A Practical Study of Data Requirements for Self-Supervised Learning in Medical Image Analysis

Deep learning has become the mainstream approach for medical image analysis, but the availability of annotated datasets remains constrained in many clinical scenarios. Even when using such limited data, however, self-supervised learning (SSL) has yielded promising results. Nevertheless, comprehensive investigations into the number of images required for effective pretraining and fine-tuning in SSL frameworks remain inadequate. Especially lacking are such studies assessing factors such as class diversity, dataset scale, and task complexity. For this study, we evaluated the performance of widely adopted contrastive SSL models, particularly assessing their applicability in limited data settings by examining how the number of images used for pretraining and fine-tuning influences the accurate execution of binary classification tasks. Our findings indicate that among the three studied methods, Simple Siamese Representation Learning (SimSiam) achieved high accuracy based on only four training images. Achieving such performance would likely require at least 10,000 pretraining images. These findings offer practical insights into optimizing SSL-based pipelines for medical image analysis, particularly in scenarios involving rare diseases or severely scarce data.