(39) Automatic Impacted Tooth Segmentation Using Artificial Intelligence Model on Cone-beam Computed Tomography Images

Abstract: Statement of the problem: Tooth impaction is a common pathological dental condition. The etiology behind tooth impaction can be systemic, local, or genetic. Some other reasons for impaction are lack of space in the arch and hinderance in the eruption path of the tooth. Different pathologies associated with impacted teeth are pericoronitis, orofacial pain, TMJ disorders, pathological fractures, cysts, and neoplasms. Considering the complexity of the location and pathology associated with impacted teeth, Cone Beam Computed Tomography (CBCT) is considered as imaging of choice for detection of impacted tooth and its relations to surrounding structures. In addition to this, with the idea of reducing the errors in subjective evaluations related to the person, artificial intelligence applications can play a key role in detecting impacted teeth on advanced imaging such as CBCT. The aim of the study is to assess the success of artificial intelligence (AI) model for the automatic impacted tooth segmentation on Cone-Beam Computed Tomography (CBCT) images.

Materials and

Methods: Deep learning based Convolutional Neural Network (CNN) model was developed to segmentation of impacted tooth on CBCT images. 4060 axial slice images (4605 impacted tooth labels) obtained from 300 CBCT images. Labeling was made using CranioCatch annotation software (CranioCatch, Eskişehir-Türkiye) for impacted tooth. The dataset was randomly divided into train (3249 axial slice images, 3618 label), validation (406 axial slice images, 462 label), and test (405 axial slice images, 525 label). The Mask R-CNN ResNet 101 model were used for algorithm development.

Methods of data analysis: In this retrospective study, 300 CBCT images were used obtained from archive of Eskisehir Osmangazi University Faculty of Dentistry Department of Oral and Maxillofacial Radiology between 2016 and 2022 years. The success of the AI model was calculated with the confusion matrix. Also, Receiver Operating Characteristic (ROC) and Area Under Curve (AUC) analysis were made for detail analysis.

Results: In evaluating the performance of the AI model to segmentation of impacted tooth, True Positive (TP), False Positive (FP), False Negative (FN) values were found as 496, 15, 29, respectively. Sensitivity, precision, and F1 score values were calculated as 0.9447, 0.9706, 0.9575, respectively. Area Under Curve (AUC) value is calculated as 0.8682.

Outcomes Data: Mask R-CNN ResNet 101 model has promising results for automatic segmentation of impacted tooth on CBCT images.

Conclusions: AI can play a role as a decision support system for dentists in the diagnosis and treatment planning of impacted teeth on CBCT images. More powerful AI models can be obtained with studies to be carried out with large databases in the future.

References:
1. Orhan K, Bilgir E, Bayrakdar IS, Ezhov M, Gusarev M, Shumilov E. Evaluation of artificial intelligence for detecting impacted third molars on cone-beam computed tomography scans. Journal of Stomatology, Oral and Maxillofacial Surgery. 2021 Sep 1;122(4):333-7.

2. Başaran M, Çelik Ö, Bayrakdar IS, Bilgir E, Orhan K, Odabaş A, Aslan AF, Jagtap R. Diagnostic charting of panoramic radiography using deep-learning artificial intelligence system. Oral Radiol. 2022 Jul;38(3):363-369.

Figure 1. Actual and predictive images for impacted tooth segmentation (True positive-TP)
Figure 2. Actual and predictive images for impacted tooth segmentation (False positive-FP)
Figure 3. Actual and estimated images for impacted tooth segmentation (False negative-FN)