FedU-KAN: Cloud-Enhanced Privacy-Preserving Federated Learning for Medical Image Segmentation Based on U-KAN

Abstract

Machine learning is gradually transforming medical image segmentation. However, its accuracy often relies on large-scale medical datasets, while centralized data collection raises serious privacy concerns. To address this issue, federated learning (FL) enables collaborative model training without sharing raw data, thus effectively protecting patient privacy. Despite this advantage, commonly used segmentation models, such as U-Net and its variants, typically have large parameter sizes, making them inefficient for local training on FL clients. To overcome this challenge, we propose FedU-KAN, a framework built upon the lightweight U-KAN architecture, tailored for federated medical image segmentation tasks. Moreover, we design an adaptive differential privacy mechanism that dynamically adjusts gradient clipping based on feature importance. This approach helps preserve anatomical details while reducing the risk of privacy leakage. We evaluate FedU-KAN on the CVC-ClinicDB and Kvasir-SEG datasets, where it achieves IoU scores of 87.09% and 83.38%, respectively—outperforming standard FL baselines. These results demonstrate that FedU-KAN can effectively balance privacy protection and model performance in real-world medical segmentation scenarios.

Ähnliche Arbeiten