FedASAP: Activation Statistics-driven Structured Adaptive Pruning for Efficient Personalized Federated Learning for Lesion Segmentation on brain MRI

Abstract

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.