Micro–Computed Tomography–Guided Artificial Intelligence for Pulp Cavity and Tooth Segmentation on Cone-beam Computed Tomography

Abstract

IntroductionThis study proposes a novel data pipeline based on micro–computed tomographic (micro-CT) data for training the U-Net network to realize the automatic and accurate segmentation of the pulp cavity and tooth on cone-beam computed tomographic (CBCT) images. MethodsWe collected CBCT data and micro-CT data of 30 teeth. CBCT data were processed and transformed into small field of view and high-resolution CBCT images of each tooth. Twenty-five sets were randomly assigned to the training set and the remaining 5 sets to the test set. We used 2 data pipelines for U-Net network training: one manually labeled by an endodontic specialist as the control group and one processed from the micro-CT data as the experimental group. The 3-dimensional models constructed using micro-CT data in the test set were taken as the ground truth. The Dice similarity coefficient, precision rate, recall rate, average symmetric surface distance, Hausdorff distance, and morphologic analysis were used for performance evaluation. ResultsThe segmentation accuracy of the experimental group measured by the Dice similarity coefficient, precision rate, recall rate, average symmetric surface distance, and Hausdorff distance were 96.20% ± 0.58%, 97.31% ± 0.38%, 95.11% ± 0.97%, 0.09 ± 0.01 mm, and 1.54 ± 0.51 mm in the tooth and 86.75% ± 2.42%, 84.45% ± 7.77%, 89.94% ± 4.56%, 0.08 ± 0.02 mm, and 1.99 ± 0.67 mm in the pulp cavity, respectively, which were better than the control group. Morphologic analysis suggested the segmentation results of the experimental group were better than those of the control group. ConclusionsThis study proposed an automatic and accurate approach for tooth and pulp cavity segmentation on CBCT images, which can be applied in research and clinical tasks.