Hybrid SSL-Driven ASD Detection: A Study of DINOv2, MoCo, BYOL and SimCLR with CNN Integration

Abstract

In this work, we investigate modern iterations of self-supervised learning (SSL) and contrastive learning approaches for classifying Autism Spectrum Disorder (ASD) from neuroimaging data. This study employs these techniques by extensively using strong pre-trained models like DINOv2 and MoCo, SimCLR and BYOL, and also effectively leveraging backbones like EfficientNetB0 and ResNet50 for ASD classification through novel hybrid approaches. Utilizing those sophisticated models with simple feed-forward classifiers/neural networks have yields high classification accuracies, ranged from 73.18% to 98.01%, depending on the framework and dataset used. A cross-breed approach with DINOv2, MoCo and SimCLR accomplishes a classification exactness of 91.06% showing that the utilize of a combination of vision transformers and contrastive learning systems for preparing in therapeutic imaging from prior errands complements each other. The second approach utilizes DINOv2 and SimCLR with an EfficientNetB0 backbone and achieves an impressive accuracy of 98.01%, signifying the merit of using always SSL algorithms alongside using feature projection for clinical decision support systems. Moreover, the best performance of 92.72% top-1 accuracy can also be achieved by the combination of DINOv2 and MoCo with ResNet50 backbone, which also illustrates the effectiveness of self-supervised learning and transfer learning for generating powerful features. For example, with MRI data, this study shows that DINOv2, BYOL, and MoCo extraction features outperformed each other achieving above 73% in classification accuracy, the metrics of DINOv2 were reported as... It showcases the strength of SSL models in effectively leveraging non-label data to generate features and demonstrates the wide applicability across heterogeneous dataset. Such detailed analysis, across these techniques, helps cement the high precision, recall and robustness provided in detecting ASD by these methods, as a significant insight into the usefulness of SSL across several diverse medical imaging tasks. Our work serves as a benchmark for future studies, and also highlights the changing landscape of neuroimaging-based ASD detection with the introduction of SSL model.