The aim of this study was to develop a deep learning model to automatically detect and segment unobturated mesial buccal 2 (MB2) canals on endodontically obturated maxillary molars depicted in CBCT studies. Fifty-seven deidentified CBCT studies of maxillary molars with clinically confirmed unobturated MB2 canals were retrieved from a dental institution radiology database. One-hundred and two maxillary molar roots with and without unobturated MB2 canals were segmented using ITK-SNAP. The data were split into training and testing samples designated to train and evaluate the performance, respectively, of a convolutional neural network (CNN), U-Net. The detection performance revealed a sensitivity of 0.8, a specificity of 1, a high PPV of 1, and a NPV of 0.83 for the testing set, along with an accuracy of 0.9. The segmentation performance of unobturated MB2 canals, assessed using the custom metric, rendered a mean value of 0.3018 for the testing set. The current AI algorithm has the potential to identify obturated and unobturated canals in endodontically treated teeth. However, the AI algorithm is still somewhat affected by metallic artifacts, variations in canal calcifications, and the applied configuration. Thus, further development is needed to improve the algorithm and validate the accuracy using external validation data sets.

Jafari Z, et al. Restor Dent Endod. 2018 Apr;43:e27. https://doi.org/10.5395/rde.2018.43.e27

The aim of this study was to determine if a convolutional neural network (CNN) can be trained to automatically detect and localize cervical carotid artery calcifications (CACs) in CBCT. A total of 56 CBCT studies (15,257 axial slices) were utilized to train, validate, and test the deep learning model. The study comprised of two steps: Step 1: Localizing axial slices that are below the C2–C3 disc space. For this step the openly available Inception V3 architecture was trained on the ImageNet dataset of real-world images, and retrained on 40 CBCT studies. Step 2: Detecting CACs in slices from step 1. For this step, two methods were implemented; Method A: Segmentation neural network trained using small patches at random coordinates of the original axial slices; Method B: Segmentation neural network trained using two larger patches at fixed coordinates of the original axial slices with an improved loss function to account for class imbalance. Our approach resulted in 94.2% sensitivity and 96.5% specificity. The mean intersection over union metric for Method A was 76.26% and Method B improved this metric to 82.51%. The proposed CNN model shows the feasibility of deep learning in the detection and localization of CAC in CBCT images.