high-density in-vivo recording from neural tissue. In: Proceedings of BioCAS, pp 1–4. https://doi.org/10.1109/BioCAS.2015.7348358 8. Thewes R, Bertotti G, Dodel N, Keil S, Schroder S, Boven KH et al (2016) Neural tissue and brain interfacing CMOS devices—an introduction to state-of-the-art, current and future challenges. In: Proceedings of IEEE-ISCAS, pp 1826–1829. https://doi.org/10.1109/ISCAS. 2016.7538925 9. Jun JJ, Steinmetz NA, Siegle JH, Denman DJ, Bauza M, Barbarits B et al (2017) Fully integrated silicon probes for high-density recording of neural activity. Nature 551(7679):232–236. https:// doi.org/10.1038/nature24636 10. Landhuis E (2017) Neuroscience: big brain, big data. Nature 541(7638):559–561. https://doi. org/10.1038/541559a 11. Vassanelli S, Mahmud M (2016) Trends and challenges in neuroengineering: toward “intelligent” neuroprostheses through brain-“brain inspired systems” communication. Front Neurosci 10(438). https://doi.org/10.3389/fnins.2016.00438 12. Mahmud M, Bertoldo A, Girardi S, Maschietto M, Vassanelli S (2012) SigMate: a matlab-based automated tool for extracellular neuronal signal processing and analysis. J Neurosci Methods 207(1):97–112. https://doi.org/10.1016/j.jneumeth.2012.03.009 13. Mahmud M, Pulizzi R, Vasilaki E, Giugliano M (2014) QSpike tools: a generic framework for parallel batch preprocessing of extracellular neuronal signals recorded by substrate microelectrode arrays. Front Neuroinform 8(26). https://doi.org/10.3389/fninf.2014.00026 14. Mahmud M, Pulizzi R, Vasilaki E, Giugliano M (2014) QSpikeTools: an open source toolbox for parallel batch processing of extracellular neuronal signals recorded by substrate microelectrode arrays. In: Proceedings of ICEEICT, pp 1–6. https://doi.org/10.1109/ICEEICT.2014.6919177 15. Stevenson IH, Kording KP (2011) How advances in neural recording affect data analysis. Nat Neurosci 14(2):139–142. https://doi.org/10.1038/nn.2731 16. Mahmud M, Kaiser MS, Rahman MM, Rahman MA, Shabut A, Al-Mamun S, Hussain A (2018) A brain-inspired trust management model to assure security in a cloud based IoT framework for neuroscience applications. Cogn Comput 10:864–887. https://doi.org/10.1007/s12559-0189543-3 17. Mahmud M, Vassanelli S (2019) Open-source tools for processing and analysis of in vitro extracellular neuronal signals. In: In vitro neuronal networks. Springer, Cham, pp 233–250. https://doi.org/10.1007/978-3-030-11135-910 18. Ali HM, Kaiser MS, Mahmud M (2019) Application of convolutional neural network in segmenting brain regions from MRI data. In: Liang P, Goel V, Shan C (eds) Brain informatics. Springer, Cham. https://doi.org/10.1007/978-3-030-37078-714 19. Orojo O, Tepper J, McGinnity TM, Mahmud M (2019) A multi-recurrent network for crude oil price prediction. In: Proceedings of IEEE SSCI, pp 2953–2958. https://doi.org/10.1109/SSC I44817.2019.9002841 20. Yahaya SW, Lotfi A, Mahmud M (2019) A consensus novelty detection ensemble approach for anomaly detection in activities of daily living. Appl Soft Comput 83:105613. https://doi. org/10.1016/j.asoc.2019.105613 21. Noor MBT, Zenia NZ, Kaiser MS, Mahmud M, Al Mamun S (2019) Detecting neurodegenerative disease from mri: a brief review on a deep learning perspective. In: Liang P, Goel V, Shan C (eds) Brain informatics. Springer, Cham, pp 115–125. https://doi.org/10.1007/978-3030-37078-712 22. Rabby G, Azad S, Mahmud M, Zamli KZ, Rahman MM (2020) TeKET: a tree-based unsupervised keyphrase extraction technique. Cogn Comput. https://doi.org/10.1007/s12559-019-097 06-3, [epub ahead of print]

188

M. Mahmud et al.

23. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, Van Den Driessche G et al (2016) Mastering the game of go with deep neural networks and tree search. Nature 529(7587):484. https://doi. org/10.1038/nature16961 24. Maschietto M, Mahmud M, Stefano G, Vassanelli S (2009) A high resolution bi-directional communication through a brain-chip interface. In: Proceedings AT-EQUAL, pp 32–35. https:// doi.org/10.1109/AT-EQUAL.2009.18 25. Girardi S, Maschietto M, Zeitler R, Mahmud M, Vassanelli S (2011) High resolution cortical imaging using electrolyte-(metal)-oxide- semiconductor field effect transistors. In: Proceedings of NER, pp 269–272. https://doi.org/10.1109/NER.2011.5910539 26. Vassanelli S, Felderer F, Mahmud M, Maschietto M, Girardi S (2012) CyberRat probes: highresolution biohybrid devices for probing the brain. In: Proceedings of conference on biomimetic biohybrid systems, vol 7375 LNAI. https://doi.org/10.1007/978-3-642-31525-124 27. Hutzler M, Lambacher A, Eversmann B, Jenkner M, Thewes R, Fromherz P (2006) Highresolution multitransistor array recording of electrical field potentials in cultured brain slices. J Neurophysiol 96(3):1638–1645. https://doi.org/10.1152/jn.00347.2006 28. Cecchetto C, Schroder S, Keil S, Mahmud M, Brose E, Dogan O et al (2016) Imaging local field potentials in the rat barrel cortex. In: Proceedings of ICIIBMS, pp 296–299. https://doi. org/10.1109/ICI-IBMS.2015.7439533 29. Eversmann B et al (2003) A 128 x 128 CMOS biosensor array for extracellular recording of neuralactivity. IEEE J Solid-State Circ 38(12):2306–2317 30. Mahmud M, Bertoldo A, Girardi S, Maschietto M, Vassanelli S (2010) SigMate: a MATLABbased neuronal signal processing tool. In: Proceedings of EMBC, pp 1352–1355. https://doi. org/10.1109/IEMBS.2010.5626747 31. Mahmud M, Bertoldo A, Maschietto M, Girardi S, Vassanelli S (2010) Automatic detection of layer activation order in information processing pathways of rat barrel cortex under mechanical whisker Stimulation. In: Proceedings of EMBC, pp 6095–6098. https://doi.org/10.1109/ IEMBS.2010.5627639 32. Mahmud M, Pasqualotto E, Bertoldo A, Girardi S, Maschietto M, Vassanelli S (2011) An automated method for detection of layer activation order in information processing pathway of rat barrel cortex under mechanical whisker stimulation. J Neurosci Methods 196(1):141–150. https://doi.org/10.1016/j.jneumeth.2010.11.024 33. Mahmud M, Bertoldo A, Girardi S, Maschietto M, Pasqualotto E, Vassanelli S (2011) SigMate: a comprehensive software package for extracellular neuronal signal processing and analysis. In: Proceedings of NER, pp 88–91. https://doi.org/10.1109/NER.2011.5910495 34. Mahmud M, Maschietto M, Girardi S, Vassanelli S (2012) A matlab based tool for cortical layer activation order detection through latency calculation in local field potentials recorded from rat barrel cortex by brain-chip interface. In: Proceedings of BRC, pp 1–4. https://doi.org/ 10.1109/BRC.2012.6222170 35. Mahmud M, Cecchetto C, Vassanelli S (2016) An automated method for characterization of evoked single-trial local field potentials recorded from rat barrel cortex under mechanical whisker stimulation. Cogn Comput 8(5):935–945. https://doi.org/10.1007/s12559-016-9399-3 36. Mahmud M, Girardi S, Maschietto M, Rahman MM, Vassanelli S (2009) Noise characterization of electrophysiological signals recorded from high resolution brain-chip interface. In: Proceedings of ISBB, pp 84–87 37. Mahmud M, Girardi S, Maschietto M, Rahman MM, Bertoldo A, Vassanelli S (2009) Slow stimulus artifact removal through peak-valley detection of neuronal signals recorded from somatosensory cortex by high resolution brain-chip interface. IFMBE Proc 25(4):2062–2065. https://doi.org/10.1007/978-3-642-03882-2547 38. Mahmud M, Girardi S, Maschietto M, Vassanelli S (2012) An automated method to remove artifacts induced by microstimulation in local field potentials recorded from rat somatosensory cortex. In: Proceedings of BRC, pp 1–4. https://doi.org/10.1109/BRC.2012.6222169 39. Mahmud M, Girardi S, Maschietto M, Bertoldo A, Vassanelli S (2010) Processing of neuronal signals recorded by brain-chip interface from surface of the S1 brain cortex. In: Proceedings of NEBEC, pp 1–2. https://doi.org/10.1109/NEBC.2010.5458211.

Towards Automated Processing and Analysis of Neuronal Big Data …

189

40. Mahmud M, Travalin D, Bertoldo A, Girardi S, Maschietto M, Vassanelli S (2010) A contour based automatic method to classify local field potentials recorded from rat barrel cortex. In: Proceedinhs of CIBEC, pp 163–166. https://doi.org/10.1109/CIBEC.2010.5716087 41. Mahmud M, Girardi S, Maschietto M, Pasqualotto E, Vassanelli S (2011) An automated method to determine angular preferentiality using LFPs recorded from rat barrel cortex by brain-chip interface under mechanical whisker stimulation. In: Proceedings of EMBC, pp 2307–2310. https://doi.org/10.1109/IEMBS.2011.6090580 42. Mahmud M, Travalin D, Bertoldo A, Girardi S, Maschietto M, Vassanelli S (2012) An automated classification method for single sweep local field potentials recorded from rat barrel cortex under mechanical whisker stimulation. J Med Biol Eng 32(6). https://doi.org/10.5405/ jmbe.923 43. Mahmud M, Travalin D, Hussain A, Girardi S, Maschietto M, Felderer F et al (2012) Single LFP sorting for high-resolution brain-chip interfacing. In: Proceedings of BICS. 7366 LNAI, pp 329–337. https://doi.org/10.1007/978-3-642-31561-937 44. Diamond ME et al (2008) ’Where’ and ’what’ in the whisker sensorimotor system. Nat Rev Neurosci 9(8):601–612. https://doi.org/10.1038/nrn2411 45. Mahmud M, Travalin D, Hussain A (2012) Decoding network activity from LFPS: a computational approach. In: Proceedings ICONIP7663 LNCS, pp 584–591. https://doi.org/10.1007/ 978-3-642-34475-670 46. Birmingham K et al (2013) Bioelectronic medicines: a research roadmap. Nat Rev Drug Discov 13(6):nrd4351. https://doi.org/10.1038/nrd4351 47. Cianci E, Lattanzio S, Seguini G, Vassanelli S, Fanciulli M (2012) Atomic layer deposited TiO2 for implantable brain-chip interfacing devices. Thin Solid Films 520(14):4745–4748. https:// doi.org/10.1016/j.tsf.2011.10.197 48. Schroder S et al (2015) Three dimensional ALD of TiO2 for in-vivo biomedical sensor applications. In: Proceedings of IWASI, pp 21–24. https://doi.org/10.1109/IWASI.2015.718 4978 49. Gupta I, Serb A, Khiat A, Zeitler R, Vassanelli S, Prodromakis T (2018) Sub 100 nW volatile nano-metal-oxide memristor as synaptic-like encoder of neuronal spikes. IEEE Trans Biomed Circ Syst 12(2):351–359 50. Serb A, Corna A, George R, Khiat A, Rocchi F, Reato M et al (2020) Memristive synapses connect brain and silicon spiking neurons. Sci Rep 10(1):1–7 51. Mahmud M, Hawellek D, Bertoldo A (2010) EEG based brain-machine interface for navigation of robotic device. In: Proceedings of BioRob, pp 168–172. https://doi.org/10.1109/BIOROB. 2010.5627015 52. Daly JJ, Huggins JE (2015) Brain-computer interface: current and emerging rehabilitation applications. Arch Phys Med Rehabil 96(3):S1–S7. https://doi.org/10.1016/j.apmr.2015.01.007 53. Rajba S, Raif P et al (2013) Wireless sensor networks in application to patients health monitoring. In: Proceedings of CICARE, pp 1–4. https://doi.org/10.1109/CICARE.2013.658 3075 54. Shih JJ, Krusienski DJ, Wolpaw JR (2012) Brain-computer interfaces in medicine. Mayo Clin Proc 87(3):268–279. https://doi.org/10.1016/j.mayocp.2011.12.008 55. Jackson A (2016) Spinal-cord injury: neural interfaces take another step forward. Nature 539(7628):177–178. https://doi.org/10.1038/539177a 56. Sejnowski TJ, Churchland PS, Movshon JA (2014) Putting big data to good use in neuroscience. Nat Neurosci 17(11):1440–1441. https://doi.org/10.1038/nn.3839 57. Mahmud M, Kaiser MS, Hussain A (2020) Deep learning in mining biological data. arXiv preprint arXiv:2003.00108. 2020 Feb 28. Available at: https://arxiv.org/abs/2003.00108 58. Anonymous (2014) Focus on big data. Nat Neurosci 17(11):1429. https://doi.org/10.1038/nn. 3856 59. Al-jawahiri R, Milne E (2017) Resources available for autism research in the big data era: a systematic review. Peer J (1):e2880. https://doi.org/10.7717/peerj.2880 60. Mahmud M, Rahman MM, Travalin D, Raif P, Hussain A (2012) Service oriented architecture based web application model for collaborative biomedical signal analysis. Biomed Tech (Berl) 57(SUPPL. 1). https://doi.org/10.1515/bmt-2012-4412

190

M. Mahmud et al.

Mufti Mahmud is a Senior Lecturer of Computing at the Nottingham Trent University, UK. He received Ph.D. degree in Information Engineering from University of Padova—Italy, in 2011. A recipient of the Marie-Curie postdoctoral fellowship, he served at various positions in the industry and academia in India, Bangladesh, Italy, Belgium, and the UK since 2003. An expert in computational intelligence and big data technologies, Dr. Mahmud aims to develop secure and intelligent tools to advance healthcare access in low-resource settings. Dr. Mahmud serves as Associate Editor to the Cognitive Computation, IEEE Access, Big Data Analytics, and Brain Informatics journals. A senior member of IEEE and ACM, he is currently serving as Vice Chair of the Intelligent System Application Technical Committee of IEEE CIS, Member of the IEEE CIS Task Force on Intelligence Systems for Health and the IEEE R8 Humanitarian Activities Subcommittee, and Project Liaison Officer of the IEEE UK and Ireland SIGHT committee. Dr. Mahmud is also serving as Local organising chair of IEEE-WCCI2020; General Chair of BI2020 and BI2021; and Programme Chair of IEEE-CICARE2020 and IEEE-CICARE2021. Claudia Cecchetto is a biophysicist specialized in in vivo electrophysiological recordings from the rodent primary somatosensory cortex. She obtained her Ph.D. in Bioengineering from the University of Padova (Italy) in 2016, under the supervision of Prof. Stefano Vassanelli. Her doctoral thesis focused on the characterization of sensory information encoding by neuronal microcircuits in the rat brain, investigated through an innovative bi-directional and high-resolution brain-chip interface. In 2017 she has been awarded a Canon Foundation Research Fellowship and a Marie Curie Global Fellowship. She is currently a Postdoctoral Research Fellow at OIST (Okinawa Institute of Science and Technology Graduate University), in Okinawa (Japan) under the supervision of Prof. Bernd Kuhn. The aim of her current project “GRACE” is to develop an innovative and advanced dual approach to study how whisker deflections are encoded into the barrel cortex, combining Voltage Sensitive Dye (VSD) imaging and high-resolution electrical recordings. Marta Maschietto received her Master Degree in Biological Sciences in 2003 and her Ph.D. in Genetics and Developmental Molecular Biology in 2007 at the University of Padova. Since 2003 she has been working in Stefano Vassanelli’s team, first as a Ph.D. student and then from 2008 as a research assistant. Her skills span from molecular to cellular biology and in vivo recordings in the field of neuroscience research. She is mainly involved in the neurosurgical procedures and in vivo recording/microstimulation of neuronal signals from different brain areas of anesthetized rodents (brain cortex, hippocampus, cerebellum). She is skilled in recordings and stimulations with different types of commercial and non-commercial multisite probes. She currently cultivates primary neurons dissociated from rodent hippocampi on planar microchips featuring multi-electrodes arrays. She is also involved in developing a capacitive electroporation technique on mammalian cell lines and primary neuronal cells growing in adhesion on microchips. Roland Thewes received the Ph.D. degree in Electrical Engineering from University of Dortmund, Germany, in 1995. Since 2009, he is a full professor at TU Berlin. Since 1994, he has served at renowned companies including Siemens AG, Infineon Technologies, and Qimonda. He also served as a consultant of the Max-Planck Society in the area of CMOS-based neural interfacing between 2005 and 2009. He has (co-)authored more than 160 technical publications and a similar number of patents. He has been serving on various TPCs including IEEE ISSCC, IEEE IEDM, and IEEE ESSCIRC. He served as TP Chair of IEEE ESSCIRC, IEEE BioCAS, and IEEE ICECS. He will be General Chair of IEEE BioCAS 2021. He is a recipient of the German President’s Future Award (2004), the ISSCC 2002 Jack Raper Award (2003), and nine more paper and conference awards. Dr. Thewes served as elected member of the IEEE SSCS AdCom. Stefano Vassanelli received the M.D. and the Ph.D. degrees in molecular and cellular biology and pathology from the University of Padova, Padua, Italy, in 1992 and 1999, respectively. From 1993 to 2000, he was with the Oregon Graduate Institute of Science and Technology, Portland,

Towards Automated Processing and Analysis of Neuronal Big Data …

191

OR, USA, for one year, where he was investigating the mitochondrial proton transporter uncoupling protein, and the Max-Planck Institute for Biochemistry, Martinsried, Germany, for five years, where he was involved in the first neuron-transistor capacitive coupling. In 2001, he founded the NeurChip Laboratory, University of Padova. Since 2001, he has been involved in the development of neurotechnologies for recording, stimulation, and processing of signals generated by neuronal networks. He is currently a Professor of physiology and teaching both for the engineering and medical faculties.

Conclusion and Future Work

Conclusion and Future Work Vassiliy Tsytsarev

All achievements of mankind in science, art and technology exist thanks to the mind container which is the brain. All products of our civilization are adapted and limited by physiological properties of the brain. The brain is able to receive only certain types of information, process it only in limited volume and with limited velocity, but in the world we live in with all of our scientific and technological advances in the modern neurobiology, we are still limited in our understanding of how the brain works in normal and pathological mode. In the twentieth century, the invention of magnetic resonance imaging (MRI) was a very important step on the path to progress in understanding the physiology of the information processing in the brain. Functional MRI (fMRI) was established later, based on the monitoring of the blood-oxygen level-dependent (BOLD) signal, which is a neural correlate and does not require exogenous contrast [1]. Being a logical developing of MRI, fMRI currently is a sort of gold standard of the functional brain imaging. fMRI is based on the noninvasive technology that uses a strong magnetic field and high frequency electromagnetic waves to generate 3D images of the investigated object. This ability to visualize noninvasively not only the structure but the function of the brain’s areas is a major scientific advancement in neuroimaging. Nevertheless, being a very powerful neuroimaging method fMRI not only has its advantages, but disadvantages as well. A comprehensive analysis and interpretation of large volumes of electrophysiological data is not an easy task before neurobiologists since the parallel recording of signals at multiple scales generate a huge amount of data. The chapter written by Drs. Mufti Mahmud, Claudia Cecchetto, Marta Maschietto and Stefano Vassanelli, presents an overview of different generations of high resolution neural probes capable of providing high spatiotemporal electrical imaging of neural activities in the brain. V. Tsytsarev (B) University of Maryland, Baltimore, MD, USA e-mail: [email protected] © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020 V. Tsytsarev et al. (eds.), Functional Brain Mapping: Methods and Aims, Brain Informatics and Health, https://doi.org/10.1007/978-981-15-6883-1_9

195

196

V. Tsytsarev

Authors also describing the use of high spatial resolution Multi Transistor Arrays (MTA) as neuronal probes in recording neurophysiological signals to study brain functions, as well as processing and analyzing the recorded neural signals. Interpretation of obtained imaging data is complicated since the brain processes are complex and often non-localized. It is important that the BOLD signal is an indirect reflection of neural activity, and is therefore can be affected by non-neural changes in the organism. The temporal resolution of fMRI is limited. It takes a second to observe significant BOLD response after neural firing [2]. Nevertheless, fMRI allows for noninvasively recording brain signals without chemical or radiational risks and is now a widely used standardized method which allows neuroscientists to compare results across the world. Brain functional optical imaging has grown intensively within last years. In contrast with fMRI, it is based on the idea that photon absorption is highly responsive to tiny biochemical changes of the brain tissue. Optical imaging takes advantage of the various wavelengths in order to monitor different properties of the metabolic activity at the same time. Photons of different wavelengths can be used to observe brain function in vivo in the exposed brain and even through the skin and bone, e.g. transcranially. Main factor of limitation of this method is a high level of light absorption and scattering light in the living tissue. There a plenty of types of the functional brain optical imaging. Thus, optical coherence tomography (OCT) is a method for obtaining 3D images below the tissue surface [3]. Photoacoustic imaging (PA) uses laser pulses to generate heat, rapidly expanding the living tissues and enabling their properties to be imaged [4]. The technique can be used for a number of clinical and preclinical applications including blood oxygenation for the functional brain mapping. Functional near-inferred spectroscopy (fNIRS) and diffuse optical tomography (DOT) employ near-infrared photons which goes through the skull and reaching the brain tissue noninvasively [5]. The photon absorption reveals information about local metabolism in the brain. Molecular imaging methods answer questions about physiological activities of cells, transmembrane channels, receptors and neurotransmitters and how these are altered by disorders and traumas. Molecular imaging use various types of optic techniques and plenty of optical probes which are compounds that have been specially labeled to emit photons of various wavelengths to select the target cells from surrounding tissue. The manuscripts collection presented in this book captures some of the neurobiological and translational research. As most of the high impacted papers they are based on the interaction of science, engineering, and medical research to further our understanding of the brain biology and preventing neurological diseases. Chapter “Intrinsic signal optical imaging (ISOI): state-of-the-art with emphasis on pre-clinical and clinical studies” by Dr. Ron Frostig describing intrinsic signal optical imaging. Being “classical” functional brain optical imaging technique ISO remains one of the most powerful imaging methods for functional brain mapping. Being originally developed for fundamental research with animal objects it has been successfully adopted in clinical and preclinical neuroscience. Dr. Frostig’s review

Conclusion and Future Work

197

highlight recent developments and topics that were not covered in most of the other reviews. General description of the existing situation in fMRI research is presented in Dr. Hidenao Fukuyama’s book. The method of fMRI is rapidly moving from to wide clinical and preclinical application but there are a number of questions regarding the method that remains unclear. Some of these are observed in Dr. Fukuyama’s chapter. Some of the results obtained by fMRI study are presented here and illustrating the power of system approaches in experiments of cognitive processes and clinical studies. The connectome uses quantitative metrics to evaluate functional and anatomical information about neural pathways [6]. Taking together structural and functional evaluations represent the anatomic and physiological properties establishing a new paradigm for understanding the brain functioning by looking at neural connections [7]. This new approach is called connectomics and it is observed in the chapter of Drs. Jean Faber, Priscila Antoneli, Daniel Leal and Esper Cavalheiro. A very promising application of fMRI involves the measurement of the extent to which brain areas are functionally connected. Resting and activation fMRI studies have reported similar disruptions in functional connectivity in which might account in part for the decreased local glucose metabolism found with positron emission tomography (PET) scanning [8–10]. Another relatively recent application of MRI is diffusion tensor imaging (DTI). This method provides information about neuronal connectivity in the form of quantitative data on the directionality water diffusion which can show fiber orientation in the white matter. DTI has proven itself a powerful method for clinical and preclinical studies of the mechanisms underlying brain pathologies. The success of DTI in clinical neuroscience has demonstrated its great potential for studies animal models of neural disorders [11]. One of the fundamental questions of neuroscience—roles that glia may play in synaptic plasticity and brain function—is observed in the chapter “Neurons and Plasticity: what do glial cells have to do with this?” written by Drs. Nicolangelo Iannella and Michel Condemine. The brain is not simply composed of neurons alone: being a complicated organ it contains lot of glial cells. Astrocytes oligodendrocytes, and other glial cells have been estimated to make at least half of the brain. Electrophysiological studies demonstrated that glial cells are essentially electrically inert and therefore not involved into the processing of information in the brain. But recently have been obtained an evidence indicating that glial cells may not behave as the brain’s passive elements but play more active roles. Authors reviewed glial cell physiology, followed by a discussion of neural-glial signaling and current efforts in modeling neural-glial communications [12]. It is a serious technical problem to monitor neural activity by imaging methods under freely moving conditions for long time. To solve this problem it is necessary to use a minimally invasive or non-invasive imaging technology in an awake animal. Technology that combines optogenetics for controlling neurons and an imaging to recording neural activity definitely will meet the abovementioned demand. Reviewee

198

V. Tsytsarev

of such devices that can be applied to the brain non-invasively has been presented in the chapter of Dr. Kiyotaka Sasagawa. Photostimulation and fluorescence imaging can be performed simultaneously in the freely moving animal if the implantable device has a dual light emission diode (LED) and a miniature image sensor [13–15]. This type of device enables bidirectional communication with the neural network in vivo by means of light since it is succeeded in activating particular neurons by local light—stimulation, and the intracellular Ca2+ releasing by fluorescence imaging. Theoretical and practical principles of the data analysis of functional near infrared spectroscopy (fNIRS) are presented in the chapter “Data analysis method for neuroimaging data: Task-related component analysis and its applications to fNIRS data”, by Drs. Hirokazu Tanaka, Takusige Katura, and Hiroki Sato. Functional brain optical imaging data frequently contain external components that are not pertinent directly to brain activity. Therefore, data analysis plays a critical role in removing such noise and extracting relevant information. Authors provided an overview of signal processing methods with an emphasis on fNIRS [16, 17]. Recently developed method, called task-related component analysis (TRCA) allow maximizes the block-by-block reproducibility of a signal in one condition is proposed. The concept of signal reproducibility originally were developed for fNIRS data analysis but definitely has a wide range of applications in neuroscience data analysis. During the past few decades there have been huge advancements in clinical and preclinical neuroimaging which have made it possible to visualize activity of the living brain in normal and pathological conditions. Not only functional but structural imaging is related with neural and mental disorders. Thus, disruption of anatomical connectivity has been shown by diffusion tensor imaging (DTI) in schizophrenia, depression and autism spectrum disorders [9]. Being a method for detecting regional brain activity response to the cognitive task, event-related functional MRI (fMRI) has an even better perspective in clinical application. Biological meaningful molecules, like N-acetylaspartate compounds, choline, creatine can be detected in vivo by proton magnetic resonance spectroscopy [9]. Biologically active molecule compounds, labeled with positron-emitting radioisotope, such as C11 , are commonly used for the positron emission tomography (PET). This method allows for the use of various molecular targets, of which the activity can be visualized by preparing the target specific radiolabeled compound [9]. This method allows to monitor the local concentration of various neurotransmitter and enzyme as well as diseases’ markers like amyloid beta and tau peptides [10, 18]. In regards to preclinical neuroimaging studies, we definitely have to mention terahertz brain imaging (T-ray) [19]. Terahertz-wave (THz-wave) lies between the infrared and microwave parts of the electromagnetic spectrum. Recently, THz imaging techniques have been tested as candidates for functional brain imaging as they provide a high sensitivity for the detection of biomarkers combined with good biosafety, since T-ray is non-ionized [19–22]. Thus, it was shown that traumatic brain injury (TBI) area on the animal model can be well distinguished by T-ray imaging which have been compared with MRI results. T-ray absorption coefficients increased with the aggravation of brain damage, whereas the cell density decreased as the order of mild, moderate, and severe TBI.

Conclusion and Future Work

199

Being suitable for high temporal and spatial resolution imaging of the neural activity in animal experiments, voltage-sensitive dyes are currently replaced due to their toxicity by genetically encoded voltage-sensitive proteins. Thus, recently engineered novel genetical encoded voltage-sensitive probe “Voltron” has been successfully tested for in vivo imaging of the neural activity in mice, zebrafish, and fruit flies [23]. Showing pretty good signal-to-noise ratio, Voltron doesn’t require averaging and allows for single-trial recording of spikes and subthreshold voltage signals. This improvement enables neural activity imaging studies not only on the anesthetized but also with freely moving animals and will definitely help in the better understanding for correlation between neural activity and behavioral acts. As was mentioned in this book by many authors, brain mapping has made significant strides through development of advanced methods. Optical imaging methods are not only used in animal studies of the model of epileptic seizures but also been in clinical trials. Nevertheless, translation of such fundamental neuroscience application of functional brain mapping technologies into clinical setting remains complicated [24, 25]. In the chapter “Transcranial Dynamic Fluorescence Imaging for the Study of the Epileptic Seizures” by Dr. Kalchenko et al., authors reviewed current advances in the field, along with one clear focus on laser speckle contrast imaging. They concluded that functional brain optical imaging will play a key role in bridging between morphology and functional mapping of the brain, and will contribute to more accurate diagnostics and improved efficacy of the antiepileptic therapy. Coupling brain optical imaging with selective biomarkers is also making early diagnosis more effective for variable clinical and preclinical tasks. Since the initial application of optical and non-optical brain functional imaging methods in the twentieth century the field of variable imaging studies of the brain has been continuously growing. Data analysis and technological developments as well as novel areas of application keep advancing the field of clinical and preclinical neuroscience. Without a doubt, methods based on two- and three-dimensional visualization of metabolic processes will be improved, and the use of these methods in clinical and fundamental research will also grow.

References 1. Lu W, Dong K, Cui D, Jiao Q, Qiu J (2019) Quality assurance of human functional magnetic resonance imaging: a literature review. Quant Imaging Med Surg 9(6):1147–1162 2. Drew PJ (2019) Vascular and neural basis of the BOLD signal. Curr Opin Neurobiol 58:61–69 3. Wyl˛egała A (2018) Principles of OCTA and applications in clinical neurology. Curr Neurol Neurosci Rep 18(12):96 4. Yao J, Wang LV (2014) Breakthroughs in photonics 2013: photoacoustic tomography in biomedicine. IEEE Photonics J 6(2):1–6 5. Kim HY, Seo K, Jeon HJ, Lee U, Lee H (2017) Application of functional near-infrared spectroscopy to the study of brain function in humans and animal models. Mol Cells 40(8):523–532 6. Meoded A, Huisman TAGM (2019) Connectomics in brain malformations. Neuroimaging Clin N Am 29(3):435–444

200

V. Tsytsarev

7. Gore JC et al (2019) Functional MRI and resting state connectivity in white matter—a minireview. Magn Reson Imaging 63:1–11 8. Watabe T, Hatazawa J (2019) Evaluation of functional connectivity in the brain using positron emission tomography: a mini-review. Front Neurosci 13:775 9. Suhara T (2016) Neuroimaging in psychiatry: current methods and future direction. Psychiatry Clin Neurosci 70(7):259–260 10. Cecchin D et al (2017) Brain PET and functional MRI: why simultaneously using hybrid PET/MR systems? Q J Nucl Med Mol Imaging 61(4):345–359 11. Jang SH, Seo JP, Lee SJ (2019) Diffusion tensor tractography studies of central post-stroke pain due to the spinothalamic tract injury: a mini-review. Front Neurol 10:787 12. Liao L-D et al (2013) Neurovascular coupling: in vivo optical techniques for functional brain imaging. Biomed Eng Online 12(1):38 13. Kobayashi T et al (2016) ‘Optical communication with brain cells by means of an implanted duplex micro-device with optogenetics and Ca2+ fluoroimaging.’ Sci Rep 6(1):21247 14. Tokuda T et al (2016) CMOS-based opto-electronic neural interface devices for optogenetics. In: 2016 38th annual international conference of the IEEE engineering in medicine and biology society (EMBC), vol 2016, pp 6319–6322 15. Haruta M et al (2019) Chronic brain blood-flow imaging device for a behavioral experiment using mice. Biomed Opt Express 10(4):1557 16. Hong K-S, Zafar A (2018) Existence of initial dip for BCI: an Illusion or Reality. Front Neurorobot 12:69 17. Lee CW, Cooper RJ, Austin T (2017) Diffuse optical tomography to investigate the newborn brain. Pediatr Res 82(3):376–386 18. Zhang K et al (2014) Comparison of cerebral blood flow acquired by simultaneous [15 O] water positron emission tomography and arterial spin labeling magnetic resonance imaging. J Cereb Blood Flow Metab 34(8):1373–1380 19. Zhao H et al (2018) High-sensitivity terahertz imaging of traumatic brain injury in a rat model. J Biomed Opt 23(3):1–7 20. Oh SJ et al (2014) Study of freshly excised brain tissues using terahertz imaging. Biomed Opt Express 5(8):2837–2842 21. Gusev SI, Demchenko PS, Litvinov EA, Cherkasova OP, Meglinski IV, Khodzitsky MK (2018) Study of glucose concentration influence on blood optical properties in THz frequency range. Nanosyst Phys Chem Math 9(3):389–400 22. Gusev SI, Borovkova MA, Strepitov MA, Khodzitsky MK (2015) Blood optical properties at various glucose level values in THz frequency range, vol 9537, p 95372A 23. Abdelfattah AS et al (2019) Bright and photostable chemigenetic indicators for extended in vivo voltage imaging. Science 365(6454):699–704 24. Kalchenko V, Israeli D, Kuznetsov Y, Meglinski I, Harmelin A (2015) A simple approach for non-invasive transcranial optical vascular imaging (nTOVI). J Biophotonics 8(11–12):897–901 25. Kalchenko V, Israeli D, Kuznetsov Y, Harmelin A (2015) Transcranial optical vascular imaging (TOVI) of cortical hemodynamics in mouse brain. Sci Rep 4(1):5839

Vassiliy Tsytsarev holds a Ph.D. in Neuroscience from Saint Petersburg State University, Russia. Soon after graduation he moved to Japan and began working at the Brain Science Institute of RIKEN, and the Human Brain Research Center, Kyoto. Functional brain mapping, neural circuits and different types of brain optical imaging are his main scientific interests. In Japan, Vassiliy worked in the field of auditory neuroscience using intrinsic optical imaging (IOS) and voltagesensitive dye imaging. After seven years in Japan he moved to the United States, where he has worked at several universities; for the past six years, at the University of Maryland School. His current focus is on functional brain mapping, epileptic studies and neural network function in the rodent somatosensory system, which offers a perfect specimen for many types of neuroscience research, including models of neural diseases. Vassiliy is the author and co-author of more than 40

Conclusion and Future Work

201

publications in peer-reviewed magazines, and several book chapters. He is a senior editor for the Journal of Neuroscience and Neuroengineering, serves on the board of directors of the Society for Brain Mapping and Therapeutics (SBMT), and on the editorial boards of other scientific journals.