AI / ML Skin Cancer Detection System with XAI
A novel Classification Model achieving 86% accuracy, featuring an interpretable Grad-CAM visualization with fine-tuned MED-GEMMA LLM explanation.
OVERVIEW
An end-to-end skin cancer diagnosis pipeline combining a high-accuracy segmentation CNN with explainable AI. Grad-CAM++ produces visual heatmaps highlighting suspicious regions, while a fine-tuned MED-GEMMA VLM generates clinician-readable textual explanations — bridging black-box deep learning with actionable medical insight.
Key features
85.66% segmentation accuracy on HAM10000 and ISIC datasets
Grad-CAM++ saliency maps for pixel-level decision transparency
Fine-tuned MED-GEMMA LLM generating natural language diagnoses
Deployed as an interactive HuggingFace Space for real-time inference
Supports 7 skin lesion subtypes including melanoma and basal cell carcinoma
Engineering
- Integrated a CNN segmentation backbone with a vision-language model in a single inference pipeline
- Custom Grad-CAM++ implementation renders heatmaps at original image resolution
- Optimized model quantization for CPU-efficient Spaces deployment without GPU
Impact
- Provides dermatologists with explainable AI assistance for faster, more confident triage
- Published and cited as a contribution to interpretable medical AI research