DentalRegiformer: A Fully Automated Deep Learning Framework for Coarse-to-Fine Registration between Intraoral Scan and CBCT Image

Purpose: Precise registration between intraoral scans (IOSs) and CBCT images is essential for many digital dentistry applications. The purpose of this study was to develop DentalRegiformer, a fully automated, end-to-end deep learning framework that ensures accurate registration between IOS and CBCT images. Materials and Methods: The network was trained on 88 paired CBCT-IOS datasets from 44 patients, with ground-truth alignments obtained through manual registration performed by trained dental students. To enhance clinical accuracy and efficiency, a multi-stage coarse-to-fine strategy is proposed to integrate segmentation, classification, and registration into a unified pipeline. First, tooth segmentation is performed on CBCT volumes using 3D nnU-Net to generate segmented tooth regions for alignment. Simultaneously, IOS datasets are classified as maxillary or mandibular dentition using a hybrid representation that combines geometric descriptors with point cloud features extracted via a DGCNN network. Registration is then conducted in two phases: coarse alignment using a DCP framework for feature encoding and probabilistic correspondence estimation to compute the initial rigid transformation, followed by fine alignment using GeoTransformer's local-to-global iterative refinement, which alternates between local feature aggregation and global context integration to progressively enhance correspondence accuracy and achieve precise rigid registration. Results: The segmentation model achieved robust results, with a Dice coefficient of 0.924, precision of 0.916, and recall of 0.934. The DGCNN classifier perfectly distinguished maxillary and mandibular arches, achieving sensitivity, specificity, ROC, and accuracy of 1.000. For registration, DentalRegiformer accurately predicted transformation between IOS and CBCT images, yielding an overall RMSE of 0.556 mm; MAEs of 0.231mm, 0.329 mm, and 0.283 mm along the x-, y-, and z-axes, respectively; and inference time of 19.623 seconds. In comparison, the manual method based on combined pair-points and ICP had an RMSE of 0.615 mm; MAEs of 0.244 mm, 0.347 mm, and 0.298 mm along the x-, y-, and z-axes, respectively; and processing time of 431.6 seconds. Discussion and Conclusion: Therefore, DentalRegiformer is a fully automated, deep learning-based solution that minimizes manual intervention and improves clinical efficiency. The framework streamlines digital dental workflows, reduces operator-dependent errors, and facilitates automated, personalized dental care.