AI-Driven Melanoma Detection
A PyTorch deep-learning model for classifying dermoscopic melanoma images, built for early-detection use in healthcare.
Overview
Built and evaluated a PyTorch model to classify dermoscopic images of melanoma at the Australian Institute of Health Innovation. Engineered domain-specific transformations such as rotational invariance and noise reduction to improve generalisation, lifting accuracy by 12% on imbalanced data. Used StepLR dynamic learning-rate scheduling to cut convergence time by 25% while holding precision at 93%.
Ran rigorous evaluation across more than 50 epochs, reporting AUC-ROC and F1 for transparency against clinical-grade expectations, and processed over 15,000 high-resolution images on NCI GADI HPC at 88% GPU utilisation. Added Grad-CAM visualisations so clinicians could inspect the model’s decision regions.