Enhancing ASD Diagnosis with Contrastive and Non-Contrastive Models from Neuroimaging Data

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition which impacts social interaction, communication, and behavior. This study introduces a self-supervised learning approach for ASD prediction using resting-state functional MRI (rs-fMRI) scans from the ABIDE I dataset, employing contrastive and non-contrastive models. The contrastive model enhances learning by maximizing agreement between augmented views of the same scan, while the non-contrastive model reconstructs input data through an autoencoder. Both models, pre-trained on ImageNet, were fine-tuned for ASD classification using Resnet18. Our results show that the contrastive model achieved 97.6% accuracy, with a precision of 95.7%, recall of 99.2%, and F1 score of 97.4%. The non-contrastive model reached 95.4% accuracy, with precision, recall, and F1 scores of 99.3%, 96.46%, and 97.82%. These findings highlight the performance of self-supervised learning in ASD diagnosis, showcasing robust pattern-capturing capabilities for neuroimaging data.