Automatic Annotation of Cephalometric Landmarks using a Two-Stage Multi-Regional Context-Enhanced Framework in Cone-Beam Computed Tomography Volumes.

Objectives

In this retrospective study, we aimed to develop an incomplete-observation-aware framework for automatic three-dimensional cephalometric landmark localization in cone-beam computed tomography (CBCT) volumes, capable of explicitly modelling landmark visibility, reconstructing missing anatomical context, and effectively leveraging both fully and partially annotated clinical data.

Methods

An incomplete-observation-aware framework was developed for CBCT cephalometric landmark localization. The framework explicitly models landmark visibility under limited field-of-view conditions and incorporates latent anatomical completion within a two-stage, coarse-to-fine localization strategy. A total of 144 CBCT scans obtained for routine orthodontic diagnosis and treatment planning, including 90 fully annotated and 54 partially annotated, were retrospectively collected, with 115 cases for training and 29 for testing. Performance was evaluated using the mean radial error (MRE) and successful detection rate (SDR).

Results

The proposed framework achieved an MRE of 1.01 mm ± 0.70 mm on the complete-landmark test set and 1.01 mm ± 0.72 mm on the mixed-landmark test set. SDRs exceeded 91% within 2 mm and reached 100% within 6 mm for both datasets. The comparable performance observed across the complete and mixed test sets demonstrates the robustness of the proposed framework in handling incomplete observations and partially annotated data.

Conclusions

The proposed framework achieved robust and accurate performance for automatic three-dimensional (3D) cephalometric landmark localization on CBCT volumes by modelling landmark visibility and incorporating latent anatomical completion, enabling reliable localization under incomplete observations. CLINICAL RELEVANCE The proposed framework enables automated 3D cephalometric landmark localization in CBCT volumes and is robust to incomplete field-of-view and missing landmarks through visibility-aware modelling and latent anatomical reconstruction.