FedASAP：基于激活统计的结构化自适应剪枝，用于脑MRI病变分割的高效个性化联邦学习

中文摘要 AI 辅助

在医学影像中，开发稳健的深度学习模型需要来自不同领域的数据。然而，保护患者隐私的监管政策限制了数据共享。联邦学习（FL）通过在不集中数据的情况下实现协作模型训练来解决这一问题。然而，各客户端中心之间的数据异质性要求个性化模型以提升性能。资源有限的客户端可能难以高效训练或部署大型深度神经网络。自适应模型剪枝通过减小模型尺寸并实现个性化联邦学习适应来应对这两个挑战。尽管已有研究探索了联邦学习中的自适应剪枝方法，但其在分割等密集预测任务上的有效性仍未得到探索。为填补这一空白，我们提出了基于激活统计的结构化自适应剪枝（FedASAP），该方法利用基于激活的特征来指导每个客户端的滤波器移除。通过在逐滤波器激活统计上学习一个轻量级分类器，FedASAP优化了重要性评分排名，并在异构联邦设置中生成紧凑、个性化的分割模型。我们的结果证明了该方法的有效性，在两个多中心脑病理分割数据集：联邦肿瘤分割（FeTS）和白质高信号（WMH）上，分别取得了0.796和0.743的Dice分数，优于现有最先进方法，同时参数数量分别减少了45%和73%。

英文摘要

In medical imaging, developing robust deep learning models requires data from various domains. However, regulatory policies protecting patient privacy restrict data sharing. Federated learning (FL) addresses this by enabling collaborative model training without centralizing data. Yet, data heterogeneity across client centers requires personalized models to enhance performance. Clients with limited resources may struggle to efficiently train or deploy large deep neural networks. Adaptive model pruning tackles both challenges by reducing model size while enabling personalized FL adaptation. Although research has investigated adaptive pruning methods in FL, their effectiveness on dense prediction tasks like segmentation remains unexplored. To address this gap, we propose Activation Statistics-driven structured Adaptive Pruning (FedASAP), which uses activation-based features to guide filter removal for each client. By learning a lightweight classifier on per-filter activation statistics, FedASAP refines importance-score rankings and produces compact, personalized segmentation models in heterogeneous federated settings. Our results show the approach's effectiveness by achieving better Dice scores of 0.796 and 0.743, than state-of-the-art, while reducing parameter count by 45% and 73% on two multi-centric brain pathology segmentation datasets: Federated Tumour Segmentation (FeTS) and White Matter Hyperintensities (WMH), respectively.

发表机构

• Indian Institute of Science（印度科学研究所）

机构由 AI 辅助整理，请以论文原文为准。