Root canal segmentation from cone-beam computed tomography guided by micro-computed tomography based on deep learning

Accurate root canal segmentation from cone-beam computed tomography (CBCT) is essential for endodontic diagnosis and treatment planning. This study aims to explore the feasibility of using deep learning (DL) models, trained on CBCT images of extracted teeth guided by micro-computed tomography (µCT), for clinical CBCT image segmentation. A dataset of 56 extracted teeth with diverse root canal complexities was constructed, combining CBCT and µCT scans. Ground truth annotations were derived from µCT-based masks and registered to CBCT images. DL models based on U2-Net architecture were trained and evaluated for tooth and root canal segmentation, comparing µCT-guided and manual-label-based approaches. We further investigated the impact of voxel size and image resampling on performance. Finally, the trained models were applied to segment root canals in clinical CBCT images, with the results validated by endodontic specialists. The µCT-guided AI segmentation method outperformed the manual-label-based approach. Utilizing a smaller native voxel size (80 μm), coupled with image resampling, proved particularly advantageous for capturing intricate anatomical details. In clinical validation, the model delivered rapid and accurate root canal segmentation, with 94% of single-rooted teeth and 100% of molars rated as “excellent” or “good”. Results demonstrated the potential of µCT-guided AI models for enhancing root canal segmentation in clinical practice, offering a promising tool for digital dentistry. Not applicable.