FL-CheX: A Federated Learning Benchmark for Chest X-ray Classification with Multi-Source Heterogeneity

FL-CheX is a federated learning benchmark dataset designed for chest X-ray disease classification under realistic multi-source heterogeneity. The dataset simulates three types of real-world distribution shifts: 1. Demographic heterogeneity (age and gender imbalance)2. Disease-based non-IID hospital specialization3. Scanner/device heterogeneity with varying noise levels The benchmark is built from CheXpert-derived label embeddings and DenseNet-121 feature extraction using TorchXRayVision, followed by dimensionality reduction using Johnson-Lindenstrauss projection. The framework supports evaluation of Federated Learning algorithms including FedAvg and FedProx, with fairness analysis using demographic parity, equalized odds, and AUC gap metrics. This repository includes:- Federated learning training code- 15 pre-partitioned FL client nodes- Precomputed embeddings- Evaluation results and visualization figures This work aims to support reproducible research in federated medical imaging and fairness-aware machine learning.