A feature-driven registration method for low-overlap partial matching of CBCT and IOS dental models

Accurate registration of cone-beam computed tomography (CBCT) and intraoral scanner (IOS) dental models is essential for diagnosis and preoperative planning. However, reliable alignment remains challenging because IOS data primarily represent the crown surface whereas CBCT provides complete tooth morphology with lower surface fidelity, resulting in substantial cross-modal discrepancy and limited geometric overlap. This study proposes a training-free spectral-topological feature-driven registration framework (STFR) for robust CBCT-IOS alignment under low-overlap conditions. STFR resamples the CBCT point cloud using Divergence Index and Improved Euclidean Clustering Rules (DI-IECR), then establishes reliable coarse correspondences via curvature-topology features and geometric-topological confidence domains. During fine registration, Laplacian spectral features are integrated with geometric-topological descriptors in an adapted iterative closest point framework to improve global structural consistency and reduce convergence to anatomically incorrect local optima. A neighborhood-curvature-based strategy then improves crown-root continuity. The framework was evaluated on 14 paired CBCT-IOS tooth models covering different morphologies and overlap conditions, using surface-distance metrics, spatial overlap ratio, and landmark-based target registration error (TRE) at the case level. STFR achieved a mean RMSE of 0.475 mm and a mean landmark TRE of 0.621 mm, outperforming five representative baselines (ICP, NDT, CPD, FPFH-RANSAC, and a fine-tuned DCP), with RMSE reductions of 29%-59% and significantly lower error in every paired comparison (all p < 0.01). STFR thus provides an interpretable, robust approach for low-overlap CBCT-IOS registration, although prospective clinical validation remains necessary before routine use.