Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

pFLFE: Cross-silo Personalized Federated Learning via Feature Enhancement on Medical Image Segmentation

Luyuan Xie, Manqing Lin, Siyuan Liu, ChenMing Xu, Tianyu Luan, Cong Li, Yuejian Fang, Qingni Shen, Zhonghai Wu

TL;DR

pFLFE tackles client drift in cross-silo federated medical image segmentation by introducing Local Feature Enhancement and decoder-based personalization. It employs a four-stage FL framework with two global aggregations, plus a fast-converging variant FC-pFLFE that reduces communication rounds. Across optic disc/cup, polyp, and prostate segmentation tasks on 17 datasets, pFLFE achieves state-of-the-art or near-centralized performance with improved stability and faster convergence, and demonstrates robust generalization to unseen clients. The approach improves the separation of foreground and background features as evidenced by lower KL divergence and demonstrates that decoder personalization can yield stable, scalable FL in privacy-sensitive medical settings.

Abstract

In medical image segmentation, personalized cross-silo federated learning (FL) is becoming popular for utilizing varied data across healthcare settings to overcome data scarcity and privacy concerns. However, existing methods often suffer from client drift, leading to inconsistent performance and delayed training. We propose a new framework, Personalized Federated Learning via Feature Enhancement (pFLFE), designed to mitigate these challenges. pFLFE consists of two main stages: feature enhancement and supervised learning. The first stage improves differentiation between foreground and background features, and the second uses these enhanced features for learning from segmentation masks. We also design an alternative training approach that requires fewer communication rounds without compromising segmentation quality, even with limited communication resources. Through experiments on three medical segmentation tasks, we demonstrate that pFLFE outperforms the state-of-the-art methods.

pFLFE: Cross-silo Personalized Federated Learning via Feature Enhancement on Medical Image Segmentation

TL;DR

pFLFE tackles client drift in cross-silo federated medical image segmentation by introducing Local Feature Enhancement and decoder-based personalization. It employs a four-stage FL framework with two global aggregations, plus a fast-converging variant FC-pFLFE that reduces communication rounds. Across optic disc/cup, polyp, and prostate segmentation tasks on 17 datasets, pFLFE achieves state-of-the-art or near-centralized performance with improved stability and faster convergence, and demonstrates robust generalization to unseen clients. The approach improves the separation of foreground and background features as evidenced by lower KL divergence and demonstrates that decoder personalization can yield stable, scalable FL in privacy-sensitive medical settings.

Abstract

In medical image segmentation, personalized cross-silo federated learning (FL) is becoming popular for utilizing varied data across healthcare settings to overcome data scarcity and privacy concerns. However, existing methods often suffer from client drift, leading to inconsistent performance and delayed training. We propose a new framework, Personalized Federated Learning via Feature Enhancement (pFLFE), designed to mitigate these challenges. pFLFE consists of two main stages: feature enhancement and supervised learning. The first stage improves differentiation between foreground and background features, and the second uses these enhanced features for learning from segmentation masks. We also design an alternative training approach that requires fewer communication rounds without compromising segmentation quality, even with limited communication resources. Through experiments on three medical segmentation tasks, we demonstrate that pFLFE outperforms the state-of-the-art methods.
Paper Structure (10 sections, 3 equations, 4 figures, 7 tables)

This paper contains 10 sections, 3 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. pFLFE Framework
  4. Local Feature Enhancement
  5. Fast Converging pFLFE
  6. Experiment Setup
  7. Results and Discussion
  8. Conclusion
  9. Acknowledgements.
  10. Disclosure of Interests.

Figures (4)

  • Figure 1: Previous personalized federated learning approaches suffer from client drifting problems. (a). Training progress of previous approaches (red and blue line) compared with our approach (green line). (b)(c) Foreground (cyan points) and background (red points) feature distribution t-SNE for FedRep and LG-FedAvg. (d) Feature distribution t-SNE for our pFLFE. The stability of previous works' training processes is not ideal and their foreground and background feature distribution are evidently overlapped. For comparison, pFLFE has a more stable training process, evidently split foreground and background feature distribution, and better segmentation accuracy.
  • Figure 2: The overview of (a) pFLFE framework, (b) Local Feature Enhancement and (c) FC-pFLFE framework.
  • Figure 3: Visualized comparison of personalized methods on three datasets. From each dataset, we randomly select two samples from different clients to form the visualization. (a-j) Segmentation results by models trained with FedAVG, SCAFFOLD, FedProx, Ditto, APFL, LG-FedAvg, FedRep, FedSM, LC-Fed and our method pFLFE; (k) Ground truths (denoted as ‘GT’);
  • Figure 4: Training progress of our pFLFE compared with previous results on 3 tasks. The green and black line is our pFLFE and FC-pFLFE training progress, the red, blue, and yellow lines are training progress of previous approaches. The black dashed line is the result of the centralized method. It is easy to observe that our pFLFE has better performance and more stable, faster-converged training progress. The FC-pFLFE (red vertical line) reaches the near-optimal solution is significantly lower than that of pFLFE (blue vertical line).