Multi-view deep learning for mandibular landmark localization

Accurate localization of anatomical landmarks on the mandible is crucial for maxillofacial surgery and orthodontic treatment planning. This study aims to develop and validate a novel multi-view deep learning framework to enhance the accuracy and efficiency of landmark localization on CBCT-derived 3D mandibular surface models.

We propose a multi-view stacked hourglass convolutional neural network (MVSH-CNN) that localizes 19 anatomical landmarks on 3D mandibular surface models reconstructed from cone beam computed tomography (CBCT) scans. A total of 140 mandibular scans from adult Han Chinese individuals were used, with 100 cases for training/validation and 40 cases (20 normal, 20 asymmetry) for independent testing. Manual annotations served as the reference standard. Localization performance was compared with the MeshMonk non-rigid registration method using Euclidean distance error and computational time.

MVSH-CNN achieved a mean localization error of 1.13 ± 0.85 mm in the normal group and 1.10 ± 0.79 mm in the asymmetry group, significantly outperforming MeshMonk (1.42 ± 1.28 mm and 1.43 ± 1.14 mm, respectively; P < 0.05). Processing time per mandible was reduced from 356 s to 19.65 s. Over 94% of landmarks localized by MVSH-CNN had an error < 2 mm, meeting the predefined clinical threshold.

The MVSH-CNN framework provides accurate, robust, and time-efficient semi-automated 3D landmark localization directly on STL-based mandibular models, outperforming conventional registration-based approaches.

MVSH-CNN offers a semi-automated and clinically viable solution for digital orthodontic assessment, virtual surgical planning, and intelligent craniofacial analysis, significantly reducing manual workload while enhancing reproducibility and standardization.