In medical imaging, the segmentation of skin lesions plays a vital role in detecting and treating skin cancer. Deep learning demonstrates its efficacy in this process. However, it largely relies on extensive and well-annotated datasets that are often limited by healthcare agencies privacy restrictions and institutional data silos. Federated Learning (FL) emerged as a boon, enabling collaborative training without sharing data. Yet, in a real-world setting, healthcare bodies may possess various computational capacities that can affect the consistency of the FL framework, posing the requirement of generalising the model architecture. This study proposes a Decentralised Federated Learning (DFL) framework to improve model generalisation for Skin Lesion Segmentation (SkinSeg). It incorporates a novel Voting (V)-based client selection mechanism to identify the most suitable local model based on performance metrics and dataset size. The selected model is then subjected to Gradual Pruning (GP) via a modified Lottery Ticket Hypothesis (LTH) to reduce model complexity while preserving segmentation quality. The pruned model is then broadcast to all clients for further training. The VGPDFL-SkinSeg substantially improved over State-Of-The-Art FL frameworks on benchmark datasets HAM10K, ISIC-2016/17/18 and DermIs+DermQuest. It achieved a client-wise average Dice Coefficient (DSC) of 90.09%, 96.60% Accuracy, 82.45% meanIOU, 13.63% HD95 and 5.20% ASSD. Initially, each client starts with different segmentation models, reflecting practical diverse systems, and gradually converges towards homogeneity. The study shows that gradual pruning up to 40% yields better segmentation quality than fixed pruning at the beginning and is consistent with client scaling.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号