FairFedMed: Benchmarking Group Fairness in Federated Medical Imaging with FairLoRA

---

🎯 Motivation

Fairness in medical FL remains underexplored due to heterogeneous data and lack of demographic-aware benchmarks. We introduce:

📊 FairFedMed Dataset

First medical FL benchmark with demographic annotations

• FairFedMed-Oph: 2D/3D ophthalmology images with 6 demographic attributes
• FairFedMed-Chest: Cross-institutional CheXpert + MIMIC-CXR with 3 attributes

🚀 FairLoRA Framework

Fairness-aware FL via SVD-based low-rank adaptation

• Customizes singular values per demographic group
• Shares singular vectors for efficiency
• Achieves superior performance-fairness trade-offs

---

📦 Dataset

📥 Download: FairFedMed Dataset on Google Drive

🏥 FairFedMed-Oph

• Modalities: Paired 2D SLO fundus images and 3D OCT B-Scans
• Scale: 15,165 patients for glaucoma detection
• Demographics: 6 attributes (age, gender, race, ethnicity, preferred language, marital status)

🫁 FairFedMed-Chest

• Sources: CheXpert + MIMIC-CXR
• Setup: 2 clients simulating real cross-institutional FL
• Demographics: 3 attributes (age, gender, race)

📊 Dataset Statistics

🗂️ Data Structure

DATA/
  ├── fairfedmed/
  │   ├── all  # a dir that stores all raw data files
  │   │   ├──filename1.npz 
  │   │   ├──filename2.npz 
  │   │   └── ...
  │   ├── meta_all.csv
  │   ├── meta_site{k}_language_train.csv
  │   ├── meta_site{k}_language_test.csv
  │   ├── meta_site{k}_language.csv
  │   ├── meta_site{k}_race_train.csv
  │   ├── meta_site{k}_race_test.csv
  │   ├── meta_site{k}_race.csv
  │   ├── meta_site{k}_ethnicity_train.csv
  │   ├── meta_site{k}_ethnicity_test.csv
  │   ├── meta_site{k}_ethnicity.csv
  │   └── ...
  ├── fedchexmimic/
  │   ├── CheXpert-v1.0/  # symlink to CheXpert dataset
  │   ├── mimic/  # symlink to MIMIC-CXR dataset
  │   ├── meta_chexpert_age.csv
  │   ├── meta_chexpert_age_train.csv
  │   ├── meta_chexpert_age_test.csv
  │   ├── meta_chexpert_gender.csv
  │   ├── meta_chexpert_gender_train.csv
  │   ├── meta_chexpert_gender_test.csv
  │   ├── meta_chexpert_race.csv
  │   ├── meta_chexpert_race_train.csv
  │   ├── meta_chexpert_race_test.csv
  │   ├── meta_mimic_age.csv
  │   ├── meta_mimic_age_train.csv
  │   ├── meta_mimic_age_test.csv
  │   ├── meta_mimic_gender.csv
  │   ├── meta_mimic_gender_train.csv
  │   ├── meta_mimic_gender_test.csv
  │   ├── meta_mimic_race.csv
  │   ├── meta_mimic_race_train.csv
  │   └── meta_mimic_race_test.csv

---

🧬 Methodology

FairLoRA: A group fairness-aware federated learning model using SVD-based low-rank adaptation.

---

🏋️ Model Training

Download the dataset first, then run the training scripts:

🏥 FairFedMed-Oph (Ophthalmology)

# 2D SLO fundus images
sh scripts/fairfedlora_fairfedmed.sh       # ViT-B/16 backbone
sh scripts/fairfedlora_fairfedmed_rn50.sh  # ResNet50 backbone

# 3D OCT B-Scan images
sh scripts/fairfedlora_fairfedmed_oct.sh       # ViT-B/16 backbone
sh scripts/fairfedlora_fairfedmed_oct_rn50.sh  # ResNet50 backbone

🫁 FairFedMed-Chest (Chest X-ray)

sh scripts/fedchexmimic/fairfedlora_fedchexmimic.sh  # ViT-B/16 backbone

---

📊 Evaluation Metrics

Metric	Description
AUC	Area Under ROC Curve
ESAUC	Equalized Selection AUC
Group-wise AUC	AUC per demographic group
EOD	Equalized Odds Difference
SPD	Statistical Parity Difference

📄 Implementation

---

📈 Experimental Results

We compare FairLoRA with:

• Traditional FL: FedAvg, FedHEAL
• Prompt-based FL: PromptFL, FedOTP

---

🙏 Acknowledgements

This code is partially derived from:

• FedOTP - Federated Optimal Transport Prompting
• DASSL - Domain Adaptation/Generalization Library

---

📝 Citation

If you find this work useful, please cite:

@ARTICLE{11205878,
  author={Li, Minghan and Wen, Congcong and Tian, Yu and Shi, Min and Luo, Yan and Huang, Hao and Fang, Yi and Wang, Mengyu},
  journal={IEEE Transactions on Medical Imaging}, 
  title={FairFedMed: Benchmarking Group Fairness in Federated Medical Imaging with FairLoRA}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Biomedical imaging;Federated learning;Data models;Artificial intelligence;X-ray imaging;Three-dimensional displays;Robots;Ophthalmology;Benchmark testing;MIMICs;Group Fairness;Federated Learning;Medical Imaging;Low-rank Approximation (LoRA)},
  doi={10.1109/TMI.2025.3622522}}

---

⭐ Star this repo if you find it helpful!