skinDisease — Dermatology Classification

Skin conditions are among the most common reasons people see a doctor, yet dermatologists cluster in big cities. A family in a rural district can wait weeks and travel hours for an answer to something a photo could triage. I wanted to see how far an honest, explainable model could go toward closing that gap — not to replace a doctor, but to help decide who needs one soon.

I built it alone. Every part — how the data is split, how the model is trained, how results are measured and how they are explained — was a decision I made and can defend, and I wrote it as a reproducible codebase with tests rather than a single notebook that only runs on my machine.

Medical datasets are brutally imbalanced: a few conditions have thousands of images, others only a handful. A model that ignores the rare classes can still look accurate on paper while being useless for the patients who need it most. I fought that with a weighted sampler so rare conditions appear every epoch, a class-weighted loss for a stronger gradient on them, and skin-aware augmentation (including elastic deformation) to squeeze more signal from limited images.

I chose EfficientNet-B0 for its accuracy-to-size ratio and fine-tuned it in two phases — freeze the pretrained backbone to warm up a new head, then unfreeze to adapt gently — with a warmup-plus-cosine learning-rate schedule and mixed-precision training for roughly a 2× speedup. To be sure the choice was principled, I benchmarked it against ResNet-50 and ConvNeXt rather than trusting one run.

A label alone is dangerous in medicine. Grad-CAM overlays where the model was looking, so a clinician can tell whether it focused on the lesion or on a shadow — and reject it when it's wrong. I am deliberate about the framing too: this is a screening aid to prioritise, never a diagnosis, and I keep the words 'aid' and 'diagnose' apart on purpose.

On this dataset EfficientNet-B0 reaches an expected top-1 around 75% and top-5 around 93%, edging out ResNet-50 — but I report these as expected ranges, not verified claims, because they shift with hardware and random seed and I have not run an independent clinical evaluation. Saying '~75%, pending verification' is less impressive than a bold number, and far more trustworthy — which, for a medical tool, is the only thing that counts.

The goal was never a leaderboard score; it was triage for people far from a specialist. Next, I want a proper held-out clinical evaluation with a doctor in the loop, probability calibration so the model knows when it's unsure, and a lightweight on-device version that works without a connection.