This paper investigates self-supervised learning approaches for medical image diagnosis within the AD-CARE framework, focusing on Alzheimer's Disease. It evaluates SimCLR, MAE, and DINO on standard benchmarks and synthetic medical data, including a multi-cohort assessment and fairness analysis. The study finds SimCLR competitive with supervised methods and highlights demographic performance disparities.
Key findings
SimCLR achieves near-supervised performance after fine-tuning on CIFAR-10.
Self-supervised methods show better cross-domain transferability.
Fairness analysis reveals significant performance disparities across demographic groups.
Limitations & open questions
Future work needed to scale to real ADNI datasets.
Development of fairness-aware objectives is required for clinical deployment.