Deep learning algorithms for detecting fractured instruments in root canals

The primary aim of our study is to compare the performance of different deep learning algorithms in detecting broken instruments on periapical radiographs taken at the Department of Endodontics, Faculty of Dentistry, Trakya University. Periapical radiographs taken for any reason between January 2022 and December 2023 were screened, and after applying exclusion criteria, 360 radiographs containing broken instruments were included in the study. To ensure a meaningful comparison, radiographs of 340 teeth with root canal treatment that did not contain broken instruments but met the inclusion criteria were also included in the study. In the CVAT program, broken instruments were labeled in blue, and root canal treatments were labeled in red. Training, validation, and test sets were created. Confusion matrices were generated for each model based on the results obtained from the test sets. Accuracy, F1 score, Matthew's correlation coefficient (MCC), and ROC-AUC values were calculated from the confusion matrices, and the performance of the models was compared. DenseNet, ResNet, and WideResNet models showed the highest success rates, while the MaxViT model was the least successful among the eight models compared. This study, which uses deep learning models for detecting broken instruments, has yielded promising results. However, more successful performances can be achieved by addressing the shortcomings in the study design.