Preiskovanje povezav med psihiatričnimi motnjami in nevroslikovnimi podatki z uporabo globokega učenja

Identifying the brain regions associated with certain cognitive functions and neurological disorders is an important focus of neuroimaging research. With the latest neuroimaging tools, researchers have been able to investigate the neural basis of various psychiatric disorders like never before. But the analytic capabilities of these tools still have to catch up to the magnitude of the intricate and detailed three-dimensional imaging data. With neural networks, and in particular, Graph Convolutional Networks, comes the ability to develop unique and creative ways to analyze the brain as a graph. This paper studies the utilization of Graph Convolutional Networks on neuroimaging datasets to determine the patterns linked to various psychiatric disorders. Graph Convolutional Networks permit the fusion of spatial and relational data by defining brain areas as nodes and using the functional connectivity matrix as the edges. We try to take advantage of structural information which is potentially hidden in functional connectivity matrices by applying Graph Convolutional Networks on the dataset consisting of 636 participants, including 150 diagnosed with bipolar disorder, 119 with schizoaffective disorder, 167 with schizophrenia, and 202 without diagnosis. We trained our models for three tasks: binary and multiclass classification of diagnosis and regression of indicators of psychopathology. The ANN baseline outperformed the GCN: it achieved strong accuracy in the binary task and produced better multiclass discrimination between diagnostic groups. The GCN underperformed relative to the ANN, suggesting that the chosen graph construction and GNN architecture did not capture discriminative signals as effectively. Both models performed poorly on PANSS regression, indicating symptom severity was not well predicted from the used functional connectivity representations.