A Customized Convolutional Neural Network for Dental Bitewing Images Segmentation

Abstract

Bitewing images are useful for recognizing the most common dental diseases, like tooth decay and periodontal bone loss. Besides providing important details like the condition of fillings and the presence of calculus or tartar. Due to the wide variety of topologies, the complexity of medical structures, and the poor image quality caused by problems like low contrast, noise, irregularities, and fuzzy edges borders, segmentation of dental images is difficult and often unsuccessful. Recent advances in deep learning models improve the performance of analyzing dental images. In this study, we build a customized Convolutional neural network (CNN) to segment the bitewing image. The bitewing radiographs, which will be used as input to the CNN model, are imported into MATLAB where the image is first enhanced before being segmented to create a binary mask image that excludes the background from the original images. Those masks are used as a target for the deep learning model. By training the proposed system with 456 bitewing images, the best accuracy we achieved on unseen images is 97.3% of accuracy, 88.27% of Fl-score.