Selective Downsampling for Fast and Accurate 3D Global Registration with Applications in Medical Imaging

Robust global point-cloud registration remains a key challenge in robotic neurosurgery, particularly for markerless patient registration, where anatomical surface acquisition can be incomplete and noisy. This paper proposes practical pre-processing steps, defines performance criteria, and evaluates the keypoint-based 4-Points Congruent Set (K4PCS) and Super4PCS algorithms for global registration. Experiments are conducted on surface point clouds segmented from real patient head CT scans, while all measurement errors are synthetically simulated by applying clinically relevant perturbations, including large initial misalignment, Gaussian (CT-like) and non-Gaussian (camera-like) noise injection, and partial scans, across 30 different poses. Registration performance is quantified using pose errors and noise-aware surface-distance/overlap measures, while run-time is assessed under a newly developed selective downsampling strategy and compared to standard voxel downsampling. Results show that both algorithms reliably converge from substantial misalignment and remain robust after noise injection, with computation times ranging from 0.1 s to >15 min. Partial-to-whole registration achieves accuracy comparable to whole-to-whole registration (errors <10−3 mm), but typically exceeds real-time run-times. Selective downsampling provides a clear improvement in precision and, in most cases, also improves speed compared to the voxel-based downsampling method. Overall, the results indicate that robust and real-time markerless head registration is feasible under clinical conditions.