The study on precise segmentation of micro-carious lesions in CBCT using SBC-Net

In dentistry, accurate segmentation of carious lesions is crucial for clinical diagnosis. Conventional approaches rely on clinician experience, while most studies focus on 2D panoramic radiographs, which cannot capture spatial structures or boundary details. This study proposes a two-stage method for segmenting carious regions on dental cone beam computed tomography (CBCT). First, a deep learning model extracts the dental foreground, computes its boundary, and crops the volume to reduce background interference and emphasize spatial priors. Second, we introduce small-object and boundary-sensitive caries segmentation 3D U-Net (SBC-Net) for micro-lesion segmentation. Comprehensive comparative and ablation experiments demonstrate that our model achieves a superior Dice coefficient of 81.89%, significantly outperforming mainstream baselines and single-module variants. The results indicate that imposing boundary constraints on the dentition region, rather than simple dilation, better guides the model to exploit spatial structure for localizing carious areas. Moreover, using larger patch sizes to expand the model’s receptive field coverage effectively improves capture of carious regions and significantly boosts segmentation performance. We also performed lesion-level stratified analyses across affected surfaces and found that primary caries achieved higher segmentation accuracy than secondary caries. Occlusal and smooth surface lesions were easier to segment, whereas proximal, root surface, and hidden lesions were harder, supporting the model’s stability and clinical applicability across scenarios. The proposed two-stage 3D automatic segmentation approach facilitates precise volumetric localization and quantitative analysis of dental caries in clinical practice, advancing personalized treatment.