Medical-imaging datasets are often smaller and more expensive to label than general image collections. Transfer learning helps by starting with a model that has already learned useful visual representations and adapting it to a new clinical task.
Why pretrained features help
Early neural-network layers detect patterns such as edges, shapes, and textures. Many of these features remain useful across domains. A model pretrained on a large dataset can therefore provide a stronger starting point than random initialization, especially when labeled medical examples are limited.
A common training workflow
First, replace the original classification head with a new output layer for the target task. Freeze most of the backbone and train the new head. Then unfreeze selected deeper layers and fine-tune them using a smaller learning rate. Gradual unfreezing reduces the risk of destroying useful pretrained weights too quickly.
Domain shift still matters
Natural photographs and radiological images have different visual characteristics. Even two hospitals may use different scanners, acquisition protocols, or patient populations. Transfer learning improves initialization, but it does not remove the need for representative data and external validation.
Avoid data leakage
Images from the same patient must not appear in both training and validation sets. Otherwise, the model may recognize patient-specific patterns and produce overly optimistic results. Splits should be created at the patient level before augmentation or preprocessing.
Evaluate beyond one score
AUROC alone may hide poor calibration, weak minority-class performance, or sensitivity to acquisition changes. Report several metrics, confidence intervals, error analysis, and subgroup results. Transfer learning is most valuable when it is combined with careful fine-tuning, leakage prevention, and evaluation that reflects the intended clinical setting.