Efficient 3D Mandibular Landmark Detection Via MVSH-CNN

Accurate localization of anatomical landmarks on the mandible is critical for orthodontic diagnosis and maxillofacial surgical planning. This study proposes a novel and efficient automated approach, MVSH-CNN, leveraging a multi-view stacked hourglass convolutional neural network to enhance precision and efficiency in 3D mandibular landmark detection. A total of 100 mandibular surface models were reconstructed from cone beam computed tomography (CBCT) scans of adult Han Chinese individuals (80 normal, 20 with mandibular asymmetry). The MVSH-CNN model was trained using this dataset and evaluated on an independent test set of 40 cases (20 normal, 20 asymmetric). Nineteen anatomical landmarks were annotated. Model performance was compared with the MeshMonk registration-based method. Accuracy was assessed via mean localization error, and efficiency was measured by processing time. All procedures were approved by the institutional ethics committee. MVSH-CNN achieved an average localization error of 1.13 ± 0.85 mm, showing approximately 25% improvement in accuracy compared to MeshMonk. The average processing time per case was 19.65 seconds, significantly lower than the baseline (356 seconds). No statistically significant difference in localization accuracy was observed between normal and asymmetric mandibles (p > 0.05). MVSH-CNN demonstrates superior accuracy and efficiency in 3D mandibular landmark localization. This method offers a reliable and automated solution for clinical applications, supporting intelligent workflows in orthodontics and maxillofacial surgery.