Offline-first, explainable AI for thoracic disease screening from chest X-rays (TB, pneumonia, COVID-19, no findings).

• Streamlit UI — Upload PNG/JPEG/WebP/TIFF/BMP or DICOM (.dcm), view predictions, confidence, Grad-CAM++ overlays, PDF reports, SQLite admin.
• FastAPI — POST /predict (PyTorch + heatmap), POST /predict/onnx (ONNX, fast CPU edge path without saliency).
• Training — scripts/train_finetune.py for folder-based fine-tuning; scripts/evaluate_model.py for validation metrics.
• Deployment — ONNX export, optional INT8 quantization, Docker, env overrides (XRAYNET_WEIGHTS, XRAYNET_ONNX_PATH).

pip install -r requirements.txt
python scripts/init_demo_model.py   # optional: save starter .pth
streamlit run app/main.py

SSL errors on macOS/conda: the app runs src/utils/ssl_setup.py before downloading ImageNet weights (tries certifi, then falls back to unverified TLS for local dev only so PyTorch model URLs work).

API:

uvicorn api.main:app --host 0.0.0.0 --port 8000

See data/xray_finetune/DATA_LAYOUT.txt. Evaluate on a labeled val/ tree:

python scripts/evaluate_model.py --data-dir data/xray_finetune --split val

Download instructions: data/nih_chest_xray/KAGGLE_SETUP.txt.

python scripts/train_nih_xraynet.py --nih-root data/nih_chest_xray --epochs 5 --max-per-class 2000 --class-weights

NIH does not contain true TB or COVID labels; the script maps findings to the 4 app classes using documented proxies (see src/data/nih_xraynet_dataset.py).

Unified fine-tune (NIH + TB + COVID Kaggle): see data/UNIFIED_FINETUNE_SETUP.txt.

python scripts/train_unified_finetune.py --nih-root data/nih_chest_xray --tb-root data/kaggle_tb_chest --covid-root data/kaggle_covid_chest --dry-run
python scripts/train_unified_finetune.py --nih-root data/nih_chest_xray --tb-root data/kaggle_tb_chest --covid-root data/kaggle_covid_chest --epochs 10 --class-weights --cap-per-class 8000

For research and decision-support demos only — not a certified medical device. Clinical use requires validation, regulatory clearance, and qualified oversight.