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

Objective: The aim of this study is to detect the C-shaped canal formation in mandibular second molars on panoramic radiographs trained with the aid of conebeam computed tomography (CBCT) using different deep learning algorithms. Method: Our study included images of 592 patients with at least one of their 37th or 47th teeth present in the mouth, archived in the Department of Oral, Dental, and Maxillofacial Radiology, Faculty of Dentistry, Pamukkale University. These images consisted of digital panoramic radiographs and cone-beam computed tomography (CBCT) scans. The dataset included 289 panoramic radiographs with C-shaped canals and 303 panoramic radiographs without C-shaped canals. From the 289 panoramic radiographs with C-shaped canals, a total of 422 teeth (both unilateral and bilateral) were labeled, and an equal number of 422 teeth from the 303 panoramic radiographs without C-shaped canals were labeled. This resulted in a dataset of 844 labeled panoramic radiographs for teeth numbered 37 and 47, with and without C-shaped canals. CBCT images were considered the gold standard for confirming the presence of a C-shaped canal. To detect C-shaped canals in the 844 panoramic images comprising our dataset, the detection accuracy performance of 11 different deep learning models (InceptionV3, VGG19, Xception, ResNet152V2, EfficientNetB1, EfficientNetB0, VGG16, ResNet50, ResNet50V2, DenseNet169, and DenseNet201) was investigated. The models were applied to original and preprocessed panoramic images of teeth numbered 37 and 47, divided into two separate groups: crown-root and root-only. Additionally, the performance metrics were evaluated using confusion matrices based on random majority images. Results: For the crown-root dataset, the highest average accuracy values for original and preprocessed images were found to be 0.885 (88.5%) and 0.886 (88.6%), respectively. For the root-only dataset, the highest average accuracy values for original and preprocessed images were 0.892 (89.2%) and 0.887 (88.7%), respectively. The highest accuracy performance metrics for random majority images in the crown-root and root-only datasets were 0.902 (90.2%) and 0.897 (89.7%), respectively. When preprocessing was applied to the crown-root dataset, the average accuracy values increased for the deep learning architectures InceptionV3, VGG19, Xception, and ResNet152V2, while they decreased for EfficientNetB1, EfficientNetB0, VGG16, ResNet50, ResNet50V2, DenseNet169, and DenseNet201. For the root-only dataset, preprocessing increased the average accuracy for EfficientNetB0, ResNet50, ResNet50V2, Xception, and ResNet152V2, whereas it decreased for InceptionV3, EfficientNetB1, VGG16, VGG19, DenseNet169, and DenseNet201. Conclusion: High-performance values were achieved through the combined use of deep learning architectures. Our study is significant for the detection of C-shaped canals in terms of the success of endodontic treatments, and deep learning models are sufficiently capable of assisting clinicians. Keywords: C-shaped tooth root canal, endodontic treatment, deep learning, preprocessed images