Divergence-Aware Training with Automatic Subgroup Mitigation for Breast Tumor Segmentation

Eleonora Poeta¹ · Luisa Vargas² · Daniele Falcetta² · Vincenzo Marciano'²

Eliana Pastor¹ Tania Cerquitelli¹ · Elena Baralis¹ · Maria A. Zuluaga²

¹Politecnico di Torino, Italy ²EURECOM, France

🏆 Winner of the Best Paper Award at MAMA-MIA Challenge at MICCAI 2025

📌 Overview

Deep learning models for breast tumor segmentation in DCE-MRI often exhibit performance disparities across different demographic and clinical subgroups. This research addresses these fairness concerns by proposing a subgroup-aware in-processing mitigation strategy.

Key Features:

• Divergence-based Regularization: Integrates a fairness-aware loss directly into the training loop.

• Interpretable Metadata: Leverages clinical data (age, menopausal status, breast density) to identify underperforming subgroups.

• Dynamic Loss Weighting: Automatically assigns higher weights to samples from subgroups that diverge from average performance.

• No Post-processing Required: Improves fairness and segmentation quality during the training phase without needing external data.

📊 Results

Our method was evaluated on the MAMA-MIA Challenge dataset, demonstrating:

• Significant improvements in fairness scores across subgroups;

• Maintained or improved segmentation quality (Dice score) for the overall population;

• Enhanced clinical trustworthiness by reducing performance gaps in harder-to-segment subpopulations.

---

⚡️ Fast setup with uv (recommended)

We recommend using uv for a fast, reproducible, and dependency-safe setup.

0️⃣ Prerequisites

• Python 3.10 (recommended)
• Git

1️⃣ Clone the repository

git clone https://github.com/robustml-eurecom/FairMedSeg.git
cd FairMedSeg

2️⃣ Create and activate the virtual environment

uv init
source .venv/bin/activate

3️⃣ Install Torch and all the other dependencies

uv add torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
uv add -r requirements.txt
uv pip install pyradiomics==3.0.1 --no-build-isolation
uv sync

📁 Data layout expected by the loaders

The MONAI data loader (dataloader/load_monai_metadata.py) expects the dataset to be organized as follows:

• Images
  One folder per patient ID, containing the DCE-MRI phase volumes (.nii.gz).

• Segmentations
  One segmentation file per patient, named <patient_id>.nii.gz.

• Metadata
  An Excel file containing clinical and imaging information, with at least a patient_id column.

• Train / test split
  A CSV file specifying the dataset split, with train_split and test_split columns listing patient IDs.

Example directory structure

<DATA_ROOT>/
├── images/
│   └── PAT_0001/
│       ├── phase_0.nii.gz
│       ├── phase_1.nii.gz
│       ├── phase_2.nii.gz
│       └── ...
├── segmentations/
│   └── expert/
│       ├── PAT_0001.nii.gz
│       └── ...
├── patient_info_files/
│   ├── PAT_0001.json
│   └── ...
├── clinical_and_imaging_info.xlsx
└── train_test_splits.csv

🗣️ Citation

If you use this repository, please cite the associated paper:

@inproceedings{poeta2025divergence,
  title     = {Divergence-Aware Training with Automatic Subgroup Mitigation for Breast Tumor Segmentation},
  author    = {Poeta, Eleonora and Vargas, Luisa and Falcetta, Daniele and Marciano', Vincenzo and Pastor, Eliana and Cerquitelli, Tania and Baralis, Elena and Zuluaga, Maria A.},
  booktitle = {Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care (Deep-Breath 2025)},
  series    = {Lecture Notes in Computer Science},
  volume    = {16142},
  pages     = {52--62},
  year      = {2025},
  publisher = {Springer},
  doi       = {10.1007/978-3-032-05559-0_6}