A critical appraisal of computer vision in orthodontics

Objective To evaluate the precision of a computer vision (CV) and augmented reality (AR) pipeline for orthodontic applications, specifically in direct bonding and temporary anchorage device (TAD) placement, by quantifying system accuracy in six degrees of freedom (6DOF) pose estimation. Methods A custom keypoint detection model (YOLOv8n-pose) was trained using over 1.5 million synthetic images and a supplemental manually annotated dataset. Thirty anatomical landmarks were defined across maxillary and mandibular arches to maximize geometric reliability and visual detectability. The system was deployed on a Microsoft HoloLens 2 headset and tested using a fixed typodont setup at 55 cm. Pose estimation was performed in “camera space” using Perspective-n-Point (PnP) methods and transformed into “world space” via AR spatial tracking. Thirty-four poses were collected and analyzed. Errors in planar and depth estimation were modeled and experimentally measured. Results Rotational precision remained below 1°, while planar pose precision was sub-millimetric (X: 0.46 mm, Y: 0.30 mm), except for depth (Z), which showed a standard deviation of 5.01 mm. These findings aligned with theoretical predictions based on stereo vision and time-of-flight sensor limitations. Integration of headset and object pose led to increased Y-axis variability, possibly due to compounded spatial tracking error. Sub-pixel accuracy of keypoint detection was achieved, confirming high performance of the trained detector. Conclusion The proposed CV-AR system demonstrated high precision in planar pose estimation, enabling potential use in clinical orthodontics for tasks such as TAD placement and bracket positioning. Depth estimation remains the primary limitation, suggesting the need for sensor fusion or multi-angle views. The system supports real-time deployment on mobile platforms and serves as a foundational tool for further clinical validation and AR-guided procedures in dentistry.