Cortical surface-based methods for analyzing the human brain morphology: Graph theoretical analysis and geometric deep learning

Abstract

Recently, graph theory has attracted many interests in various area of science. Graph can represent many complex systems such as social networks, protein-protein interaction networks, knowledge graphs, citation, Internet, neuroscience, etc. Graphs are the data structure in the non-Euclidean space that are consisted of vertices and edges. Vertices is the fundamental unit of the graph. And edges are association between two nodes. Brain surface model is considered as graph, because it is forming high resolution meshes and association between vertices was defined through discrete triangular elements. And it is represented brain network as graph which consist of parcellated brain regions and associations between these regions. The aim of the dissertation is proposing the novel procedures for analyzing graph defined from brain surface model. First, I proposed a method that construct network from brain surface model using morphological properties. Based on the Jensen Shannon divergence, a similarity between morphological distributions of two brain regions was defined. And this definition was expanded to define regional distribution from 1D to 2D, 3D. This allowed constructing brain networks that maintains properties for each morphological feature. To evaluate a similarity between brain regions, I compared the Jensen Shannon divergence with various measurements based on distance and divergence. Second, I investigated with respect to gender differences and hemispheric differences using brain morphological network. To evaluate constructed networks of young healthy controls using proposed framework, I investigated gender and hemispheric effects on morphological brain network comparing with the previous studies. In other word, the hemispheric difference within and between the gender groups was investigated in the global and local scale. These findings can provide indirect evidence of the topological differences in hemispheric morphology networks and the behavioral differences associated with gender. Third, I proposed deep learning model, surface model-based graph convolutional neural network architecture. Brain surface model forming high resolution meshes was defined as graph. And association between vertices was defined through discrete triangular elements. Various morphological features such as cortical thickness, surface area, cortical volume, sulcal depth, mean curvature were used to input of graph. I evaluated the performance of the graph convolutional neural network in gender classification using cortical thickness of both hemispheres. The overall results of dissertation showed that brain morphology analysis based on graph can be a useful approach for exploring new aspects of scientific investigations and clinical applications.