Anatomy Aware Cascade Network: Bridging Epistemic Uncertainty and Geometric Manifold for 3D Tooth Segmentation

TL;DR

This work tackles the persistent challenge of accurate 3D tooth segmentation from CBCT by addressing boundary ambiguity and topological consistency. The authors introduce AACNet, a coarse-to-fine cascade that jointly learns an epistemic-uncertainty-driven boundary refinement (AGBR) and a geometry-aware attention mechanism guided by a Signed Distance Map (SDMAA). Through Stage I global localization and Stage II fine segmentation, AACNet achieves state-of-the-art performance (DSC 90.17%, HD95 3.63 mm on internal data; HD95 2.19 mm on external data) and demonstrates robust generalization and clinically relevant measurements for sinus and inferior alveolar canal proximity. The approach leverages implicit geometric priors to preserve anatomical topology while focusing refinement on uncertain regions, offering a clinically viable path toward automated digital dental workflows. Code is available at the provided repository.

Abstract

Accurate three-dimensional (3D) tooth segmentation from Cone-Beam Computed Tomography (CBCT) is a prerequisite for digital dental workflows. However, achieving high-fidelity segmentation remains challenging due to adhesion artifacts in naturally occluded scans, which are caused by low contrast and indistinct inter-arch boundaries. To address these limitations, we propose the Anatomy Aware Cascade Network (AACNet), a coarse-to-fine framework designed to resolve boundary ambiguity while maintaining global structural consistency. Specifically, we introduce two mechanisms: the Ambiguity Gated Boundary Refiner (AGBR) and the Signed Distance Map guided Anatomical Attention (SDMAA). The AGBR employs an entropy based gating mechanism to perform targeted feature rectification in high uncertainty transition zones. Meanwhile, the SDMAA integrates implicit geometric constraints via signed distance map to enforce topological consistency, preventing the loss of spatial details associated with standard pooling. Experimental results on a dataset of 125 CBCT volumes demonstrate that AACNet achieves a Dice Similarity Coefficient of 90.17 \% and a 95\% Hausdorff Distance of 3.63 mm, significantly outperforming state-of-the-art methods. Furthermore, the model exhibits strong generalization on an external dataset with an HD95 of 2.19 mm, validating its reliability for downstream clinical applications such as surgical planning. Code for AACNet is available at https://github.com/shiliu0114/AACNet.

Figures (11)

• Figure 1: Illustration of typical adhesion artifacts. (a) Inter proximal adhesion between adjacent teeth. (b) Inter arch adhesion at the occlusal interface. In both scenarios, low grayscale contrast at contact zones leads to severe boundary ambiguity and erroneous fusion in conventional segmentation.
• Figure 2: Overview of the proposed AACNet framework. The method adopts a coarse-to-fine decomposition strategy where the green line illustrates Stage I generating a global ambiguity field from the input volume, shown as (a) the original CBCT, resulting in the probabilistic priors displayed as (b) the coarse localization maps ($P_{upper}, P_{lower}$). Subsequently, the yellow line depicts Stage II fusing these priors with the original image. By utilizing two key mechanisms, AGBR and SDMAA, the network hierarchically addresses boundary ambiguity and enforces topological consistency, yielding (c) the final segmentation result.
• Figure 3: Detailed architecture of the fundamental building blocks used in the Residual U-Net backbone. (a) The Encoder Residual Block, featuring stacked convolutional layers with Instance Normalization and LeakyReLU to preserve feature identity during deep extraction. (b) The Decoder Basic Block, designed for spatial recovery, utilizing transposed convolutions for upsampling followed by feature concatenation and refinement convolutions.
• Figure 4: Visualization of the ambiguity field $A(v)$ derived from Stage I predictions. Bright regions indicate voxels of maximal epistemic uncertainty, which correspond precisely to the blurred occlusal interfaces and inter proximal contact zones. By applying a high threshold ($\tau=0.95$), we generate a sparse binary gating mask from this field, directing the AGBR module to focus refinement efforts exclusively on these hardest-to-segment boundaries.
• Figure 5: Detailed of the AGBR module. The module quantifies voxel-level epistemic uncertainty from Stage I predictions to generate a binary gating mask via thresholding. This mask acts as a dynamic switch, activating the residual refinement branch $\mathcal{R}(\cdot)$ exclusively in high-entropy transition zones. This mechanism performs targeted feature rectification on ambiguous boundaries while preserving the integrity of high-confidence regions.