A Deep Learning Approach for Classification of Dentinal Tubule Occlusions

Abstract

This study aimed to develop a novel deep learning model for reliable quantification of dentinal tubule occlusions instead of manual assessment techniques, and the performance of the model was compared to other methods in the literature. Ninety-six dentin samples were cut and prepared with desensitizing agents to occlude dentinal tubules on different levels. After obtaining images via scanning electron microscope (SEM), 2793 single dentinal tubule images with 48 × 48 resolution were segmented and labeled. Data augmentation techniques were applied for improvement in the learning rate. The augmented data having a total of 10700 images belonging to five classes were used as the network training dataset. The proposed convolutional neural network (CNN) is a class of deep learning model and was able to classify the degree of dentinal tubule occlusions into five classes with an overall accuracy rate of 90.24%. This paper primarily focuses on developing a CNN architecture for detecting the level of dentin tubule occlusions imaged by SEM. The results showed that the proposed CNN architecture is an immensely successful alternative and allowed for objective and automatic classification of segmented dentinal tubule images.