Federated Learning Framework for Privacy-Preserving Depth-Based AR Surgical Registration

Abstract

The increasing use of AR in surgical navigation raises concerns over patient data privacy when training AI-enhanced registration models. We introduce a federated learning framework that enables distributed training of depth-based markerless registration networks across multiple hospitals without sharing raw patient data. Each local node trains a CNN-Transformer hybrid for point cloud alignment, and only encrypted weight updates are aggregated on a central server. Experiments conducted across three institutions with 2,000 intraoperative scans demonstrated a 32% reduction in generalization error compared with single-site models. Registration accuracy improved from 2. 1 mm to 1.3 mm, while training convergence time decreased by 27% due to adaptive aggregation. This work confirms the feasibility of collaborative yet privacy-preserving AR surgical registration pipelines.