Preinterventional Third-Molar Assessment Using Robust Machine Learning

Abstract

Machine learning (ML) models, especially deep neural networks, are increasingly being used for the analysis of medical images and as a supporting tool for clinical decision-making. In this study, we propose an artificial intelligence system to facilitate dental decision-making for the removal of mandibular third molars (M3M) based on 2-dimensional orthopantograms and the risk assessment of such a procedure. A total of 4,516 panoramic radiographic images collected at the Center of Dental Medicine at the University of Zurich, Switzerland, were used for training the ML model. After image preparation and preprocessing, a spatially dependent U-Net was employed to detect and retrieve the region of the M3M and inferior alveolar nerve (IAN). Image patches identified to contain a M3M were automatically processed by a deep neural network for the classification of M3M superimposition over the IAN (task 1) and M3M root development (task 2). A control evaluation set of 120 images, collected from a different data source than the training data and labeled by 5 dental practitioners, was leveraged to reliably evaluate model performance. By 10-fold cross-validation, we achieved accuracy values of 0.94 and 0.93 for the M3M-IAN superimposition task and the M3M root development task, respectively, and accuracies of 0.9 and 0.87 when evaluated on the control data set, using a ResNet-101 trained in a semisupervised fashion. Matthew's correlation coefficient values of 0.82 and 0.75 for task 1 and task 2, evaluated on the control data set, indicate robust generalization of our model. Depending on the different label combinations of task 1 and task 2, we propose a diagnostic table that suggests whether additional imaging via 3-dimensional cone beam tomography is advisable. Ultimately, computer-aided decision-making tools benefit clinical practice by enabling efficient and risk-reduced decision-making and by supporting less experienced practitioners before the surgical removal of the M3M.

Keywords: algorithms; deep learning; humans; mandible / diagnostic imaging; panoramic; radiography.

Figure 1.

Summary of the end-to-end pipeline. Clinical pipeline: depiction of the 3 main stages that include detection of the mandibular third molar (M3M), its characterization with respect to superimposition with inferior alveolar nerve (IAN) and root development, and the final clinical outcome to require or not require an additional diagnostic method. Machine learning pipeline: depiction of the available annotated and nonannotated data and its usage to train the machine learning models that will provide the necessary outcomes for the clinical pipeline. More precisely, the spatially dependent U-Net (SDU-Net) relies on orthopantogram (OPG) images and masks data (red connecting line) and outputs the location of the M3M; the ResNet-101 is first pretrained with nonannotated images (dark blue line) and then fine-tuned with OPG images and the class labels (light blue line).

Figure 2.

Therapy planning characterization and additional details on labelling procedure. (A) Matrix depicting the need for additional diagnostic intervention based on the combination of the potential outcomes from the 2 classification tasks. (B) Depiction of the class labels used in the annotation process of the mandibular third molar (M3M): the alveolar nerve superimposition task considers “no superimposition,” “superimposition <50%,” and “superimposition >50%,” whereas the root development task considers “complete root development,” “no root development,” and “uncertain root development.” (C) Distribution across label assignment for all tasks.

Figure 3.

Overall results of model performance for the mandibular third molar (M3M) detection task. (A) Violin plot of the 10-fold cross-validation results for all architectures on the training data set. Each point represents 1 iteration of the cross-validation for each model. (B) Confusion matrix of the best-performing model, the spatially dependent U-Net (SDU-Net) architecture evaluated on the external evaluation data set (out-of-distribution evaluation). (C) Table of all performance metrics (accuracy, F1-score, precision, recall, and Matthew’s correlation coefficient [MCC]) for all models evaluated on the out-of-distribution data. *Performances computed taking into consideration the samples where the M3M is shifted to mesial. (D) Example of a shifted M3M where the model did not recognize the existence of a M3M. (E) Examples of 4 successfully detected M3Ms using the SDU-Net.

Figure 4.

Overall results for the superimposition of the mandibular third molar (M3M) with the inferior alveolar nerve (IAN) and the M3M root development classification tasks. (A, F) Violin plots of the 10-fold cross-validation results for ResNet-101 and ViT-B, trained with supervised and semisupervised learning, and evaluated on the training and validation data sets. Each point represents 1 iteration of the cross-validation for the respective model. (B, G) Receiver operator characteristics (ROC) curves of the models evaluated on the external evaluation data set. (C, H) Confusion matrices for the ResNet-101 trained with semisupervision and evaluated on the external evaluation data set. (D, I) Table of all performance metrics (accuracy, F1-score, precision, recall, and Matthew’s correlation coefficient [MCC]) for all models evaluated in the external evaluation data set.