Skin Lesion Classification with GAN Augmentation

Overview

A full-stack application classifying skin lesions (melanoma, benign nevi, basal cell carcinoma, actinic keratosis, benign keratosis). It achieves 97.23% accuracy and a 95.39% macro-F1 on the HAM10000 dataset.

Problem

Class imbalance and hair/artifact noise in medical-image datasets degraded accuracy on minority classes.

Solution

ACGAN (class-conditional) and DCGAN generate synthetic data for imbalance, while a black-hat transform + inpainting removes hair artifacts. An enhanced SE-ResNet adds SE blocks across all ResNet layers with residual connections after each block.

Highlights

• Accuracy: base ResNet50 67% → 91% with DCGAN → 97.23% with SE-ResNet + ACGAN.
• ImageNet-pretrained ResNet-50 base, SE attention modules, 25.2M parameters.
• FastAPI + PyTorch backend; inference under 3 seconds.
• Visualization frontend with React, Vite, Tailwind and Chart.js.

Screenshots