Towards fully automated third molar development staging in panoramic radiographs.

Area of Science:

• Forensic Odontology
• Radiology
• Artificial Intelligence

Background:

• Third molar development staging is crucial for age assessment in sub-adults.
• Current manual methods are time-consuming and subject to operator variability.

Purpose of the Study:

• To develop a fully automated deep learning system for third molar developmental staging.
• To enhance the accuracy and efficiency of age assessment using panoramic radiographs.

Main Methods:

• Utilized convolutional neural networks (CNNs) for automated localization, segmentation, and staging of third molars.
• Employed transfer learning and data augmentation on a dataset of 400 panoramic radiographs (OPGs).
• Validated the three-step automated workflow using fivefold cross-validation.

Main Results:

• Automated localization achieved an average Euclidean distance of 63 pixels.
• Third molar segmentation yielded an average Dice score of 93%.
• Developmental stage classification achieved 54% accuracy, 0.69 mean absolute error, and 0.79 kappa coefficient.
• The automated workflow computed in an average of 2.72 seconds per OPG.

Conclusions:

• The fully automated deep learning approach shows promising results for third molar staging, offering significant speed advantages over manual methods.
• Despite a limited dataset, the proposed method demonstrates potential for reliable age assessment in forensic and clinical contexts.