A Deep Learning Model for Data-Driven Discovery of Functional Connectivity

Functional connectivity (FC) studies have demonstrated the overarching value of studying brain and its disorders through undirected weighted graph functional magnetic resonance imaging (fMRI) correlation matrix. However, most work with FC depends on way is computed, it further manual post-hoc analysis matrices. In this work, we propose a deep learning architecture BrainGNN that learns structure as part to classify subjects. It simultaneously applies graphical neural network learned select sparse subset regions important prediction task. We demonstrate model’s state-of-the-art classification performance schizophrenia fMRI dataset how introspection leads disorder relevant findings. The graphs are by model exhibit strong class discrimination consistent literature.