An Anatomically Guided AI System for Automated Crown–Root Segmentation in Dental CBCT

About the Project

Cone-beam computed tomography (CBCT) provides three-dimensional images of teeth and surrounding structures, supporting detailed dental diagnosis and treatment planning. However, most available dental CBCT datasets label each tooth as a single structure without separating its crown from its root, limiting the development of automated systems for analyzing root morphology, root position and its relationship with the surrounding bone. Creating separate crown and root annotations manually is also time-consuming and requires substantial expert effort. This project developed an anatomically guided synthetic-annotation pipeline that transforms existing whole-tooth masks into separate crown and root labels. Stage 1 used principal component analysis and CBCT intensity to estimate each tooth’s long axis, determine its crown-root orientation and automatically sample reliable crown and root prompt points, achieving an overall prompt accuracy of 99.6%. Stage 2 generated 902 synthetic masks from 451 teeth and applied structural validation and quality review, with 98.2% of evaluated teeth accepted for dataset construction. The accepted labels were then used to train a lightweight voxel-level neural network. When evaluated on 82 teeth from three independent clinical cases with manual annotations, the model achieved mean Dice scores of 0.8924 for crowns, 0.9149 for roots and 0.9037 overall, demonstrating the potential of synthetic annotation to reduce reliance on fully manual labeling.