Federated Learning for Histopathological Image Segmentation: A Privacy-Preserving Study Using the EBHI-Seg Dataset

Abstract

Federated learning (FL) enables collaborative model training without centralizing sensitive medical data, making it attractive for digital pathology. This study evaluates FL for histopathological image segmentation using the EBHI-Seg dataset under independent and identically distributed (IID) and classstratified non-IID client partitions. We simulate six decentralized clients and compare three segmentation architectures (UNet, DeepLabV3, FCN) with three FL strategies (FedAvg, FedProx, FedOptimizer), yielding 18 experimental configurations. Performance is assessed using Dice coefficient, Intersection over Union (IoU), loss, and betweenround convergence volatility. The results show that the tested aggregation algorithms achieve similar final segmentation accuracy, while class-stratified non-IID partitioning lowers Dice scores and substantially increases convergence oscillation. These findings indicate that client-level heterogeneity, rather than the choice among the tested aggregators, is the dominant challenge for privacy-preserving federated histopathology segmentation.