A Deep Learning-Based Method for Automatic Registration of Craniofacial CBCT and Multi-planar Temporomandibular Joint MRI.

OBJECTIVES

To develop a fully automated multi-stage registration framework for temporomandibular joint (TMJ) imaging that enables spatial alignment between large-field-of-view cone-beam computed tomography (CBCT) and multi-planar small-field-of-view magnetic resonance imaging (MRI).

METHODS

The framework combined landmark-based rigid initial registration with nnU-Net-based mask-guided refinement registration. Four anatomical landmarks, including the most superior, medial, and lateral ​points and the gonion point, were automatically detected to estimate the initial rigid transformation between CBCT and MRI. Condylar segmentation masks were subsequently used to refine the local alignment. The segmentation performance was qualitatively and quantitatively evaluated using the Dice similarity coefficient, Intersection over Union, precision, and recall. Landmark localization accuracy was evaluated quantitatively using the modality-specific landmark localization error (LLE) against manually annotated landmarks, whereas the final registration results were qualitatively reviewed by a clinician experienced in TMJ imaging.

RESULTS

The segmentation models showed high performance for automated condyle and articular disc segmentation. The mean LLE was 2.920 ± 1.460 mm for CBCT landmark localization and 3.520 ± 1.520 mm for MRI landmark localization. A qualitative review by a clinician suggested improved local anatomical alignment of the condyle and surrounding structures following mask-guided refinement.

CONCLUSIONS

The proposed multi-stage framework enables automated registration between CBCT and multi-planar TMJ MRI by combining landmark-based coarse alignment and segmentation-mask-guided refinement registration.

CLINICAL SIGNIFICANCE: The proposed framework enables the automated spatial integration of CBCT-derived osseous information and MRI-derived soft tissue information, potentially facilitating comprehensive TMJ assessment while reducing the manual effort required for multimodal image alignment.