🌸 MAMA-MIA Challenge 2025: Primary Tumour Segmentation in DCE-MRI

Authors: Beyza Zayim¹, Aissiou Ikram², Boukhiar Naima³
¹ Université de Bourgogne, Dijon 21000, France
² Algeria
³ University of Algiers 1 Ben Youcef Ben Khedda, Algeria

🧬 Overview

This repository contains our submission for the MICCAI 2025 MAMA-MIA Challenge, focusing on primary tumor segmentation in Dynamic Contrast-Enhanced MRI (DCE-MRI) breast cancer data. Our approach leverages the nnU-Net framework with a selective training strategy based on image quality and center-specific variability.

Key Findings

• Data Quality Matters: Including low-quality ISPY scans impaired segmentation performance, even with advanced preprocessing
• Center-Specific Strategy: Training on high-quality DUKE and NACT data with early-phase images (0000–0002) yielded more robust results
• Best Performance: Achieved validation Dice score of 0.72 using multi-phase (phases 1-3) DUKE+NACT data
• Postprocessing: Keeping only the largest connected component significantly improved results

📁 Repository Structure

.
├── sample_code_submission/
│   ├── model.py                    # Main model implementation
│   ├── Dataset105_full_image/      # nnU-Net model directory
│   │   └── nnUNetTrainer_nnUNetPlans_3d_fullres/
│   │       ├── fold_0/
│   │       ├── fold_1/
│   │       ├── fold_2/
│   │       ├── fold_3/
│   │       └── fold_4/
└── ReadMe.md

⚙️ Method

Pipeline Overview

1. Dataset Preparation: Multi-center DCE-MRI data from DUKE, NACT, ISPY1, and ISPY2
2. Preprocessing: Isotropic resampling (1mm³) and Z-score normalization
3. Model Training: nnU-Net 3D full-resolution with 5-fold cross-validation
4. Postprocessing: Largest connected component filtering

Key Technical Details

• Model: nnU-Net 3D full-resolution configuration
• Input: 3 temporal phases (0000-0002) for optimal performance
• Training Data: Selective use of high-quality DUKE (247 cases) and NACT (64 cases) data
• Optimization: Adam optimizer (lr=1e-4, weight_decay=3e-5)
• Hardware: NVIDIA A100 80GB GPU
• Framework: PyTorch 2.6.0, CUDA 11.8

💻 Installation

Prerequisites

# Create conda environment
conda create -n mama-mia python=3.9
conda activate mama-mia

# Install PyTorch
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

# Install nnU-Net v2
pip install nnunetv2

Usage

Environment Setup

export nnUNet_raw="/path/to/your/nnUNet_raw"
export nnUNet_preprocessed="/path/to/your/nnUNet_preprocessed"
export nnUNet_results="/path/to/your/nnUNet_results"

Running Inference

from model import Model

# Initialize model
model = Model(dataset=your_dataset, 
              dataset_id="Dataset105_full_image", 
              config="3d_fullres")

# Run prediction
output_dir = model.predict_segmentation("/path/to/output")

Model Features

• Automatic Preprocessing: Isotropic resampling and normalization
• Multi-phase Support: Handles 3 temporal DCE phases
• Breast Region Masking: Uses provided breast coordinates for focused segmentation
• Robust Postprocessing: Largest connected component filtering
• Error Handling: Comprehensive logging and fallback mechanisms

📊 Results

Performance Metrics

Experiment	Data	Phases	Validation Dice	DUKE_001	ISPY1_1183	ISPY2_332	NACT_64
Final Model	DUKE+NACT	3 (1-3)	0.72	0.9394	0.7640	0.8967	0.9580
Single Phase	DUKE+NACT	1 (phase2)	0.62	0.8625	0.7196	0.8111	0.9514
All Centers	1200 cases	1 (phase2)	0.45	0.8894	0.6739	0.5227	0.9334

Key Insights

1. Quality > Quantity: Selective high-quality data outperformed larger, mixed-quality datasets
2. Multi-phase Benefits: Using 3 temporal phases improved generalization
3. Center Variability: DUKE and NACT data showed superior consistency compared to ISPY datasets
4. Postprocessing Impact: Largest connected component filtering eliminated false positives

Technical Contributions

• Quality-Aware Training: Demonstrated that selective, high-quality data outperforms larger mixed datasets
• Multi-phase Integration: Showed benefits of temporal information in DCE-MRI segmentation
• Robust Pipeline: Implemented comprehensive error handling and fallback mechanisms
• Postprocessing Innovation: Applied connected component analysis for improved segmentation

MAMA-MIA Dataset:

@article{garrucho2025,
  title={A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations},
  author={Garrucho, Lidia and Kushibar, Kaisar and Reidel, Claire-Anne and Joshi, Smriti and Osuala, Richard and Tsirikoglou, Apostolia and Bobowicz, Maciej and Riego, Javier del and Catanese, Alessandro and Gwoździewicz, Katarzyna and Cosaka, Maria-Laura and Abo-Elhoda, Pasant M and Tantawy, Sara W and Sakrana, Shorouq S and Shawky-Abdelfatah, Norhan O and Salem, Amr Muhammad Abdo and Kozana, Androniki and Divjak, Eugen and Ivanac, Gordana and Nikiforaki, Katerina and Klontzas, Michail E and García-Dosdá, Rosa and Gulsun-Akpinar, Meltem and Lafcı, Oğuz and Mann, Ritse and Martín-Isla, Carlos and Prior, Fred and Marias, Kostas and Starmans, Martijn P A and Strand, Fredrik and Díaz, Oliver and Igual, Laura and Lekadir, Karim},
  journal = {Scientific Data},
  year = {2025},
  doi = {10.1038/s41597-025-04707-4},
  pages = {453},
  number = {1},
  volume = {12}
}

Acknowledgments

This work was supported by the MAMA-MIA Challenge 2025. We thank the challenge organizers, data contributors, and the broader medical imaging community for their open-source tools and resources.

License

This project is licensed under the CC BY-NC 4.0 License - see the LICENSE file for details.

Contact

• Beyza Zayim: Université de Bourgogne, Dijon, France
• Aissiou Ikram: Algeria
• Naima Boukhiar: University of Algiers 1 Ben Youcef Ben Khedda, Algeria

For questions about the implementation, please open an issue in this repository.