Artificial intelligence-assisted identification and assessment of mandibular asymmetry on panoramic radiography.

Introduction: Mandibular symmetry is crucial in orthodontic diagnosis and treatment planning. This study aimed to establish an artificial intelligence (AI) method to automatically and accurately identify mandibular landmarks and assess asymmetry via orthopantomography (OPG) radiographs.

Methods: A total of 1038 OPG radiographs (451 mixed and 587 permanent dentitions) were collected and annotated to develop the AI model for identifying mandibular landmarks. First, the mesiodistal widths of the bilateral mandibular first molars were compared to categorize images as horizontally distorted or nondistorted. Next, the efficacy and robustness of the model were assessed through landmark identification, measurement, and asymmetry diagnostics accuracy using successful detection rates and interclass correlation coefficient.

Results: The AI model achieved an average landmark detection error of 0.86 ± 0.95 mm, with 0.97 ± 0.99 mm for bony landmarks and 0.54 ± 0.84 mm for dental landmarks. The successful detection rates at 1, 2, and 3 mm were 75.33%, 93.11%, and 96.72%, respectively. The accuracy exhibits region-specific variations: vertical errors were larger in condylar landmarks, whereas horizontal errors were more pronounced in the mandibular gonial angle (P <0.05). The AI and manual methods show high consistency in all measurements (interclass correlation coefficient >0.983). Condyle landmarks were more accurate in permanent dentition, whereas mandibular angle landmarks were more precise in mixed dentition (P <0.05). Furthermore, the model achieved 82.52% and 75.24% diagnostic accuracy when using gonial angle and total ramal height.

Conclusions: The AI model accurately identifies anatomic landmarks and assesses mandibular asymmetry in OPG radiographs, demonstrating generalizability and robustness across different dentitions and showcasing potential as a promising diagnostic tool in clinical practice.