Numbering teeth in panoramic images: A novel method based on deep learning and heuristic algorithm

Abstract

Dental problems are one of the most common health problems for people. To detect and analyze these problems, dentists often use panoramic radiographs that show the entire mouth and have low radiation exposure and exposure time. Analyzing these radiographs is a lengthy and tedious process. Recent studies have ensured dental radiologists can perform the analyses faster with various artificial intelligence sup-ports. In this study, the numbering performance of Mask R-CNN and our heuristic algorithm-based method was verified on panoramic dental radiographs according to the Federation Dentaire Internationale (FDI) system. Ground-truth labelling of images required for training the deep learning algorithm was performed by two dental radiologists using the web-based labelling software DentiAssist created by the first author. The dataset was created from 2702 anonymized panoramic radio-graphs. The dataset is divided into 1747, 484, and 471 images, which serve as training, validation, and test sets. The dataset was validated using the k-fold cross-validation method (k = 5). A three-step heuristic algorithm was developed to improve the Mask R-CNN segmentation and numbering results. As far as we know, our study is the first in the literature to use a heuristic method in addition to traditional deep learning algorithms in detection, segmentation and numbering studies in panoramic radiography. The experimental results show that the mAp (@IOU = 0.5), precision, recall and f1 scores are 92.49%, 96.08%, 95.65% and 95.87%, respectively. The results of the learning-based algorithm were improved by more than 4%. In our research, we discovered that heuristic algorithms could improve the accuracy of deep learning-based algorithms. Our research will significantly reduce dental radiologists' workload, speed up diagnostic processes, and improve the accuracy of deep learning systems.(c) 2022 Karabuk University. Publishing services by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).