talhasa3285/mammogram-screening-aid

Mammogram Screening Aid (Research Demo)

Binary mammogram classifier — Malignant vs Benign/Normal — with a FastAPI backend and a React (Vite) frontend, served as a single Docker container.

⚠️ Research & screening-aid demonstration only. Not a medical device and not for clinical diagnosis. Predictions may be wrong.

What it does

Upload a mammogram (PNG, JPEG, or DICOM) and get a classification, a confidence score, and a highlighted most-suspicious region. Uploads are processed in memory and never stored; DICOM patient tags are stripped on ingest.

Model

ResNet18 (ImageNet) tile features + an attention-MIL head, trained on the public mini-MIAS dataset (322 images). Honest performance: whole-image AUC ≈ 0.56 (near-chance) — mini-MIAS is tiny, so this is a proof-of-concept. See the technical report for the full methodology, metrics, and privacy framework.

Endpoints

• —/ — web UI
• —/docs — interactive API documentation (Swagger)
• —/health — health check
• —/predict — POST an image, returns the classification payload

(Full documentation and the technical report are added in Phase 5.)