Automatic Detection of Mandibular Canal in Cone-Beam Computed Tomography Images Using a Deep Convolutional Neural Network

Abstract

Objectives: Deep learning techniques have been successfully applied in numerous domains, including medical diagnostic imaging, proving their viability. Research into automatic diagnostic and prediction techniques is ongoing in response to the rising demand in the healthcare sector. The automatic detection of the lower dental canal has drawn a lot of attention, particularly in dentistry for a variety of reasons. To prevent nerve damage during surgery, it is crucial to consider where the inferior alveolar nerve (IAN), one of the major components in the lower dentition, is located. This study looked into how a deep convolutional neural network might automatically segment the lower dental canal in CBCT images. Materials and Methods: Deep convolutional neural networks were used to segment the lower dental canal. Due to the high expenses of expert labeling, the difficulties in gaining patients' informed consent, and the need to use the target network in all CBCT devices—a factor that had not been taken into account in earlier studies—the Multi-shot segmentation method was employed for this purpose. This technique makes use of the ResNet deep neural network, a particular kind of convolutional network. In addition to ResNet, there are two sub-networks in this network. Predetermined tasks were used to train the main network, and unlabeled photos were fed into the everything segmentation model to extract the image's features. The main network was trained using ten label samples. In order to ensure that the learning process can be properly carried out due to the small number of samples, over-fitting of the network was also avoided by using the analysis of singular values. Self-supervised learning was also employed as a learning strategy. Results: The experiments' outcomes were assessed numerically and qualitatively. The IoU criterion was applied in the event of a quantitative comparison, and the result was 0.6331. Better outcomes were attained when this procedure was compared to the basic fine tuning techniques. The performance of few-shot lower dental canal detection in CBCT images was greatly enhanced by the proposed approach. Results among related investigations were greatly enhanced. Conclusion: The automatic canal recognition system developed through deep learning will considerably aid in the planning of effective treatment and lessen patient discomfort, according to the findings of the studies.