How conn-finnegan/soma_skin_cancer works
Project type and purpose
• This is a privacy-focused AI/ML pipeline tool for classifying dermoscopic skin-lesion images as benign or malignant.
• It provides:
• Model training (PyTorch + ResNet-18)
• Model evaluation
• Inference (single / batch image prediction + Grad-CAM overlays)
• A CLI-style usage pattern (no images are stored or logged).
File structure
Root files:
• README.md – project overview and usage
• requirements.txt – Python dependencies (PyTorch, torchvision, Grad-CAM libs)
• Lumen_Skin_Cancer.code-workspace – VS Code workspace config
models/
• skin_cancer_resnet18_v1.mlmodel (to be deprecated – drop this once iOS feature is removed)
src/
• load_data.py – data ingestion/preprocessing
• dataset.py – dataset wrapper for HAM10000
• train_model.py – training script (fine-tunes ResNet-18)
• evaluate.py – runs evaluation on a held-out set
• predict_image.py – single-image inference + Grad-CAM
• test_loader.py – sanity checks on data loaders
README for additional insights
• Emphasises “all inference runs in-memory; no images stored.”
• Deployment target: Hugging Face Spaces (Gradio) – not part of this file tree, but consider as an external demo UI.
[ ] Lay out the seven components listed in 4a
[ ] Encapsulate any external demo or integration as a separate module