Evaluation of the accuracy of detecting C-shaped canals in mandibular second molars identified by cone-beam computed tomography on panoramic radiographs using artificial intelligence algorithms developed with deep learning methods.

objectiveThe aim of this study is to detect the C-shaped canal formation in mandibular second molars on panoramic radiographs based on different Deep Convolutional Neural Networks (DCNNs) trained using panoramic radiographs.

methodThis study includes images of 592 patients, consisting of digital panoramic radiographs and cone-beam computed tomography (CBCT) scans of patients with at least one mandibular second molar, archived in the Department of Oral, Dental, and Maxillofacial Radiology, Faculty of Dentistry, Pamukkale University. To confirm the presence of a C-shaped canal, CBCT images were analyzed and set as the gold standard. From 289 panoramic radiographs with C-shaped canals, a total of 422 mandibular second molars were labeled, and an equal number of 422 mandibular second molars were labeled from 303 panoramic radiographs without C-shaped canals, resulting in a total dataset of 844 labeled panoramic radiographs. To detect C-shaped canals in the 844 panoramic images comprising our dataset, the detection accuracy performance of 11 different deep learning models was investigated. These models were applied to the preprocessed and non-preprocessed panoramic images of mandibular second molars divided into two separate groups as "crown-root" and "root". For the first time in the literature, to the best of our knowledge, model prediction results were fused using majority voting for the detection of C-shaped canals in mandibular second molars. Then, corresponding performance measurements were evaluated in terms of accuracy, precision, recall, specificity and confusion matrices.

resultsFor the crown-root dataset, the highest average accuracy metrics for preprocessed and non-preprocessed images were obtained as 0.886 (88.6%) and 0.885 (88.5%), respectively. For the root dataset, the highest average accuracy values for preprocessed and non-preprocessed images were 0.887 (88.7%) and 0.892 (89.2%), respectively. The highest accuracy performance metrics, on the other hand, obtained by the fusion of different DCNNs decisions with the application of majority voting, yielded as 0.902 (90.2%) and 0.897 (89.7%) for crown-root and root dataset groups, respectively.

conclusionHigh-performance values were achieved through the use of combined deep learning architectures. Obtained results show that the proposed method is significant for the detection of C-shaped canals in terms of the success of endodontic treatments, and use of deep learning models are sufficiently capable of assisting clinicians.