Evaluation of mandibular asymmetry using artificial intelligence algorithms developed with deep learning on orthodontic photographs

Aim: This study aims to enable the automatic detection of mandibular asymmetry through the development of a deep learning–based artificial intelligence algorithm using orthodontic photographs. The proposed system is intended to standardize the diagnostic process by reducing observer-dependent variability. Material and Method: A total of 2598 standardized resting frontal facial photographs were used. The data were divided into training (80%), validation (10%), and test (10%) sets. The model was trained for 300 epochs with an input size of 640 pixels and a batch size of 16. Data augmentation techniques were applied to address class imbalance. Model performance was evaluated using accuracy, precision, recall, and F1 score. Results: The overall accuracy of the model was obtained as 70.0%. In the "No Asymmetry" class, the model demonstrated high diagnostic performance, with an F1 score of 0.82, a precision of 0.71, and a recall of 0.96. In contrast, in the "Asymmetry Present" class, recall remained at 0.10 and the F1 score was calculated as 0.17. These findings indicate that the model largely missed the samples belonging to the asymmetrypresent class and showed limited performance in identifying asymmetric cases. Conclusion: The findings indicate that deep learning models are applicable for detecting mandibular asymmetry using resting frontal facial photographs. However, a direct classification-based approach may be insufficient for capturing the subtle geometric differences that characterize asymmetry. Future models enriched with landmark-based measurements and more extensive augmentation strategies may improve performance.