Semantic segmentation of breast ultrasound images using a custom U-Net architecture.
Trained on the BUSI dataset · Served via FastAPI · GDPR-compliant · Deployed with Docker · CI/CD with GitHub Actions.
| Metric | Score |
|---|---|
| Dice Coefficient | 0.73 |
| IoU (Jaccard) | 0.66 |
| Precision | 0.80 |
| Recall (Sensitivity) | 0.71 |
| Pixel Accuracy | 0.96 |
Evaluated on the BUSI test set (benign + malignant + normal classes).
Training tracked with W&B — view experiment dashboard.
Input (1×128×128 grayscale ultrasound)
↓
Encoder (4× conv blocks + MaxPool)
64 → 128 → 256 → 512 channels
↓
Bottleneck (1024 channels)
↓
Decoder (4× ConvTranspose2d + skip connections)
512 → 256 → 128 → 64 channels
↓
Output (1×128×128 binary segmentation mask)
Custom U-Net with:
- BatchNorm after every convolution
- Skip connections between encoder and decoder
- BCEWithLogitsLoss for training stability
- Adam optimizer with lr=5e-4
This project integrates medical-ai-middleware — a production-grade middleware package built for medical AI APIs:
| Feature | Details |
|---|---|
| GDPR Compliance | Consent enforcement, right to erasure, audit logging |
| Data Anonymization | EXIF stripping, IP anonymization before S3 upload |
| S3 Storage | Anonymized uploads auto-deleted after 24h, audit logs after 90 days |
| Prometheus Metrics | Request count, latency, error rate at /metrics |
| Rate Limiting | 10 req/min on /predict, 30 req/min default |
| Security Headers | HSTS, CSP, X-Frame-Options, X-Request-ID on every response |
Breast-Cancer-Segmentation/
├── .github/
│ └── workflows/
│ └── ci.yml ← GitHub Actions: lint → test → docker → deploy
├── Data/
│ ├── Dataset_BUSI_with_GT/ ← BUSI dataset (not committed, tracked by DVC)
│ └── data_loader.py
├── Docker/
│ └── Dockerfile ← Production container
├── src/
│ ├── model.py ← U-Net architecture
│ ├── train.py ← Training loop with W&B + MLflow
│ ├── evaluate.py ← Evaluation + quality gates
│ ├── validate.py ← Validation loop
│ ├── metrics.py ← Dice, IoU, precision, recall, F1
│ ├── config.py ← Hyperparameters and paths
│ ├── mlflow_utils.py ← Learning curves, threshold sweep
│ ├── utils.py ← Visualization helpers
│ ├── inference.py ← MLflow model registry inference
│ └── app.py ← FastAPI serving + middleware
├── hf_spaces_app.py ← Gradio demo (HuggingFace Spaces)
├── tests/
│ ├── test_suite.py ← Model, metrics, preprocessing tests
│ └── test_api.py ← API + middleware integration tests
├── checkpoints/ ← Model weights (not committed)
├── requirements.txt
└── README.md
git clone https://github.com/moebouassida/Breast-Cancer-Segmentation.git
cd Breast-Cancer-Segmentation
docker build -f Docker/Dockerfile -t breast-seg-api .
docker run --rm -p 8000:8000 \
-e S3_ENABLED=false \
breast-seg-apiOpen:
- API docs: http://localhost:8000/docs
- Health check: http://localhost:8000/health
- Prometheus metrics: http://localhost:8000/metrics
git clone https://github.com/moebouassida/Breast-Cancer-Segmentation.git
cd Breast-Cancer-Segmentation
python -m venv venv && venv\Scripts\Activate.ps1 # Windows
pip install -r requirements.txt
uvicorn src.app:app --host 0.0.0.0 --port 8000All endpoints on /predict and /explain/predict require the consent header:
X-Data-Consent: true
{
"status": "healthy",
"metrics": true
}curl -X POST http://localhost:8000/predict \
-H "X-Data-Consent: true" \
-F "file=@ultrasound.png"{
"device_used": "cpu",
"threshold": 0.5,
"request_id": "550e8400-e29b-41d4-a716-446655440000",
"coverage_pct": 12.4,
"mask_png_b64": "iVBORw0KGgo...",
"overlay_png_b64": "iVBORw0KGgo..."
}Returns segmentation mask + Grad-CAM heatmap showing which regions drove the prediction.
{
"mask_png_b64": "...",
"overlay_png_b64": "...",
"coverage_pct": 12.4,
"xai": {
"heatmap_b64": "...",
"method": "gradcam",
"clinical_note": "Red regions indicate areas that most strongly influenced the segmentation decision."
},
"request_id": "550e8400-..."
}Returns GDPR compliance status and data retention policy.
Permanently deletes all data associated with a request ID.
{ "erased": true, "s3_deleted": true }Prometheus metrics — request counts, latency histograms, error rates.
# All tests
pytest tests/ -v
# Model + metrics only
pytest tests/test_suite.py -v
# API + middleware workflow
pytest tests/test_api.py -vTest coverage:
- U-Net output shapes, gradient flow, no-NaN outputs
- Dice/IoU/precision/recall on known inputs
- GDPR consent enforcement (403 without header)
- Security headers on every response
- Full workflow: upload → predict → explain → erase
BUSI (Breast Ultrasound Images Dataset)
780 images across 3 classes: benign, malignant, normal. Each image paired with a ground truth segmentation mask.
Data/
└── Dataset_BUSI_with_GT/
├── benign/
├── malignant/
└── normal/
python src/train.pypython src/evaluate.py --checkpoint checkpoints/best.pt Metric Value Threshold Status
----------------------------------------------------
dice 0.7300 0.7000 ✅ PASS
iou 0.6600 0.5500 ✅ PASS
precision 0.8000 0.6500 ✅ PASS
recall 0.7100 0.6500 ✅ PASS
Push to main/dev
↓
Code Quality (ruff + black)
↓
Unit + Integration Tests (pytest)
↓
Docker Build + Smoke Test (/health)
↓
Deploy to HuggingFace Spaces 🤗
| Component | Tool |
|---|---|
| Deep Learning | PyTorch |
| API Serving | FastAPI + Uvicorn |
| Production Middleware | medical-ai-middleware |
| Monitoring | Prometheus |
| GDPR & Compliance | GDPR middleware + AWS S3 |
| XAI | Grad-CAM |
| Experiment Tracking | W&B + MLflow |
| Data Versioning | DVC |
| Demo | Gradio + HuggingFace Spaces |
| Containerization | Docker |
| CI/CD | GitHub Actions |
| Dataset | BUSI (780 breast ultrasound images) |
MIT License — free to use and modify.
Moez Bouassida — AI/ML Engineer · Medical Imaging
