Self-Supervised Learning for Domain Adaptation in Medical Imaging

Abstract: Self-supervised learning (SSL) offers a transformative path for addressing domain adaptation in medical imaging, where annotated datasets are often limited and expensive to acquire. This paper explores how various SSL approaches—contrastive learning (SimCLR), masked image modeling (MAE), and transformer-based learning (DINO)—improve performance in segmentation and classification across heterogeneous medical imaging domains (MRI, X-ray, CT). Using datasets such as BraTS, CheXpert, and NIH ChestXray14, we evaluate pretraining followed by fine-tuning with minimal supervision. We demonstrate statistically significant improvements (6–15%) in Dice scores and AUC. Regression analysis shows a strong correlation between SSL representation similarity (CKA) and downstream task performance. Explainability tools such as SHAP and LIME are used to validate model reliability and transparency. Keywords: Self-Supervised Learning, Domain Adaptation, Medical Imaging, Contrastive Learning, SimCLR, DINO, Swin UNet, SHAP, LIME, Transfer Learning