Development of Intraoral Clinical Image Dataset for Deep Learning Caries Detection

A substantial percentage of people around the world suffer from dental caries, which is one of the chronic oral diseases that can affect anyone at any time. Children can get dental caries, and nearly all adults will eventually experience it. Untreated dental decay may be a sign of cardiovascular problems and cognitive decline if it is left untreated. The conventional examination of caries conducted by dentists has been deemed inefficient and time-consuming. Thus, this study introduces a novel dataset of dental clinical images designed for the purpose of caries classification using deep learning techniques. The dataset consists of two classes: caries and caries-free. To assess the effectiveness of a deep learning model, its performance on the dataset needs to be evaluated. The ResNet-50 architecture is employed and trained on the dataset with three different learning rates: $10^{-4}, 10^{-5}$, and $10^{-6}$. The training results indicate that the model trained with a learning rate of $10^{-4}$ achieved the highest accuracy of 0.7952 and a loss value of 0.4225. However, there is a notable discrepancy between the training and validation accuracy, suggesting difficulties in the model’s ability to learn from the dataset due to its significant heterogeneity. Conversely, the model trained with a learning rate of 10-6 exhibits the most favorable learning curve. This suggests that employing a lower learning rate yields more desirable learning curves. Additionally, the testing process’s confusion matrices reveal the impact of loss values obtained during training on the testing results.