Search for Gender Difference in Functional Connectivity of Resting State fMRI © Dmitry Kovalev1 © Sergey Priimenko2 © Natalya Ponomareva3 1 Federal Research Center “Computer Science and Control” of Russian Academy of Sciences, Moscow, Russia 2 Lomonosov Moscow State University, Moscow, Russia 3 Research Center of Neurology, Moscow, Russia dkovalev@ipiran.ru mior12@mail.ru ponomare@yandex.ru Abstract. During past several year huge sets of fMRI data were obtained within Human Connectome Project. Despite this, technologies for scalable analysis of large amounts of data are rarely used to analyze whole data set. Authors conducted virtual experiment on a large sample of data taken from the HCP to find the gender differences in functional connectivity. A review of methods for search for the functional connectivity is fulfilled. Further analysis of distributed use and scalability on large datasets of rfMRI data is provided with the discussion of existing libraries and suggestions of how to integrate them with a distributed system. As a result, the distributed architecture of the software based on the Apache Spark framework is developed. Being fairly complex, it includes ontology, conceptual schema and workflow. The results of this experiment may be of interest to neurophysiologists for further analysis. Keywords: data intensive research, distributed infrastructure, problem solving in neurophysiology. community. Such large-scale data warehouses could 1 Introduction serve as the beginning for the use of technologies for Today in many branches of science it is necessary to analyzing large amounts of data in the neuroimaging of solve problems associated with increasing scale of data the human brain, yet there are some limitations. One of [1–3]. This led to the development of specialized tools, the reasons why the community of neurobiologists do not which primarily focus on structured data, but are use tools to work with large amounts of data is that increasingly being adapted for more general forms [4, 5]. standard file formats, such as NIFTI[10], are binary and Yet this tools and software are not widely used in data possess additional costs to deliver to distributed file intensive research and methodology to correctly apply systems. Another problem is that many distributed them has still to be developed. Different use-cases from systems do not effectively perform iterative algorithms, multidisciplinary fields can greatly impact the evolution such as principal component analysis (PCA) and the of this methodology and tools. independent component analysis (ICA), which are actively used in the field of neuroimaging. One of the most prominent examples of data intensive domains is the field of neurophysiology, where One of significant are of research in neurophysiology the amount of data has reached petabyte scale. is the study of gender difference in functional Neurophysiology allows to visualize the structure, connectivity [11]. For example, there is a study of army functions and biochemical characteristics of the brain. In veterans that experience physical and psychiatric particular, approaches to find the functional connections complications, including craniocerebral trauma, post- of the brain departments are being explored [6]. One way traumatic stress and depression. The integration of a to do that is to measure the functional connectivity large number of women into military operations attracted between brain regions as the level of co-activation of attention to the potential sexual differences in the spontaneous functional time series of resting-state fMRI frequency and recovery from craniocerebral trauma, as [7–9]. well as from other concomitant disorders. Understanding the role of gender-related effects can provide information During past several years, major projects such as the on the needs for evaluating treatment for women, which Human Connectome Project (HCP) and the 1000 can demonstrate both similarities and differences from connectome have started with more than a thousand men. people participating. Datasets are open to the scientific This article aims at developing approach for a distributed analysis of data intensive neurophysiology Proceedings of the XIX International Conference domain. The article is structured as follows. Section 2 “Data Analytics and Management in Data Intensive surveys existing distributed methods ant tools to process Domains” (DAMDID/RCDL’2017), Moscow, Russia, and analyze neurophysiological datasets. Section 3 October 10–13, 2017 150 presents domain ontology that was created to better 2.2 Data analysis methods interact with domain experts. Section 4 describes The data of each subject is represented as a matrix distributed programming implementation on the existing (see Fig. 1), where each row represents a set of computational infrastructure, as well as output results. voxels of the brain at a particular time, and each column Section 5 concludes the article. is a time series for the corresponding voxel [12]. It is assumed that the data has already been pre-processed to remove artifacts and scaled to a standard space (coordinate system) so that the voxels are anatomically compatible for all subjects. It is also assumed that the time series of each voxel is shifted by its mean (and, possibly, normalized to the variance) [13]. If the data set consists of one object, in order to reduce the dimensionality of the data, the PCA is applied: where is the number of main components (usually much smaller than ), is the set of temporal Figure 1 Transformation of 4-D array into 2-D array eigenvectors, is the set of spatial eigenvectors, and the corresponding eigenvalues on the main diagonal of the 2 Data analysis methods matrix ( largest eigenvalues). Then, ICA is applied to the matrix , estimating a new set of spatial 2.1 Data processing components that are linear combinations of the vectors Resting-state fMRI dataset from the HCP project is of the matrix and are maximally independent of each used. The HCP consortium has developed an information other. If the data set consists of several subjects, then platform for storing raw and processed data, systematic initially all the data is combined into one large set processing and analysis of data, obtaining and consisting of s subjects, and then PCA and ICA are researching data. One of the main components of the applied. The resulting approximation will be the same as project is ConnectomeDB. ConnectomeDB provides above, but now with dimensions (see Fig. 2). database services for the storage and dissemination of With large data sets, or with a large number of datasets that are open to the scientific community. The subjects, it becomes unreasonable to form a complete set data is already preprocessed. Preprocessing consists of of data, and then apply PCA and ICA due to memory and removing spatial artifacts, distortion, surface formation time limitations. To solve this problem, several and alignment to a single standard space. algorithms were invented. Data processing is divided into two parts: data cleaning inside the brain (FMRIVolume) and on the brain surface (FMRISurface) [3]. At the FMRIVolume stage, spatial distortion removal, volume redistribution due to subject movement during the session, normalization of 4D images to the standard value and creation of the final brain mask are done. The main purpose of FMRISurface is to display time series in the standard CIFTI space. This is achieved by Figure 2 PCA for concatenated data comparing the voxels in the cortical region of the gray matter to the native surface of the cortex and transforming each subcortical region for each individual to a standard set of voxels for each data set. After processing the data, resting-state fMRI time series are stored in a special format – NIFTI. As a result, the data obtained with the resting-state fMRI yields more than 10 TB obtained for more than 1000 people. During the experiment, each patient was placed in a dark room and asked to relax, but not to fall asleep. The experiment was conducted in 4 sessions for 15 minutes. Two Figure 3 Parallel execution of PCA sessions of the fMRI device took pictures from the left side of the brain to the right side of the brain, and the In 2001, it was suggested to approximate the other two sessions from the right side of the brain to the concatenation of all data sets by first reducing each set of left. data to m main spatial vectors using PCA and then concatenating them and applying the final PCA to reduce 151 the final dataset to n components and then apply ICA 2.3 Libraries [14]. Although using a small value of limits the Nibabel [16] is a library that provides an API for memory requirements for these operations, the data size reading and writing some common file formats for is scaled linearly with the number of objects, which can neuroimaging. These formats include: ANALYZE eventually become impractically large. In addition, an (plain, SPM99, SPM2 and higher), GIFTI, NIfTI1, important piece of information can be lost if is not NIfTI2, MINC1, MINC2, MGH and ECAT, as well as relatively large (usually it should not be large). Philips PAR/REC. Different image format classes Information can be difficult to assess at the level of an provide full or selective access to header information individual subject, but it can be important at the group (meta), and access to image data is made available level (see Fig. 3). through the arrays of the numpy library. To overcome these limitations, the MELODIC's Objects of the image of nibabel consist of three Incremental Group-PCA (MIGP) algorithm was elements: proposed[15]. MIGP is an incremental approach, the 1. The n-dimensional array containing the image goal of which is to provide a very close approximation to 2. Matrix of affine transformations of size 4x4, the complete concatenation of the data set followed by which correlates the image coordinates with the the PCA, but without large memory requirements. High standard world coordinate space. accuracy is achieved due to the fact that individual sets of subjects’ data do not decrease to a small number of 3. Image metadata, stored in the header. components of PCA. The incremental approach When an image is loaded, an object of type preserves the inner space of PCA from weighted Nifti1mage is created. The file name can have an spatial eigenvectors, where is usually larger than the extension of both .nii and .nii.gz. number of time points in each individual data set. By It is worth noting that when the load function is called “weighted” is meant that the eigenvalues are included in directly, image data is not loaded into memory, since the matrix of spatial eigenvectors. The final set of m images can be stored as a numpy array or stored on a components representing the temporarily concatenated disk. To load data from a disk, you need to call the output of the PCA can then be reduced to the required get_data() function of an object of type Nifti1Image. dimension n simply by storing the upper n components This function returns an n-dimensional numpy array. and, if necessary, discarding the weighting coefficients In addition, an object of type Nifti1Image is created (eigenvalues). from numpy arrays. To do this, one should pass an n- Usually, 2–3 sets of data are first concatenated. This dimensional data array and an affine transformation data set is then fed into an m-dimensional PCA and matrix to the Nifti1Image constructor must. following matrix is obtained: . Nitime[17] is a library for the analysis of time series Each vector is multiplied by its own value. The in the field of neuroimaging. Nitime can be used to eigenvalues characterize the importance of the represent, process and analyze time series data from component here, so statistical information is not lost. experimental data. The main purpose of the library is to becomes the current evaluation of the group set and can serve as a platform for analyzing data collected in be considered as a matrix of pseudo-series consisting of neurophysical experiments. The basic principle of nitime m time points and v voxels. For each data set of each implementation is the division of time series subject, we gradually update by combining with representation and time series analysis. each data set and applying the ICA to get the updated An important feature of the nitime library is lazy , saving only m main components. Thus, the variance initialization. Most attributes of both time series and of each batch of data is preserved (see Fig. 4). analysis objects are used only when necessary. That is, the initialization of a time series object or an analysis object does not cause any intensive calculations. In addition, after the calculation starts, the object is saved and ensures that access to the results of the analysis will cause the calculation to be performed only when the analysis is performed for the first time. After that, the result of the analysis is saved for further use. Figure 4 MELODIC Incremental Group PCA One of the algorithms of the nitime library is the MIGP does not increase the memory requirement correlation analysis of brain regions. It calculates the with an increase in the number of subjects, large matrices correlation between one time series that represents a are never formed, and the computation time varies given area of the brain, with other areas that are also linearly with the number of objects. This is easily represented by a time series. To calculate the correlation parallelized by applying the approach in parallel to between regions in the nitime library, there is a subsets of entities, and then combining them using the SeedCoherencAnalyzer function that takes two time same approach of “concatenation and reduction” series inputs and returns a correlation matrix that can be described above. used for further analysis. Nilearn [18] is a Python module for statistical 152 processing of neuroimaging data. certain feature. Most often presented in the form of It uses scikit-learn module for multidimensional time series [20]. statistics with applications in intelligent modeling, • Voxel is an element of a three-dimensional image classification, decoding, and connectivity analysis. containing some value. Nilearn can work NiftiImage objects from the nibabel • Independent models - a model for investigating library. thefunctional connectivity of the entire brain. They Nilearn library has great functionality for working are designed to search for general patterns of with nii-images. It allows visualizing, decoding, functional connectivity between brain exploring the functional connectivity, and performing regions.Dependent models are a model for analyzing various manipulations, such as smoothing, marking and the correlation of a given region of the brain. advanced statistical analysis. • Brain connectivity – the structure of anatomical connections, statistical dependencies or cause-effect Nilearn provide CanICA method that is the ICA interactions between individual units within the method for analyzing fMRI data at the group level. brain's nervous system [21]. Compared to other strategies, it brings a well-controlled • Structural connectivity refers to a network of group model, as well as a threshold algorithm that physical or structural links linking sets of neurons or controls specificity and sensitivity with an explicit signal neural elements to structural biophysical features model. [22]. In order to get a time series and build a correlation • Functional connectivity is a statistical type of matrix for it, nilearn provides the NiftiMapsMasker connection between anatomically unconnected areas object. To create an object, one needs to specify an atlas of the brain that have common functional properties of the brain regions.Nilearn provides the ability to create [7]. a correlation matrix for independent components that • Effective connectivity – the combination of iscomputed by CanICA. structural and functional connectivity. It describes 3 Ontology the networks of directions of one neural element over another. The study of neuroimaging with large amounts of • The resting-state fMRI is a neural image obtained as data represents the intersection of different areas of a result of an experiment when the subject was at rest science. In order to use the same terms and concepts, and did not engage in active tasks. simple ontology was developed that describes the main • The task fMRI is the neuro-images obtained as a entities used in this work and a conceptual schema that result of the experiment, when the subject performed defines the types of data, constraints on these data types active actions, e. g., listened to music. and the means of interaction between them. Ontology is a formal specification of shared conceptualization [19]. 4 Implementation 4.1 Laboratory cluster specifications Virtual experiment was executed on the laboratory cluster (see Fig. 6). It consists of 2 master nodes and 6 slave nodes. Each master node has 32Gb of RAM, 24 threads and 2 Tb of disk space in RAID1. Slave nodes have 64Gb of RAM, 24 threads and 4 Tb of disk space attaches as JBOD. All the machines are connected to 10Gbs switch. Figure 5 Main concepts of the domain ontology The ontological specification of the subject area of neuroimaging consists of the following components (see Fig. 5): • Neuro-image – a 3-dimensional or 4-dimensional image (a series of 3-dimensional images), reflecting the distribution of metabolic activity in different Figure 6 Cluster Architecture regions of the brain in different time intervals [20]. On the cluster, the Hortonworks Data Platform (HDP) distribution package is installed. This distribution • The area of the brain is a set of voxels, sorted by a 153 represents a set of tools from the Hadoop infrastructure • Transformations are operations (for example, running Apache Ambari. mapping, filtering, merging, etc.) performed over A distributed file system (HDFS) (Hadoop RDD. The result of the transformation is a new Distributed File System) is installed file system. HDFS RDD containing its result. consists of a NameNode server and DataNode servers. • Actions are operations (eg, reduction, count, etc.) The NameNode server manages the namespace of the file that return a value that results from some system and manages the clients' access to the data. The calculations in RDD. main NameNode server is installed on the m1node and The cluster has Spark History Server installed on m1, records all transactions associated with changing the file Spark Thrift Server on m2, Livy Server on m2 and Spark system metadata to a special file called EditLog. When rt Clients on all nodes. the main NameNode server is started, it reads the HDFS For more convenient programming on a cluster, we image and applies all the changes to it. This is done once use Apache Zeppelin – a web-based notebook that allows at startup. A similar operation is performed by the to conduct interactive data analytics. It supports many Secondary NameNode, which is installed on the m2 interpreters, including the Spark interpreter and the machine. On machines s1-s4 DataNode servers are Python interpreter. installed, which are responsible for storing the data itself and keeping its integrity. Scalability. As of algorithm used, each slave machine handles several independent fMRI images, so For the sharing, scalability and reliability of the scalability increases almost linearly with using more Hadoop cluster, a resource manager YARN [5] is used. slave nodes. It is bounded by the network speed when YARN offers a hierarchical approach to the cluster transmitting initial image data into slave memory, infrastructure. The root of the YARN hierarchy is the however the transmission time is several seconds and is ResourceManager. This daemon manages the entire negligible compared to processing time. cluster and assigns applications to the underlying computing resources. It allocates resources (computing 4.2 Workflow resources, memory, and bandwidth) for the basic Workflow is depicted on Fig. 7. The program reads all NodeManager. ResourceManager interacts with files from the directory, checks the validity of the format ApplicationMaster when allocating resources and with (all data are compressed zip folders). After that, the NodeManager when starting and monitoring basic applications. ResoureManager is located on m2, and subject number is extracted from the file name and its gender is checked using an additional metadata file. NodeManager on nodes s1–s4. When the gender is known, the file is unzipped to the Another important module for the Hadoop cluster is corresponding folder. Inside the unzipped folder is a 4-D the Zookeper. ZooKeeper is a server that coordinates image in the .nii.gz format. Using the nibabel library, the distributed processing. It provides a distributed image is loaded into memory as an array of type configuration service, a synchronization service, and a numpy.array. From this array, a new array is created with registry of names for distributed systems. Distributed information about the spatial coordinates before the applications use ZooKeeper to store and notify updates value of the voxel. The new array is compressed by the of important configuration information. The Zookeper gzip algorithm and stored in HDFS. server is running on the m1 node. Due to Apache Spark limitations files larger than 2.5 Since most of the calculations are iterative GB in binary format can not be loaded. In the algorithms, Apache Spark was chosen as the uncompressed form, the sizeis 4.3 GB, so file needs to be computational backbone. Apache Spark provides a fast compression. After compression, the file occupies just and versatile platform for data processing. In comparison 700 MB. with Hadoop, Spark accelerates the work of programs by minimizing disk input-output operations. Spark task is started with the following parameters: • num-executor=4 – number of executable entities; In Spark, the concept of RDD (stable distributed data • executor-memory=25 GB – the amount of memory set) is introduced – an unchangeable fault-tolerant used for one execution process; distributed collection of objects that can be processed in • executor-cores=2 – the number of cores used for parallel. RDD can contain objects of any type. RDD is each executive entity. created by loading an external data set or distributing a • driver-memory=8 GB – the amount of memory used collection from the main program (driver program). In for the driver process, that is, where SparkContext RDD, two types of operations are supported: is initialized. Figure 7 Workflow 154 YARN creates on each node a container that receives hypotheses. As a result, a binary matrix is obtained that information from the driver. All calculations occur in two shows the deviation or acceptance of hypotheses for each streams. When metadata is received, a file with a area of the brain. compressed binary array is loaded into memory. The 4.3 Results program decompresses it and converts it to a normal array without information about the indexes. Then, using In total 50 male and 50 female subjects are used. All the resulting array and affinity transformation matrix, data is resting-state fMRI images. Nifti1Image and the CanICA object are created with the Fig. 8 depicts a binary matrix of gender differences following parameters: n_components=20; in the functional connectivity of healthy middle-aged Smoothing_fwhm=6; N_init=10; Threshold=3; people. Red spots mark areas that correlate both in men Verbose=10. and women, and blue dots indicate a lack of correlation. The CanICA object is passed to the Nifti1Image For example, this experiment shows that the upper front object and an image consisting of 20 components is (Superior Frontal Gyrus) of the brain has a significant output. This image is returned to the m1 driver. Thus, correlation with the insular cortex (Insular Cortex), but each node receives a portion of the paths to the does not have a significant correlation with the front part compressed images, processes them, and returns the (Frontal Pole) of the brain. result to the driver. The task is executed until all the files The independent components of averaged male specified for analysis on the m1 driver are processed. subject show a greater functional connectivity compared When the nodes complete the tasks, the driver comes to women. It can be seen that the main activity of the with a list of Nibabel1Image objects that contain brain of men and women occurs near its cortex. independent components. The data for all objects is averaged and a time series is created using the NiftiLabelsMasker object. A map of regions of the brain is transferred to the constructor of the NiftiLabelsMasker object. Using the ConnectivityMeasure object, which is created with the correlation parameter, the correlation matrix for the brain regions is considered.The correlation matrix for men and women is calculated separately. After this, the Fisher transform (z-transform) is applied to each matrix. After a new sample is calculated, which is obtained as the difference between the male z_m obtained and the female sample z_w. This sample will have a normal Figure 9 Averaged independent components for men (upper) and women (lower) 5 Conclusion This paper presents distributed methods and means for searching gender differences in functional connectivity of resting-state fMRI were explored. Several methods for the search for functional connectivity of functionally magnetic resonance tomography of human rest are considered. To work with large amounts of data, machine learning methods were used to identify repetitive patterns and to intelligently reduce data. Their possibilities of parallel and distributed use and scaling are investigated with large amounts of input data. For the sake of better communication with domain experts the domain ontology was specified with Figure 8 Binary matrix of functional connectivity main entities that describe this area and the necessary difference links between them. The review of existing means of preparation and distribution with a mathematical expectation of 0 and a preprocessing of data on local and distributed systems is variance of 2/(n-3). For this sample, calculates a critical carried out. At the moment there are few libraries for area with a significance level of 0.05 and c is corrected working with the NIFTI format on a distributed system, for multiple testing of the Benjamin–Hochberg so the input and output procedures for data were 155 implemented in this work. To preprocess the data, we [9] Biswal, B.B., Mennes, M., Zuo, X.-N., Gohel, S., used method compositions from the nibabel and nilearn Kelly, C., Smith, S.M., Beckmann, C.F., libraries.To solve the problem, an overview of existing Adelstein, J.S., Buckner, R.L., Colcombe, S., distributed systems was made, among which the Apache others: Toward Discovery Science of Human Brain Spark framework was most effective. For the Function. Proc. of the National Academy of experiment, a cluster of 6 machines was taken, where the Sciences, 107, pp. 4734-4739 (2010) two machines were the main nodes, and 4 the workers. [10] Cox, R.W., Ashburner, J., Breman, H., Fissell, K., On the cluster, the minimum set of programs required for Haselgrove, C., Holmes, C.J., Lancaster, J.L., the experiment, such as YARN, HDFS, ZooKeeper, Rex, D.E., Smith, S.M., Woodward, J.B., others: A Spark and Zeppelin notebook was installed and (sort of) New Image Data Format Standard: Nifti- configured. 1. Neuroimage, 22, e1440 (2004) A virtual experiment was performed in a distributed [11] McGlade, E., Rogowska, J., Yurgelun-Todd, D.: system. The time of this experiment was 4 hours for 400 Sex Differences in Orbitofrontal Connectivity in GB of data. As a result of the experiment, matrices of Male and Female Veterans With TBI. Brain connectivity between the brain regions of men and imaging and Behavior, 9, pp. 535-549 (2015) women were obtained, as well as a binary matrix of [12] Smith, S.M., Hyvärinen, A., Varoquaux, G., gender differences in functional connectivity. Miller, K.L., Beckmann, C.F.: Group-PCA for Acknowledgments Very Large fMRI Datasets. NeuroImage, 101, pp. 738-749 (2014) This research was partially supported by the Russian [13] Beckmann, C.F., Smith, S.M.: Probabilistic Foundation for Basic Research (projects 15-29-06045, Independent Component Analysis for Functional 16-07-01028). Magnetic Resonance Imaging. IEEE Transactions References on Medical Imaging, 23, pp. 137-152 (2004) [14] Calhoun, V.D., Adali, T., Pearlson, G.D., Pekar, J.: [1] Council, N.R.: Frontiers in Massive Data Analysis. A Method for Making Group Inferences from The National Academies Press, Washington, DC Functional MRI Data Using Independent (2013) Component Analysis. Human Brain Mapping, 14, [2] Hey, A.J., Tansley, S., Tolle, K.M., others eds: The pp. 140-151 (2001) Fourth Paradigm: Data-Intensive Scientific [15] Rachakonda, S., Silva, R.F., Liu, J., Calhoun, V.D.: Discovery. Microsoft Research Redmond, WA Memory Efficient PCA Methods for Large Group (2009) ICA. Frontiers in Neuroscience, 10 (2016) [3] Van Essen, D.C., Smith, S.M., Barch, D.M., [16] Gorgolewski, K., Burns, C.D., Madison, C., Behrens, T.E., Yacoub, E., Ugurbil, K., Clark, D., Halchenko, Y.O., Waskom, M.L., Consortium, W.-M.H., others: The WU-Minn Ghosh, S.S.: Nipype: A Flexible, Lightweight and Human Connectome Project: An Overview. Extensible Neuroimaging Data Processing Neuroimage. 80, 62–79 (2013) Framework in Python. Frontiers in [4] Zaharia, M., Xin, R.S., Wendell, P., Das, T., Neuroinformatics, 5 (2011) Armbrust, M., Dave, A., Meng, X., Rosen, J., [17] Rokem, A., Trumpis, M., Perez, F.: Nitime: Time- Venkataraman, S., Franklin, M.J., others: Apache Series Analysis for Neuroimaging Data. In: Proc. of Spark: A Unified Engine for Big Data Processing. the 8th Python in Science Conf., pp. 68-75 (2009) Communications of the ACM, 59, pp. 56-65 (2016) [18] Abraham, A., Pedregosa, F., Eickenberg, M., [5] Vavilapalli, V.K., Murthy, A.C., Douglas, C., Gervais, P., Mueller, A., Kossaifi, J., Gramfort, A., Agarwal, S., Konar, M., Evans, R., Graves, T., Thirion, B., Varoquaux, G.: Machine Learning for Lowe, J., Shah, H., Seth, S., others: Apache Hadoop Neuroimaging With Scikit-learn. Frontiers in yarn: Yet Another Resource Negotiator. In: Proc. of Neuroinformatics, 8 (2014) the 4th annual Symposium on Cloud Computing. p. 5. ACM (2013) [19] Sowa, J.F., others: Knowledge Representation: Logical, Philosophical, and Computational [6] Huth, A.G., Heer, W.A. de, Griffiths, T.L., Foundations. MIT Press (2000) Theunissen, F.E., Gallant, J.L.: Natural Speech Reveals the Semantic Maps that Tile Human [20] Poldrack, R.A.: Region of Interest Analysis for Cerebral Cortex. Nature, 532, pp. 453-458 (2016) fMRI. Social Cognitive and Affective Neuroscience, 2, pp. 67-70 (2007) [7] Friston, K.J.: Functional and Effective Connectivity: A Review. Brain connectivity, 1, [21] Van Den Heuvel, M.P., Pol, H.E.H.: Exploring the pp. 13-36 (2011) Brain Network: A Review on Resting-State fMRI Functional Connectivity. European Neuropsy- [8] Biswal, B.B., Kylen, J.V., Hyde, J.S.: Simultaneous chopharmacology, 20, pp. 519-534 (2010) Assessment of Flow and BOLD Signals in Resting- state Functional Connectivity Maps. NMR in [22] Sporns, O.: Discovering the Human Connectome. Biomedicine, 10, pp. 165-170 (1997) MIT Press (2012) 156