Semi-automatic individual tooth segmentation from CBCT images using anatomy-guided deep neural networks and watershed transform.

Individual tooth segmentation in cone beam computed tomography (CBCT) scan plays a crucial role in quantitatively analyzing oral diseases. However, the existing full-automatic methods exhibit limited accuracy and robustness due to the complex noise artifacts, intricate tooth morphological variability, and weak boundaries. To address this problem, we propose a semi-automatic individual tooth segmentation approach. Firstly, a semi-automatic tooth landmark detection algorithm was proposed to extract the position information of individual teeth. Secondly, we designed a deep learning network with a channel spatial attention module to segment the alveolar bone for detecting the region-of-interest (ROI) of the tooth. Next, we incorporated the recognized tooth landmark and alveolar bone into the proposed segmentation network as the anatomical information for delineating the tooth foreground. Finally, we utilized the marker-controlled watershed transform algorithm with the detected tooth landmarks to separate the individual tooth objects from the tooth foreground. In our study, 78 CBCT scans were acquired and allocated into the training set, validation set, and testing set with a ratio of 5:2:3. The proposed system yielded mean Dice coefficients (DC) of 0.91±0.03 and Average symmetric surface distance (ASSD) of 0.55±0.21 mm for alveolar bone segmentation, and mean DC of 0.91±0.01 and ASSD of 0.29±0.23 mm for individual tooth segmentation, demonstrating a good consistency with manual segmentations. The proposed semi-automatic segmentation method achieves high accuracy and efficiency in oral anatomy structure delineation from CBCT scans and has a high potential for future clinical applications.