Uncertainty Based Border-Aware Segmentation Network for Deep Caries

Deep caries is a progressive and destructive disease affecting the hard surface of teeth. If left untreated, it can lead to serious health risks such as inflammation and apical periodontitis. According to clinical evidence, carious dentin can be categorized into soft, firm, and hard dentin based on its hardness. Precise assessment of carious lesions is critical for effective treatment; however, current methods rely on subjective judgments based on tactile feedback, leading to variability in dentin removal and potential treatment failure. To address this problem and provide accurate references to the dentist, we propose a border-aware network for deep caries segmentation using clinical Microcomputed tomography (MicroCT) data. The network performs segmentation of the three types of dentin within the cavity and incorporates the Signed Distance Field (SDF) method to enhance accuracy at the borders of caries. To evaluate the clinical feasibility, we simulate CBCT images by Gaussian-blurring MicroCT images and introduce evidence theory to estimate uncertainty in ambiguous border regions, ensuring robustness against low-quality inputs. We collect a real-world dental dataset which includes 25 MicroCT scans with deep caries. Through experiments, we demonstrate that our proposed method improves segmentation accuracy, especially in images with different degrees of blurriness. Moreover, by segmenting images with different levels of Gaussian blurring, we validate the robustness of our proposed method in handling low-quality inputs, thus showing its potential for future clinical applications.