Self-supervised ensembled learning for autism spectrum classification

Abstract

PurposeDeep learning has made remarkable progress in classifying autism spectrum disorder (ASD) using neuroimaging data. However, the current methods rely mainly on supervised learning, which requires a large amount of manually labeled data, making it an expensive and difficult task to scale. MethodsTo overcome this limitation, we propose a novel ensemble-based framework that learns a transferable and generalizable visual representation from different self-supervised features for the downstream task of ASD classification. This framework dynamically learns a superior representation by aggregating complementary information in the frequency domain from independent self-supervised features with limited data. Additionally, to address the information loss caused by the dimensionality reduction of 3D fMRI data, we propose a thresholding algorithm to optimally extract the most discriminant features from 2D rs-fMRI data. ResultsExperimental results demonstrate that the proposed method outperforms previous state-of-the-art methods by 19.69% on the ABIDE-1 dataset with a 10-fold cross-validation accuracy of 94.51%. ConclusionThe proposed method learns a transferrable and generalizable ensembled representation by leveraging complementary information encoded in different self-supervised representations for ASD classification.