Learning Dynamic Connectivity with Residual-Attention Network for Autism Classification in 4D fMRI Brain Images

Abstract

Diagnosing autism spectrum disorder (ASD) is still challenging because of its complex disorder and insufficient evidence to diagnose. A recent research in psychiatry perspective demonstrates that there are no obvious reasons for ASD. However, considering a hypothesis that abnormalities in the superior temporal sulcus (STS) connected with visual cortex regions can be a critical sign of ASD, a model is required to exploit the brain functional connectivity between STS and visual cortex to reinforce the neurobiological evidence. This paper proposes a deep learning model composed of attention and convolutional recurrent neural networks that can select and extract the time-series pattern of dynamic connectivity between the two regions within the brain based on observations. By integration of extracting autism disorder features from dynamic connectivity through attention with the structure containing interlayer connections to preserve the functional connectivity loss within a neural network, the model extracts the connectivity between STS and visual cortex, leading to the increase of generalization performance. Experiments with 800 patients’ fMRI imaging data known as ABIDE (Autism Brain Imaging Data Exchange) and 10-fold cross-validation to compare its performance show that the proposed model outperforms the state-of-the-art performance by achieving a 4.90% improvement in the ASD classification. Additionally, the proposed method is analyzed by visualizing dynamic brain connectivity of the neural network layers. © 2021, Springer Nature Switzerland AG.