Brain Tumor Classification

Streamlit app that classifies brain MRI scans into four categories — glioma, meningioma, no tumor, pituitary — using a custom CNN and Xception, with OpenAI-powered clinical reports and multi-turn image chat.

• Models — Xception (transfer learning, target ≥99%) + Custom SE-ResNet CNN (target ≥98%)
• App — Streamlit with saliency map overlays and probability charts
• AI — OpenAI GPT-4o for clinical report generation and MRI chat
• Model storage — Hugging Face Hub (downloaded on first run if not present locally)

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# edit .env — set OPENAI_API_KEY and HF_MODEL_REPO
streamlit run app/app.py

For HF Spaces deployments set both in Settings → Repository secrets.

python train.py                  # train both models
python train.py --model xception # one model only
python train.py --quick          # smoke-test (few epochs)
python train.py --strict         # exit 1 if accuracy targets are missed (CI)

Training notebooks with architecture notes are in notebooks/.

Model registry and class metadata are in config.toml. Add a [[models]] entry there to register a new model — no Python changes needed.

├── app/            Streamlit app
├── notebooks/      Training notebooks
├── utils/          Inference helpers
├── models/         .keras weight files (git-ignored)
├── config.toml     Model registry and class metadata
├── train.py        Training script
└── Dockerfile      HF Spaces deployment

.streamlit/config.toml keeps enableCORS and enableXsrfProtection set to true. Do not disable either in a shared deployment — see Streamlit security docs for details.

For research and educational purposes only. Not medical advice.

Model predictions can be wrong. GPT-generated reports are AI-produced and unreviewed by clinicians. Do not use this tool as a substitute for professional diagnosis.