Abstract
Deep learning has been proven effective in predicting brain activation patterns from resting-state features. In this work, using resting state and task fMRI data from the Human Connectome Project (HCP), we replicate the state-of-the-art deep learning model BrainSurfCNN and examine new model architectures for improvement. We also examine the role of individual variability in model performance. Specifically, first, we replicated the BrainSurfCNN model and assessed how varying the input feature space impacts task contrast prediction. Second, we explored two architectural changes for improving model performance and scalability: adding a Squeeze-and-Excitation attention mechanism (BrainSERF) and using a graph neural network-based architecture (BrainSurfGCN). Third, we examined how model performance is impacted by individual variability in task performance and data quality. Overall, we present replication, potential avenues for improvements in performance and scalability, and a better understanding of how individual variability impacts prediction performance - all in the hope of advancing deep learning applications in neuroimaging.