Prognostic Evaluation of Lower Third Molar Eruption Status from Panoramic Radiographs Using Artificial...

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Prophylactic third molar extraction decisions are challenging due to conflicting guidelines and reliance on surgeon experience.
• Predicting the eruption status of impacted third molars is crucial for effective clinical management.

Purpose of the Study:

• To evaluate deep learning-based artificial intelligence (AI) for predicting the eruption status of impacted third molars.
• To develop a predictive model for final tooth positions to assist early clinical decision-making.

Main Methods:

• A retrospective study utilized 1102 panoramic radiographs annotated for eruption status (initial vs. definitive).
• Two deep learning architectures (InceptionV3, ResNet50) were trained and evaluated using hyperparameter tuning, model evaluation, and preprocessing assessment.
• Performance metrics included accuracy, recall, precision, and F1 score. Classical machine learning algorithms were also applied.

Main Results:

• The ResNet50 model with preprocessing achieved the highest performance (F1 score: 0.829).
• Model performance was better for definitive eruption cases (F1 score: 0.829) than initial cases (F1 score: 0.705).
• Clinical prediction accuracy was 83% for full eruption and 75% for impaction, with lower accuracy for partial impactions.

Conclusions:

• AI, particularly deep learning models like ResNet50, shows potential in supporting early prediction of third molar eruption status.
• These AI tools can enhance clinical decision-making regarding impacted third molars.
• Further development with larger datasets and optimization may lead to greater accuracy for routine clinical applications.