Evaluation of maxillary dental midline deviation using deep learning–based artificial intelligence algorithms on orthodontic frontal photographs

Objective: Accurate determination of maxillary dental midline deviation in orthodontic diagnosis is of great importance both functionally and aesthetically. The objective of this study was to automatically detect maxillary dental midline deviation on standardized frontal smile photographs using the deep learning-based You Only Look Once version 8 (YOLOv8) algorithm and to evaluate the performance of the obtained reference points in symmetric/asymmetric classification through various machine learning methods. Materials and Methods: In this study, frontal smile photographs of 72 individuals with maxillary dental midline deviation and 74 individuals without maxillary dental midline deviation in the permanent dentition period were examined. The photographs were annotated with the YOLOv8 algorithm according to the points glabella, subnasale, philtrum, maxillary dental midline, pogonion, and different data augmentation techniques were applied to increase data diversity. The obtained coordinates were divided into training, validation, and test groups; subsequently, classification analysis was performed using various machine learning methods. The performances of the models were evaluated by five-fold cross-validation, accuracy, precision, recall, F1 score, and area under the curve were used as performance measures. Results: As a result of the analyses, the highest classification performance was achieved with the Naive Bayes method (accuracy ≈ 75%; area under the curve ≈ 0.75), while other methods showed lower levels of performance. Linear discriminant analysis and decision tree methods achieved approximately 45% accuracy, while L1-penalized logistic regression achieved approximately 40% accuracy and was particularly insufficient in the asymmetry class. The findings demonstrated that the YOLOv8 algorithm can detect maxillary dental midline deviations with high accuracy and that, when used together with machine learning methods, it provides reliable classification results. Conclusion: This study revealed that artificial intelligence-based approaches have strong potential to be integrated into clinical diagnosis in orthodontics. The success of the YOLOv8 algorithm in automatic reference point detection and the superior performance of the Naive Bayes method in classification present a reliable and reproducible method for dental esthetic analysis. It was concluded that this approach may increase diagnostic accuracy and provide time efficiency by reducing observer-dependent variability in clinical practice. Keywords: Orthodontics, Dental Midline Deviation, Smile Esthetics, Artificial Intelligence, Deep Learning, Machine Learning