STL Net: A spatio-temporal multi-task learning network for Autism spectrum disorder identification

Abstract

Background and Objective:

The rich temporal and spatial information contained in Functional magnetic resonance imaging (fMRI) data is crucial for accurately identifying Autism spectrum disorder (ASD). Most current ASD identification methods capture temporal and spatial information in a serial manner, resulting in partial loss of information and sub-optimal outcomes. To solve this problem, we propose a heterogeneous spatio-temporal multi-task learning network (STL Net) for distinguishing between ASD patients and normal controls (NCs).

Methods:

Initially, we define two networks to extract temporal and spatial features respectively. Subsequently, the attention mechanism further capture useful features related to ASD in each network. To facilitate the interaction of spatio-temporal information, a spatio-temporal feature sharing module shares temporal and spatial features in parallel. Finally, the spatio-temporal features are aggregated for ASD identification.

Results:

We conduct experiments on five datasets from the Autism Brain Imaging Data Exchange, with the following results: Accuracy of 73.52%, 72.00%, 83.33%, 78.57% and 90.90%; Sensitivity of 66.66%, 70.00%, 80.00%, 88.88%, and 100.00%; and Specificity of 78.94%, 73.33%, 87.50%, 60.00% and 80.00%. The results show that our method outperforms other state-of-the-art ASD identification methods in Accuracy and exhibits significant competitiveness in Sensitivity and Specificity. Additionally, this method accurately identifies and points out the associated brain regions in ASD patients.

Conclusions:

This paper proposes a novel heterogeneous multi-task learning method, which offers a new perspective for more effective utilization of fMRI data in ASD identification. The proposed method can be translated into clinical applications to assist doctors in automated health screening for ASD.