Abstract

Brain is the most complex organ in human body. Understanding how different regions of the brain function and interact with one another is a challenging task. One of the most important topics in the study of the brain is the functional brain connectivity, which is defined as the correlations, each between a pair of the activation signals from the different regions of the brain. Study of the functional connectivity in the human brain provides new insights into the understanding of the healthy and diseased brains and their differences. Functional magnetic resonance imaging (fMRI) is an imaging technique that allows researchers to study the brain activity and functional connectivity. While many researchers have focused on static functional connectivity in the resting-state fMRI to study the functions of the brain, dynamic functional connectivity has received more attention recently for such a study, since it provides more detailed information about the brain functions. Within the literature for studying the dynamic brain connectivity, k-means clustering has been applied to the connectivity matrices in order to find the functional connectivity patterns. However, it is known that the k-means clustering technique is not suitable for applying it to high dimensional data such as the functional brain connectivity matrices. In this thesis, in order to overcome this problem, we propose a deep learning-based convolutional autoencoder to obtain latent representations of the connectivity matrices prior to applying to them the k-means clustering. Use of the convolutional autoencoder, not only reduces the dimension of the connectivity matrices, but also provides a more semantic representation of these matrices. It is shown that the proposed method of clustering that consists of the use of the autoencoder followed by k-means clustering results in improving the clustering of the connectivity matrices, and consequently, to a better capturing of the functional connectivity patterns. In order to show the effectiveness of the proposed clustering method, synthetic connectivity matrices for patterns, with their classes known, are generated. The proposed method is then first applied to these syntactically generated connectivity matrices and the resulting patterns are compared with that obtained by applying k-means clustering technique to the synthetic connectivity matrices. It is shown that the proposed method classifies the various patterns more accurately.

The proposed method is then used to study the dynamic functional brain connectivity by applying it to real fMRI data captured from a group of healthy subjects and another group of subjects affected by schizophrenia. For this purpose, after preprocessing the raw fMRI data for each subject in these two groups, the group independent component analysis (ICA) is applied in order to decompose the fMRI data into statistically independent components (map of the entire brain) and their corresponding time-courses. Each independent component corresponds to a specific region of the brain. The connectivity matrix whose elements corresponds to the correlation between the time-courses within a segment of the time-courses enclosed inside a sliding window is then obtained. Next, the proposed clustering method is used to cluster all the connectivity matrices, each corresponding to one segment, into a finite number of functional connectivity patterns (states). A two-sample t-test is then performed on each state in order to determine each pair of the regions in the group of the healthy control subjects for which weather or not the correlation value is significantly different from that of the corresponding pair of the regions in the group of schizophrenia patients. It is observed through this test that there are indeed pairs of the brain regions where significant differences do exist between the two groups. It is also seen that such a difference between the two groups is even more pronounced in the visual network of the brain.

Finally, in this thesis, a study is undertaken for the evaluation of the dwell time, which is defined to be the duration for a functional connectivity pattern to remain in one state before switching to another state. It is shown through this study that the dwell time for the healthy group to stay in the state with more connectivity is longer than that for the group with schizophrenia. On the other hand, the dwell time for the group with schizophrenia to stay in the state with less connectivity is longer than that for the healthy group.