Multi-view deep learning for mandibular landmark localization.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Radiology

Background:

• Accurate localization of anatomical landmarks on the mandible is essential for effective maxillofacial surgery and orthodontic treatment planning.
• Current methods for landmark identification on 3D models can be time-consuming and lack precision.

Purpose of the Study:

• To develop and validate a novel multi-view deep learning framework for enhanced accuracy and efficiency in landmark localization on 3D mandibular surface models derived from CBCT scans.

Main Methods:

• A multi-view stacked hourglass convolutional neural network (MVSH-CNN) was proposed to localize 19 anatomical landmarks on 3D mandibular models.
• The framework was trained and validated on 100 CBCT scans and tested on an independent set of 40 scans (20 normal, 20 asymmetry).
• Performance was benchmarked against the MeshMonk non-rigid registration method using Euclidean distance error and computational time.

Main Results:

• MVSH-CNN achieved a mean localization error of 1.13 ± 0.85 mm (normal) and 1.10 ± 0.79 mm (asymmetry), significantly outperforming MeshMonk.
• Processing time per mandible was drastically reduced from 356 seconds to 19.65 seconds.
• Over 94% of landmarks localized by MVSH-CNN had an error < 2 mm, meeting clinical thresholds.

Conclusions:

• The MVSH-CNN framework offers accurate, robust, and time-efficient semi-automated 3D landmark localization on STL-based mandibular models.
• It surpasses conventional registration-based methods and provides a clinically viable solution for digital orthodontics, virtual surgical planning, and craniofacial analysis.