Approach for the Detection of Early Carious Lesions Based on Intraoral Photographs Using YOLOv7 and Faster R-CNN

Abstract

Tooth decay is a global challenge due to lack of dental care and excessive sugar consumption. These generate expensive dental treatments and affect quality of life, self-esteem, and productivity. Due to this, an approach is proposed for the detection of carious pre-lesions through dental image processing and using 2 Deep learning architectures most used in the literature: YOLOv7 and Faster RCNN. The approach is developed in 4 phases: (i) acquisition of the dataset, (ii) development of architectures, (iii) performance evaluation and (iv) analysis of results. Both architectures focus on the use of a public dataset composed of a total of 9,327 images of Intraoral Photographs classified into 3 classes: “teeth with cavities” (0), “teeth without cavities” (1) and “teeth with amalgam” (2). A web system was built with the model that had the best performance. The results showed that the YOLOv7 architecture had better performance than Faster R-CNN, obtaining an average accuracy of 95.7% in the detection of teeth “without cavities,” “with cavities” and “with amalgam.”.