Semantic Segmentation of Dental Caries using Improved Deeplab V3Network

Abstract

Dental caries is a widespread health problem that can probably result to dental pulp as well as root apical infection. For patients to experience less pain, dental caries must be treated promptly and effectively. Conventional carious lesions diagnose techniques, such as naked-eye identification and panoramic radiograph evaluations, depends on qualified doctors, which can lead to misidentification and take a long time. This paper works on semantic segmentation using Improved Deeplab V3,a new deep learning design, to describe different carious lesions degrees from radiographic images. Semantic segmentation is analysis and classification of each pixel into multiple classes (labels). We begin by gathering a high-quality panoramic radiograph dataset of 2075 well-delineated caries, such as narrow, modest, and depth caries. The Deeplab V3network is then built with an additional channel and spatial attenuation module for segmenting these three types of caries from oral panoramic images. Hereafter, we make a comparison of improved DeeplabV3's semantic segmentation results with that of other previous methods. Experimental studies demonstrate that our method can segment different stages of caries with a mean Dice coefficient of 94.7% and an accuracy of 94.01 %.