Multi-level and joint attention networks on brain functional connectivity for cross-cognitive prediction.

Abstract

Deep learning on resting-state functional MRI (rs-fMRI) has shown great success in predicting a single cognition or mental disease. Nevertheless, cognitive functions or mental diseases may share neural mechanisms that can benefit their prediction/classification. We propose a multi-level and joint attention (ML-Joint-Att) network to learn high-order representations of brain functional connectivities that are specific and shared across multiple tasks. We design the ML-Joint-Att network with edge and node convolutional operators, an adaptive inception module, and three attention modules, including network-wise, region-wise, and region-wise joint attention modules. The adaptive inception learns brain functional connectivity at multiple spatial scales. The network-wise and region-wise attention modules take the multi-scale functional connectivities as input and learn features at the network and regional levels for individual tasks. Moreover, the joint attention module is designed as region-wise joint attention to learn shared brain features that contribute to and compensate for the prediction of multiple tasks. We employed the Adolescent Brain Cognitive Development (ABCD) dataset (n =9092) to evaluate the ML-Joint-Att network for the prediction of cognitive flexibility and inhibition. Our experiments demonstrated the usefulness of the three attention modules and identified brain functional connectivities and regions specific and common between cognitive flexibility and inhibition. In particular, the joint attention module can significantly improve the prediction of both cognitive functions. Moreover, leave-one-site cross-validation showed that the ML-Joint-Att network is robust to independent samples obtained from different sites of the ABCD study. Our network outperformed existing machine learning techniques, including Brain Bias Set (BBS), spatio-temporal graph convolution network (ST-GCN), and BrainNetCNN. We demonstrated the generalization of our method to other applications, such as the prediction of fluid intelligence and crystallized intelligence, which also outperformed the ST-GCN and BrainNetCNN.