Multimodal 3D Registration for Dental Imaging

Digital dentistry increasingly relies on three-dimensional data acquired from different imaging technologies, each providing complementary information about dental anatomy. Intraoral Scans (IOS) offer detailed surface representations of the visible crowns and gingiva, while Cone-Beam Computed Tomography (CBCT) provides volumetric information about tooth roots, alveolar bone, and surrounding structures. Combining these sources can therefore support a more complete representation of dentoalveolar anatomy and enable advanced applications in orthodontics, treatment planning, and computer-assisted dental procedures. This thesis addresses the registration of IOS and CBCT data, a challenging problem due to the limited anatomical overlap between the two modalities, their different geometric representations, and the presence of artifacts, missing structures, and large initial pose differences. To overcome these limitations, a coarse-to-fine framework combining learning-based anatomical landmark detection with robust geometric registration is proposed. Corresponding dental landmarks are independently predicted from the IOS mesh and the CBCT volume and used to estimate an initial rigid transformation through a confidence-weighted RANSAC-Kabsch procedure. The alignment is then refined using point-to-plane Iterative Closest Point on the reconstructed dental surfaces. This design exploits anatomical information to obtain a reliable global initialization while using local surface geometry to improve registration accuracy. The approach is evaluated on the MICCAI STS 2026 dataset. The results show that landmark-based initialization and surface refinement complement each other and improve robustness in the presence of uncertain correspondences and large initial misalignments. Overall, this work contributes to the development of automated methods for integrating heterogeneous 3D dental data, combining deep learning and geometric techniques within a unified registration pipeline.