Autism Detection using r-fMRI: Subspace Approximation and CNN Based Approach

Abstract

Autism spectrum disorder (ASD) is a heterogeneous disorder that causes impaired social interactions and altered behavioral patterns. It is currently diagnosed by assessing the developmental screenings, followed by a diagnostic evaluation of an individual using standard screening and diagnostic tools. However, a more concrete diagnostic method is required to get hold of the underlying cause of the disorder to ensure better treatment and prevention of the disorder. Recent research on ASD has shown resting state functional magnetic resonance imaging (r-fMRI) as a useful tool for classification of ASD and neurotypical subjects. However, due to the large dimensionality of fMRI scans, directly applying classification methods to the data results in high computational cost. In order to avoid the curse of dimensionality, we propose a combined multistage SAAK transform and CNN based approach which selects the most relevant and discriminative features without losing any information and uses them for a more accurate and less computationally expensive classification of ASD subjects from typically developing controls. A classification accuracy of 74.55% is achieved using the proposed method. We show that the performance of the proposed approach of classification using ABIDE dataset is comparable to that of standalone CNN while being less computationally intensive.