Skin-Lesion ClassificationNeural Architecture Search, Medical Imaging
A professor asked me to look at a skin-cancer classification project that had stopped moving. The model wasn't broken — it was being optimised for the wrong number.
The problem
The pipeline was tuned for overall accuracy on 27,000+ ISIC dermoscopy images. On a heavily imbalanced dataset that is close to meaningless: a model can look excellent while missing most malignant lesions. Recall on the class that actually matters sat at 49%. In clinical screening, a false negative is the expensive error.
The approach
I rebuilt the pipeline around recall rather than accuracy. NASNetMobile transfer learning with Keras Tuner architecture search, focal loss to stop the majority class dominating the gradient, class-weighted training, and stratified augmentation so validation splits kept the real class distribution instead of flattering the metrics.
The outcome
Recall on malignant lesions went from a 49% baseline to 77.7%, at 0.86 AUC across 27,000+ images — a model that is now useful as a screening aid rather than a well-scored classifier that misses half the cases.