APCReg AI aligns CBCT and intraoral scans with submillimeter accuracy

APCReg, developed by Xincan Zheng and colleagues (arXiv:2608.09993), tackles a persistent bottleneck in dental surgical planning: aligning cone-beam computed tomography (CBCT) with intraoral scans (IOS). These imaging modalities differ fundamentally in geometry and resolution, and clinical cases often involve limited overlap and large initial pose offsets. Existing automated methods fail on these hard cases, forcing clinicians to rely on manual alignment. APCReg's two-stage design directly addresses those failure modes. Its Multi-View Anatomical Coarse Registration (MACR) performs ordered orthogonal projection alignment from buccal, proximal, and occlusal views, decomposing the full six-degree-of-freedom search into manageable steps before 3D refinement. Then, Overlap-Aware Residual Registration (OARR) uses shared KPConv features, a folded arch-length cue, overlap-gated cross-attention, and Sinkhorn matching to refine the pose while suppressing non-overlapping regions.

The framework also introduces a dental-arch-structured hypothesis selection stage that evaluates diverse pose candidates on held-out reliable correspondences, plus a ground-truth-free coarse-retention guard that conditionally keeps a geometrically reliable coarse pose when refinement might drift. This reliability control is crucial for clinical trust. In experiments on 60 held-out jaw pairs, APCReg posted a submillimeter mean Chamfer distance of 0.87mm and a Hausdorff distance of 2.92mm under its evaluation protocol, outperforming all open-source baselines across six reported metrics. The practical impact is significant: accurate CBCT-IOS registration is a prerequisite for guided implant surgery, orthognathic planning, and digital orthodontics. By automating a step that previously required manual clinician fiddling, APCReg could reduce planning time and improve surgical precision in routine dental practice.