Morphology-adaptive feature extraction and geometric consistency evaluation for robot-assisted orthopedic surgery registration

Registration is a key technique in Robot-Assisted Orthopedic Surgery (RAOS) for aligning pre-operative models with intra-operative anatomy. However, registration presents two major challenges in clinical practice. First, the restricted surgical field of view means the intra-operative point cloud is often only a small fragment of the entire bone. Second, diverse anatomical structures exhibit significant variations in physical scale and point cloud density. To overcome these challenges, we propose GeoNet, a robust registration framework driven by morphology-adaptive and geometric consistency. GeoNet utilizes a Morphology-Adaptive Feature Extraction (MAFE) module to autonomously calibrate voxelization and receptive fields based on the intrinsic sphericity and density of different bones. GeoNet follows a coarse-to-fine registration strategy. In the coarse node correspondence, we strategically remove cross-attention to prevent feature corruption in non-overlapping regions and use geometric consistency evaluation to eliminate erroneous correspondences. In the fine point matching, a linear self-attention is used to refine the feature. Extensive experiments on a clinical orthopedic dataset demonstrate that it maintains exceptional robustness even under a 20% overlap ratio, improving registration accuracy by a significant margin compared to baseline methods and showcasing immense potential for clinical applications. On standard public datasets, GeoNet surpasses the state-of-the-art (SOTA) methods, achieving a 0.7% performance improvement in registration accuracy.