MammoAI — Multi-Modal Breast Cancer Detection

A three-stage interpretable AI system for breast cancer detection, fusing clinical BI-RADS features with deep learning on CBIS-DDSM mammography.

---

Results

Stage	Method	AUC-ROC	Sensitivity	Specificity
Stage 1	Gradient Boosting — 11 clinical features	0.8678	72.1%	83.2%
Stage 2	EfficientNet-B4 + TTA ×5	0.8294	87.3%	63.6%
Stage 3	Late-Fusion (512-dim CNN + clinical)	0.8825	78.6%	82.0%
Stage 1+	Multi-Dataset GB (CBIS-DDSM + VinDr-Mammo, 18,864 cases)	0.9925	—	—
Stage 5	Calibrated Ensemble (ECE = 0.052, Brier = 0.146)	0.8722	—	—
Sub-class	BENIGN risk stratification	0.9729	97.2%	78.9%

---

Four Novel Contributions

1. Dual Explainability — SHAP (clinical stage) + GradCAM (image stage) shown simultaneously
2. Late-Fusion Architecture — 512-dim CNN embeddings + 11 clinical features → AUC 0.8825
3. BENIGN Sub-Class Stratification — predicts biopsy need from benign cases (AUC 0.9729)
4. Cross-Dataset Generalisation — zero-shot transfer protocol to VinDr-Mammo

---

Architecture

CBIS-DDSM CSV ──► Stage 1: Gradient Boosting ──► AUC 0.8678 + SHAP explanations
                      │
CBIS-DDSM DICOM ──► Stage 2: EfficientNet-B4 ──► AUC 0.8294 + GradCAM heatmaps
                      │
                 Stage 3: Late Fusion ──────────► AUC 0.8825 (best overall)
                   [11 clinical + 512 CNN + 1 GB_prob = 524-dim → GB meta-learner]

---

Project Structure

MammoAI/
├── app.py                        Streamlit web application
├── requirements.txt
│
├── src/
│   ├── cancer_detection.py       Stage 1: full ML pipeline + plots
│   ├── save_model.py             Train & serialize GB model
│   ├── shap_analysis.py          SHAP explainability (Contribution 1)
│   ├── benign_subclass.py        BENIGN sub-class analysis (Contribution 3)
│   ├── cross_dataset.py          VinDr-Mammo cross-dataset study (Contribution 4)
│   ├── train_multi_dataset.py    Stage 1+ multi-dataset training (CBIS + VinDr)
│   └── calibration_ensemble.py   Stage 5 calibrated ensemble with uncertainty
│
├── stage2_cnn/
│   ├── config.py                 All hyperparameters
│   ├── dataset.py                DICOM loader + CLAHE preprocessing
│   ├── model.py                  EfficientNet-B4 architecture
│   ├── train.py                  Two-phase training + resume
│   ├── evaluate.py               TTA evaluation + GradCAM
│   ├── predict.py                Single-image inference
│   └── fusion.py                 Late-fusion pipeline (Contribution 2)
│
├── models/
│   ├── gb_model.pkl              Stage 1 Gradient Boosting (CBIS-DDSM)
│   ├── gb_multi_dataset.pkl      Stage 1+ GB (CBIS-DDSM + VinDr-Mammo, AUC 0.9925)
│   ├── gb_benign_subclass.pkl    BENIGN sub-class stratifier (AUC 0.9729)
│   ├── mammo_ensemble.pkl        Stage 5 calibrated ensemble (ECE 0.052)
│   └── ensemble_metrics.json     Calibration + disagreement metrics
│
├── results/
│   ├── shap/                     SHAP plots (5 figures)
│   ├── benign_subclass/          Sub-class analysis plots
│   ├── cross_dataset/            Domain shift analysis
│   └── stage2_cnn/results/       CNN evaluation + GradCAM + fusion plots
│
├── paper/
│   ├── mammoai_paper.md          Full research paper (14 references)
│   ├── mammoai_submission.md     Submission-formatted manuscript
│   ├── MammoAI_Manuscript.docx   Word export for journal submission
│   ├── Cover_Letter.docx         Journal cover letter
│   ├── Conflict_of_Interest.docx CoI statement
│   └── figures/                  Publication-ready plots (PNG)
│
└── data/csv/                     CBIS-DDSM CSV annotations (lightweight)
    NOTE: Raw DICOM images (~152 GB) are NOT committed — download from TCIA

---

Try it in 30 seconds

No dataset download needed — 7 ready-made demo mammograms ship with the repo.

git clone https://github.com/aliaht99/MammoAI.git
cd MammoAI
pip install -r requirements.txt
./run.sh                     # opens the MammoDoctor app at http://localhost:8501

Then upload any image from demo_images/ (e.g. 1_CLEAR_MALIGNANT_spiculated_mass.png) and read off the prediction. demo_images/DEMO_CHEATSHEET.txt lists the expected result and clinical-form values for each of the 7 scenarios.

Note: the pre-trained clinical + ensemble models are bundled, so Stage 1, Sub-class and the calibrated Ensemble work out of the box. The 74 MB CNN checkpoint is not committed — the app runs fine without it and simply marks Stage 2 / GradCAM as unavailable until you train it (step 7 below).

Two apps are included:

Command	App	Best for
./run.sh	mammo_doctor.py	Full clinician-facing platform (5 tabs, uncertainty, GradCAM, PDF report)
./run.sh app	app.py	Lightweight 3-tab demo

---

Full Pipeline (reproduce from raw data)

# 1. Install dependencies
pip install -r requirements.txt

# 2. Download CBIS-DDSM from TCIA (free account required)
#    https://doi.org/10.7937/K9/TCIA.2016.7O02S9CY

# 3. Run Stage 1 (no images needed — CSV only)
python src/cancer_detection.py

# 4. Run SHAP explainability
python src/shap_analysis.py

# 5. Run BENIGN sub-class analysis
python src/benign_subclass.py

# 6. Train Stage 2 CNN (requires DICOM images + GPU/MPS + PyTorch)
pip install -r requirements-cnn.txt
cd stage2_cnn && python train.py

# 7. Evaluate Stage 2 with TTA
python evaluate.py

# 8. Run late-fusion (Stage 3)
python fusion.py

# 9. Launch web app
cd .. && ./run.sh

---

⚕️ Intended Use & Disclaimer

MammoAI is a research and educational project. It is not a medical device, has not been cleared or approved by any regulator (FDA, MHRA, CE, etc.), and must not be used for clinical diagnosis, screening decisions, or patient care. All outputs are experimental and must never replace a qualified radiologist. Do not upload identifiable patient data. Use is entirely at your own risk under the MIT License.

---

Stage 2 CNN Training Details

Hyperparameter	Value
Backbone	EfficientNet-B4 (ImageNet pretrained)
Image size	512 × 512
Batch size	8
Phase 1 — Warmup	5 epochs, LR=1e-3, backbone frozen
Phase 2 — Fine-tune	up to 25 epochs, LR=1e-4, cosine annealing
Early stopping patience	7 epochs
Label smoothing	ε = 0.05
Best checkpoint	Epoch 24, val AUC = 0.8156
TTA passes	5 (original + H-flip + V-flip + ±10° rotation)
Optimal threshold	0.39
Hardware	Apple M3 MPS

Resume interrupted training:

cd stage2_cnn && python train.py --resume

---

SHAP Feature Importance (Stage 1)

| Feature | Mean |SHAP| | |---|---| | BI-RADS Assessment | 1.29 | | Morphology Risk | 0.85 | | Mass Margin Risk | 0.33 | | Subtlety | 0.26 | | Calc Type Risk | 0.22 |

---

Cross-Dataset Protocol (VinDr-Mammo)

# 1. Get free credentialed access at physionet.org
# 2. Complete CITI training + sign data use agreement
# 3. Download CSVs only (~2 MB):
#    curl -u YOUR_USERNAME https://physionet.org/files/vindr-mammo/1.0.0/breast-level_annotations.csv \
#         -o vindr-mammo/breast-level_annotations.csv
# 4. Run cross-dataset study:
python src/cross_dataset.py --vindr /path/to/vindr-mammo/1.0.0

---

Citation

@article{hamza2026mammoai,
  author  = {Hamza, Ali},
  title   = {MammoAI: A Multi-Modal Interpretable Framework for Breast Cancer
             Detection — Fusing Clinical Features with Deep Learning on CBIS-DDSM},
  journal = {Computers in Biology and Medicine},
  year    = {2026},
  note    = {Under review},
  url     = {https://github.com/aliaht99/MammoAI}
}

Dataset:

@article{lee2017cbisddsm,
  author  = {Lee, Rebecca Sawyer and Gimenez, Francisco and Hoogi, Assaf and Rubin, Daniel},
  title   = {Curated Breast Imaging Subset of DDSM (CBIS-DDSM)},
  journal = {The Cancer Imaging Archive},
  year    = {2017},
  doi     = {10.7937/K9/TCIA.2016.7O02S9CY}
}

---

License

MIT — see LICENSE

Author

Ali Hamza — MSc Advanced Engineering Management, Leeds Beckett University
alihamza.aht.99@gmail.com | github.com/aliaht99

Acknowledgements

• Dataset: TCIA CBIS-DDSM
• EfficientNet: Tan & Le, ICML 2019
• SHAP: Lundberg & Lee, NeurIPS 2017
• GradCAM: Selvaraju et al., ICCV 2017