qhjqhj00 fl-medsegbench-eval
FL-MedSegBench: A Comprehensive Benchmark for Federated Learning on Medical Image Segmentation — Zhu et al. (2026) (arXiv:2603.11659, 2026)
What this evaluates
Evaluates federated learning methods for medical image segmentation under non-IID data distributions, measuring segmentation accuracy and robustness across multiple clinical tasks and imaging modalities. It compares generic and personalized FL approaches against local training baselines to assess client drift, fairness, and generalization.
Datasets
- Fed-Vessel — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-Prostate — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-COSAS — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-BUS — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-MG — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-Polyp — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-Pancreas — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- Fed-M&Ms — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
- FeTS2022 — total ?; splits: train (-1); repo https://github.com/meiluzhu/FL-MedSegBench
Metrics
Dice(primary) — range: [0, 100]- Dice coefficient calculated as 2 * |A ∩ B| / (|A| + |B|), where A and B are the predicted and ground-truth segmentation masks. Reported as a percentage.
Input / output format
Input: 2D or 3D medical images with corresponding ground-truth segmentation masks, partitioned across multiple federated clients with non-IID data distributions.
Output: Predicted segmentation masks for each client's local evaluation set.
Scoring recipe
def compute_dice(pred_mask, gt_mask):
intersection = np.sum(pred_mask * gt_mask)
union = np.sum(pred_mask) + np.sum(gt_mask)
return (2.0 * intersection / union) * 100 if union > 0 else 0.0
# Average Dice across test samples per client, then report per-client or global mean.
Common pitfalls
- Reporting only global average Dice, which obscures severe performance drops on specific clients under non-IID splits.
- Failing to distinguish between generic FL (gFL) and personalized FL (pFL) baselines when comparing against local training baselines.
- Ignoring the fixed communication rounds (T=100) constraint, which heavily influences convergence and fairness metrics.
Evidence (verbatim from paper)
The number of communication rounds $T$ is set to 100 for all datasets. ... TABLE V: The segmentation performance (Dice) of FL methods on Fed-COSAS and Fed-BUS datasets $(%)$.
Citation
@misc{zhu2026flmedsegbench,
title={FL-MedSegBench: A Comprehensive Benchmark for Federated Learning on Medical Image Segmentation},
author={Zhu et al. (2026)},
year={2026},
note={arXiv:2603.11659}
}
- arXiv: 2603.11659