A study on the segmentation method of dental and pulp cavities in CBCT images based on improved U-Net neural network

ZHANG Ruolin; SHAN Haoxuan; FU Yujie; CHEN Yufei; ZHANG Qi

doi:10.15956/j.cnki.chin.j.conserv.dent.2025.12.003

Abstract：Objective By improving the U-Net neural network to address the challenge of apical 4mm root canal segmentation,accurate and automated segmentation of premolar structures and intact pulp cavities in CBCT images can be achieved.This provides clinicians with complete and precise 3D root canal anatomical information for treatment.Methods Paired CBCT and Micro-CT data of 84 premolars were collected,with 70 for training and 14 for testing.For the U-Net neural network training,the control group involved manual labeling of CBCT images for training,while the experimental group used registered and labeled CBCT images based on Micro-CT,along with targeted training of apical 4mm root canal segmentation via an apical foramen localization algorithm.After training,both networks were tested using the test set,with Micro-CT-derived 3D models as the reference standard.The overall and apical 4mm root canal segmentation results of the two networks were compared to the reference standard using the Dice Similarity Coefficient for similarity assessment,and the Mean Symmetric Surface Distance and Hausdorff Distance for difference evaluation.Morphological analysis was also conducted for qualitative assessment of the segmentation results.Results The experimental group showed superior performance in all metrics.The Dice similarity coeffcient,mean symmetric surface distance,and Hausdorff distance for dental structure were(97.20±0.48)%,(0.08±0.01)mm,and(0.82±0.23)mm,respectively.For the pulp cavity,these metrics were(86.21±6.10)%,(0.19±0.04)mm,and(1.67±0.51)mm.For the apical 4 mm root canal,they were(71.62±6.13)%,(0.15±0.08)mm,and(0.51±0.12)mm.Morphological analysis also confirmed better segmentation results in the experimental group.Conclusion The modified U-Net neural network,optimized by an apical foramen localization algorithm,can achieve complete segmentation of dental structures and pulp cavities in CBCT images,offering precise 3D anatomical data for root canal therapy.