machine learning19 Apr 2026· 7 min read
Deep learning on medical images, minus the hype
Working on Alzheimer's detection and skin lesion classification taught me that data handling matters far more than picking a fancier architecture.
The split is the experiment
Any leak between train and test — the same patient in both, or augmented copies straddling the split — produces a number that looks great and means nothing. Getting the split right is most of the scientific work.
Augmentation beats architecture, usually
Comparing CNN variants was useful, but the biggest honest gains came from sensible augmentation and class balancing, not from swapping the backbone.
Report what a clinician would ask
Accuracy alone hides the failure mode that matters. Per-class recall, confusion matrices, and calibrated confidence say far more about whether a model is usable.