Quantitative level determination of fixed restorations on panoramic radiographs using deep learning

Although many studies in various fields employ deep learning models, only a few such studies exist in dental imaging. The present article aims to evaluate the effectiveness of convolutional neural network (CNN) algorithms for the detection and diagnosis of the quantitative level of dental restorations using panoramic radiographs by preparing a novel dataset.

20,973 panoramic radiographs were used, all labeled into five distinct categories by three dental experts. AlexNet, VGG-16, and variants of ResNet models were trained with the dataset and evaluated for the classification task. Additionally, 10-fold cross-validation (ie, 9 folds were separated for training and 1 fold for validation) and data augmentation were carried out for all experiments.

The most successful result was shown by ResNet-101, with an accuracy of 92.7%. Its macro-average AUC was also the highest, at 0.989. Other accuracy results obtained for the dataset were 75.5% for AlexNet, 85.0% for VGG-16, 92.1% for ResNet-18, 91.7% for ResNet-50, and 92.1% for InceptionResNet-v2.

An accuracy of 92.7% is a very promising result for a computer-aided diagnostic system. This result proved that the system could assist dentists in providing supportive preliminary information from the moment a patient's first panoramic radiograph is taken. Furthermore, as the introduced dataset is powerful enough, it can be relabeled for different problems and used in different studies.