Open source · Featured · 2023
SkinGuard - Skin Cancer Classifier.
Five models, one answer.
presentsAdmit one
Act I
The challenge
Skin-lesion datasets are painfully imbalanced, and a naive model just learns to guess the common class. The challenge was dragging accuracy well past a weak baseline with transfer learning, then serving three different models to a browser fast enough to feel interactive.
Act II
What I built
A skin-lesion classifier: a Next.js frontend uploads an image through a Next edge proxy to a Flask service that normalizes it and runs a CNN over seven HAM10000 classes. Three real models ship, and transfer learning takes accuracy from a 68.8% baseline to 83.9%.
Intermission, the demo
The product, playing itself
three CNN models to choose from, an educational tool, not a diagnosis
The cast
What actually makes it work
Read the card, then flip it to watch the mechanism run.
0x01
Transfer learning
DenseNet201 truncated at its last relu with a GlobalMaxPooling + Dense(512) + Dropout head, lifting accuracy from 68.8% to 83.9% on the imbalanced HAM10000 set.
how it works
Transfer learning
0x02
Training / serving parity
The exact dataset mean and std and each model’s input geometry are replicated in the Flask handler, so what serves matches what trained.
how it works
Training / serving parity
0x03
Browser to model plumbing
Base64 images flow through a Next edge proxy (4MB limit) into Flask, PIL, and OpenCV, with a model-selector UI on top.
how it works