[Segmentation and accuracy validation of mandibular molar and pulp cavity on cone-beam CT images by U-net neural network].

PURPOSE: To realize the automatic segmentation of mandibular molar and pulp cavity on cone-beam CT (CBCT) images by U-net convolutional neural network, and to use the 3D models reconstructed by Micro-CT data as the ground truth to validate its accuracy. METHODS： Twenty groups of small field of view(FOV) CBCT data containing complete mandibular molars were collected from the Department of Radiology, Affiliated Stomatology Hospital of Tongji University. After preprocessing, an endodontic specialist labeled teeth and pulp cavities by MITK Workbench software. These data were used as the training set for training U-net neural network. In addition, five mandibular molars and corresponding small FOV CBCT data were collected. These five CBCT were processed in the same way and used as the testing set. Then, teeth and pulp cavities on CBCT images of the testing set were segmented and reconstructed by U-net neural network and the same specialist. The isolated teeth were scanned by a Micro-CT machine after preprocessing and the results were reconstructed to 3D models, which were used as the ground truth. Then the 3D models reconstructed by the specialist's labeling, U-net network segmentation results, and the ground truth in the testing set were compared. Dice similarity coefficient(DSC), average symmetric surface distance (ASSD), Hausdorff distance (HD), and morphological analysis were used to evaluate the results. SPSS 20.0 software package was used for statistical analysis. RESULTS: Compared with the ground truth, the segmentation accuracy of the U-net neural network measured by DSC, ASSD, and AHD was (95.30±1.01)%, (0.11±0.02) mm, and (1.05±0.31) mm in teeth and (81.21±2.27)%, (0.15±0.05) mm, and (3.29±1.85) mm in the pulp cavity, respectively. Morphological analysis results showed that the U-net network segmentation results were similar to the ground truth in tooth and pulp chamber. As for the segmentation results of root canals, only thick root canals could be segmented rather than the thin root canals, such as the canals in the apical third and lateral root canals. CONCLUSIONS: Under the experimental conditions, the U-net neural network trained by the specialist's labeling realized the automatic and accurate segmentation of mandibular molar and their pulp chamber on CBCT images. For the segmentation of root canals, the results need to be further improved.