AI-assisted versus point-based registration for digital planning in CBCT images with metallic artifacts

BACKGROUND: Accurate registration between cone-beam computed tomography (CBCT) images and intraoral scans is fundamental to digital implant planning because registration errors may propagate throughout the digital workflow and affect the accuracy of computer-guided implant surgery. Conventional point-based registration relies on manually selected anatomical landmarks and may be compromised by metallic artifacts that obscure dental structures. Artificial intelligence (AI)-assisted registration has recently emerged as a promising alternative; however, evidence comparing its performance with conventional point-based registration under clinically challenging conditions remains limited. This study aimed to compare the registration accuracy (primary outcome) and registration duration (secondary outcome) of AI-assisted and conventional point-based registration in patients presenting with metallic artifacts on CBCT images. METHODS: This study included 45 eligible patients with corresponding full-arch CBCT datasets and intraoral scans. Registration accuracy was evaluated at the arch level by measuring the mean discrepancy at three predefined anatomical landmarks using Geomagic Control X software. Registration duration was recorded for each workflow. Arches were classified according to artifact burden into mild (> 0-33%), moderate (> 33-66%), and severe (> 66%) categories. Because the data were non-normally distributed, paired comparisons were performed using the Wilcoxon signed-rank test. Generalized estimating equations (GEE) were used as a sensitivity analysis to account for clustering of maxillary and mandibular arches within the same patient. Intra-observer reliability was assessed using the intraclass correlation coefficient (ICC). RESULTS: A total of 78 paired arch observations were available for registration accuracy analysis. AI-assisted registration demonstrated significantly lower registration discrepancies than conventional point-based registration (median 0.190 mm [IQR 0.121-0.275] vs. 0.268 mm [IQR 0.200-0.418]; p < 0.001). Registration discrepancy increased with increasing artifact burden for both techniques, although AI-assisted registration consistently demonstrated lower discrepancies across all artifact severity groups. Registration duration was comparable between the two workflows. The GEE sensitivity analysis confirmed the primary findings after accounting for within-patient clustering. Intra-observer reliability was excellent (ICC = 0.985; 95% CI: 0.968-0.993). CONCLUSIONS: AI-assisted registration achieved significantly lower registration discrepancies than conventional point-based registration while maintaining comparable registration duration in CBCTs with metallic artifacts on CBCT images. These findings suggest that AI-assisted registration may improve the accuracy and consistency of multimodal image registration within the digital implant planning workflow, particularly in clinically challenging cases. Further prospective multicenter studies involving multiple implant-planning software platforms and broader clinical scenarios are required to confirm these findings.