Fed-ADApt: Federated Anytime Depth Adaptation for Resource-Aware Medical Image Segmentation
This study addresses the challenge of computational heterogeneity in federated medical image segmentation, which often restricts the participation of resource-constrained institutions. To this end, we propose a deep adaptive federated learning framework built upon a UNet architecture that incorporates multi-depth supervision and hierarchical aggregation mechanisms. This design enables participating nodes to dynamically select training and inference depths according to their local computational budgets, thereby achieving joint optimization of resource allocation. Experimental results demonstrate that the proposed framework attains performance comparable to full-capacity models on 3D segmentation tasks while substantially reducing both training and inference overhead. By facilitating equitable and efficient collaboration across heterogeneous environments, this approach effectively empowers low-resource institutions to participate in federated medical analysis.