Most skin cancer datasets are collected in dermatology clinics, using dermatoscopes, full-body photography, and complete patient histories. Most people don't have any of that. They have a phone and an unusual mole.

I built a single-photo skin lesion screening model using the ISIC SLICE-3D dataset and scoped it down to something an actual person could use: one photo in, one prediction out. No account, no lesion history, no clinical context.

I'll talk through the real trade-offs: why patient-level data splitting matters, why plain accuracy is the wrong metric on a dataset that's mostly benign, etc.

I'll close with a demo, a look at what the model is actually paying attention to, and an honest account of what I built.