Investigate a CNN Pre-Trained Model for Classifying the Stages of Tooth Cavity Disease

Abstract

Dental cavity diseases are very common diseases and half of the world population suffers from it. Dental caries has been a common health issue throughout the world, which can even lead to dental pulp and root apical inflammation eventually and the most frequent dental health issue in the general population. Therefore, timely and effective treatment of dental caries is vital for patients to reduce pain. To this end, Deep learning techniques have demonstrated remarkable diagnostic capabilities within the radiology field. The aim of this study is investigate a CNN pre-trained model for classifying the stages of tooth cavity diseases and additionally, we sought to compare the classification results achieved by deep learning models with those of expert dentists. Methods: In this study, two dental experts participated in preparing and evaluating collected 4725 dental X-ray images. The combination of these images formed our reference dataset. From this dataset, we established training and validation set consisting of 4269 images, as well as a separate test set comprising 456 images. To achieve our objectives, we employed a convolutional neural network and utilized two pre-trained models, VGG16 and InceptionResNetV2. All These models were developed to detect the stages of the cavity and classify them according to their labels: Enamel, Dentin, Pulp, and Healthy tooth. Results: Based on trained models, the CNN model has achieved a Base accuracy of 0.899% and a validation rate of 0.910%, VGG16 scored an accuracy of 0.9243% and a validation rate of 0.9557%, while InceptionResNetV2 scored an accuracy of 0.977% and a validation of 0.978%. The expert and the neural network demonstrated comparable results across the metrics (F1 score, precision, and recall). In terms of cavity stage classification, InceptionResNetV2 exhibited an impressive best accuracy of 97.7% than the other two models. Furthermore, the recall results for InceptionResNetV2 in the ED/DD/PD/HT stages were 0.96%, 1.00%, 0.94%, and 1.00%, respectively. Conclusions: This paper concludes that the deep learning methods we implemented are comparable performance to experts in determining tooth cavity stages using dental panoramic radiographs. The application of these techniques could have significant implications in dental diagnostics to determine the stages of cavities in the right way.