A Graph Neural Network based fMRI classification

In this paper, we aim to describe a novel deep learning model (a machine learning subclass) to classify fMRI. We used a publicly available dataset to train and test the model which involved patients affected by depression.Our model is based on a specific deep neural network, the Graph Attention Network (GAT) which has proven its strength in dealing with graph data representation. The novelty of our approach is that it is based on the extraction from the original fMRIs of graph representation then passed to the deep learning model.We performed this crucial phase by using a Matlab based toolbox, CONN, which helped in data preparation and graph representation extraction. We then used the extracted fMRI representations to feed, train, and finally test our deep learning model.While classification results were encouraging, achieving approximately 73% accuracy, another aspect that we investigated was focused on the comparison of three architectural solutions, focusing on power consumption. We used an Apple Silicon platform compared to a NVIDIA based laptop and an edge device of the NVIDIA Jetson family.