Augmentation does not create new patients

Small medical datasets invite aggressive augmentation. Rotation, flips, contrast jitter, elastic warps, and generative models can reduce overfitting on the training folder. They cannot invent a new hospital. A GAN that learns from twenty malignant cases will reproduce the texture of those twenty cases. The network then looks robust in-sample and fails when the probe, the stain, or the population changes.

I distinguish two jobs. Geometric and photometric transforms, applied with the same settings at training only, are regularisers. Generative augmentation is a modelling choice that must be validated: train with it, test on real held-out images that the generator never saw. If the gain disappears on real test images, the GAN was decorating the training loss, not adding information.

Channel-boosted and residual pipelines in my work use generative methods as part of a larger system, not as a substitute for data. That distinction belongs in teaching. Students should keep a log of every transform, a forbidden list (do not flip laterality marks, do not crop out the lesion), and a raw-file archive. Augmentation that changes the label is not augmentation. It is relabelling by accident.

When someone asks how to fix a tiny dataset, the honest answers are: more labelled cases, a better split, a simpler model, and transfer from a related modality with a clear statement of what was pretrained. A colourful fake image is not a new patient.