Dental Caries Classification with Deep CNN on X-Ray Images

In the dental world, early detection of different oral lesions is vital to carry out an accurate and precise treatment to care for the patient's health. However, it is not always possible for dentists to make an early diagnosis due to the limited number of symptoms or lack of experience. In this context, we propose developing a classification method based on deep CNN architectures that maximize the caries classification performance on X-ray images while minimizing the model's complexity. We built and explored three architectures named DCNN1, DCNN2, and DCNN3, which were trained and validated in an experimental dataset formed by 1064 ROI samples with and without caries lesions. The DCNN3 model obtained the highest mean AUC score of 0.878 with 100 epochs during the validation in the training phase. However, the best-selected model was composed of DCNN1 with 10 epochs. It reached a mean AUC score of 0.874, which did not represent a statistical difference compared to the DCNN3 model while having the lowest complexity. Regarding the validation of the selected DCNN1 model in the test set made up of 106 samples, it obtained a mean AUC score of 0.830. The slight performance difference between the validation in the training and test stages demonstrated the learning and generalizability of the proposed model, which is reasonable for the classification of caries lesions. However, any performance improvement depends on a larger dataset, as the behavior of the loss function suggests.