Spatiotemporal Attention Autoencoder (STAAE) for ADHD Classification

Abstract

It has been of great interest in the neuroimaging community to model spatiotemporal brain function and disorders based on resting state functional magnetic resonance imaging (rfMRI). A variety of spatiotemporal methods have been proposed for rfMRI so far, including deep learning models such as convolution networks (CNN) and recurrent networks (RNN). However, the dominant models fail to capture the long-distance dependency (LDD) due to their sequential nature, which becomes critical at longer sequence lengths due to memory limit. Inspired by human brain’s extraordinary ability of long-term memory and attention, the attention mechanism is designed for machine translation to draw global dependencies and achieved state-of-the-art. In this paper, we propose a spatiotemporal attention autoencoder (STAAE) to discover global features that address LDDs in rfMRI. STAAE encodes the information throughout the rfMRI sequence and reveals resting state networks (RSNs) that characterize spatial and temporal properties of the data. Considering that the rfMRI is measured without external tasks, an unsupervised classification framework is developed based on the connectome generated with STAAE. This framework has been evaluated on 281 children with ADHD and 266 normal control children from 4 sites of ADHD200 datasets. The proposed STAAE reveals the global functional interaction in the brain and achieves a state-of-the-art classification accuracy from 59.5% to 77.2% on multiple sites. It is evident that the proposed attention-based model provides a novel approach towards better understanding of human brain.