Deep Learning-Based Evaluation of Impacted Third Molars with Open and Closed Apices on CBCT.
TL;DR
This study demonstrates the feasibility of a deep learning approach for automatically evaluating impacted third molars on CBCT, including tooth segmentation and root maturity assessment and suggests that such AI tools could assist clinicians in assessing impactions and planning extractions.
Abstract
Objective
To develop and evaluate a deep learning approach for the automated segmentation of impacted third molars on cone-beam computed tomography (CBCT) images and the classification of each tooth's root apex as open or closed. STUDY
Design
An observational study. Place and Duration of the Study: Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Marmara University, Istanbul, Turkiye, from December 2022 to December 2024. METHODOLOGY Three hundred CBCT scans containing impacted third molars were retrospectively collected. Experts categorised teeth as having an open apex (incomplete root development) or closed apex (fully formed roots). A 3D nnU-Net convolutional neural network was trained using manual segmentations of the molars (270 scans for training and 30 for testing), with the network output distinguishing between teeth with open and closed apices. Model performance was evaluated on the independent test set using segmentation overlap metrics (Dice similarity coefficient and Jaccard index) and classification metrics derived from confusion matrix components [accuracy, sensitivity, precision, and area under the ROC curve (AUC)]. All performance metrics were automatically computed within the nnU-Net v2 framework and further verified using Python.
Results
The model demonstrated high performance in segmenting closed-apex molars, achieving a Dice similarity coefficient of 0.85, a Jaccard index of 0.78, sensitivity of 0.90, precision of 0.85, and an AUC of 0.95. However, for open-apex molars, performance was substantially lower (Dice: 0.37; Jaccard: 0.30; sensitivity: 0.31; precision: 0.71; AUC: 0.65), primarily attributable to severe class imbalance in the training dataset.
Conclusion
This study demonstrates the feasibility of a deep learning approach for automatically evaluating impacted third molars on CBCT, including tooth segmentation and root maturity assessment. The model's strong performance suggests that such AI tools could assist clinicians in assessing impactions and planning extractions. KEY WORDS Impacted third molar, Cone-beam computed tomography, Deep learning, Tooth segmentation, Root apex closure, nnU-Net.