The prophylactic extraction of third molars is highly dependent on the surgeon’s experience as the common practices and guidelines contradict. The purpose of this study was to evaluate the eruption status of impacted third molars using deep learning-based artificial intelligence (AI) and to develop a model that predicts their final positions at an early stage to aid clinical decisions. In this retrospective study, 1102 panoramic radiographs (PANs) were annotated by three expert dentists to classify eruption status as either initial or definitive. A dataset was created and two deep learning architectures, InceptionV3 and ResNet50, were tested through a three-phase protocol: hyperparameter tuning, model evaluation, and assessment of preprocessing effects. Accuracy, recall, precision, and F1 score were used as performance metrics. Classical machine learning (ML) algorithms (SVM, KNN, and logistic regression) were also applied to features extracted from the deep models. ResNet50 with preprocessing achieved the best performance (F1 score: 0.829). Models performed better with definitive cases than with initial ones, where performance dropped (F1 score: 0.705). Clinically, the model predicted full eruption or impaction with 83% and 75% accuracy, respectively, but showed lower accuracy for partial impactions. These results suggest that AI can support early prediction of third molar eruption status and enhance clinical decision-making. Deep learning models (particularly ResNet50) demonstrated promising results in predicting third molar eruption outcomes. With larger datasets and improved optimization, AI tools may achieve greater accuracy and support routine clinical applications.
Guldiken et al. (Wed,) studied this question.