Hypothesis
Self-supervised learning, which cuts reliance on labels for large balanced datasets like ImageNet, would also beat ordinary supervised learning on the small, class-imbalanced datasets typical of medical imaging.
What failed
Across four binary tasks (Alzheimer's diagnosis and age prediction from brain MRI, pneumonia from chest X-rays, choroidal neovascularization from OCT; mean training sets of 771 to 33,484 images), supervised learning beat the chosen self-supervised paradigms in most small-training-set experiments, even when only part of the data was labelled.
Broken assumption
Self-supervised pretraining gains seen on large, balanced natural-image datasets carry over to small, imbalanced medical datasets.
Why
The abstract does not give a mechanism; the authors conclude that the choice of paradigm depends on training set size, label availability and class balance.
From the published paper: Espis et al. Scientific Reports, 2025. ↗CC BY