Abstract
Background: Accurate tooth segmentation from cone beam computed tomography (CBCT) images is crucial for digital dentistry but remains challenging in cases of interdental adhesions, which cause severe anatomical shape distortion. Methods: To address this, we propose a deep learning framework that integrates semantic and shape awareness for shape-preserving segmentation. Our method introduces a target-tooth-centroid prompted multi-label learning strategy to model semantic relationships between teeth, reducing shape ambiguity. Additionally, a tooth-shape-aware learning mechanism explicitly enforces morphological constraints to preserve boundary integrity. These components are unified via multi-task learning, jointly optimizing segmentation and shape preservation. Results: Extensive evaluations on internal and external datasets demonstrate that our approach significantly outperforms existing methods. Conclusions: Our approach effectively mitigates shape distortions and providing anatomically faithful tooth boundaries.
Introduction
In modern digital dentistry, the generation of high-fidelity 3D models from Cone Beam Computed Tomography (CBCT) is a prerequisite for precision procedures such as implant planning and orthodontic treatment. However, automated segmentation faces a persistent hurdle: interdental adhesion. Due to the close proximity of teeth and their similar radiographic densities, standard segmentation algorithms often fail to distinguish boundaries, leading to "fused" tooth models. These distortions compromise the clinical utility of the resulting 3D reconstructions. This paper introduces a novel framework, CMS-Net, designed to resolve these ambiguities by incorporating spatial and geometric priors.
Methods
The proposed framework employs a hierarchical three-stage pipeline to achieve precise instance segmentation.
1. Dentition Segmentation
The process begins with a global identification of the dental arch. A U-Net-based module segments the entire dentition to define a Region of Interest (ROI), effectively filtering out non-dental anatomical structures.
2. Centroid Clustering
Within the identified ROI, the system predicts the 3D centroids for each individual tooth. These centroids serve as spatial anchors, ensuring the model correctly identifies the number and location of all teeth present.
3. Semantic and Shape-Aware Segmentation
The core of the innovation lies in this final stage, which utilizes three key mechanisms:
- Centroid Prompting: Centroids act as prompts to focus the network's attention on a specific target tooth.
- Multi-Label Learning: The model is trained to simultaneously identify the target tooth, its neighbors, and the background. This explicit modeling of neighboring structures forces the network to learn the precise boundaries where teeth touch.
- Shape-Aware Learning (SDM): Instead of simple binary masks, the model predicts a Signed Distance Map (SDM). This geometric representation enforces morphological constraints, ensuring the output maintains anatomically correct, smooth tooth shapes even in noisy or low-contrast regions.
Results
The CMS-Net framework was evaluated on an internal dataset of 827 CBCT images and an external dataset of 60 images to test generalizability.
- Accuracy: The model achieved a Dice coefficient of 0.9408 on the internal dataset and 0.8934 on the external dataset, outperforming state-of-the-art (SOTA) benchmarks.
- Boundary Precision: The Hausdorff Distance (HD), which measures boundary error, was reduced to 1.19mm, a significant improvement over the 2.84mm observed in previous models.
- Robustness: High performance on the external dataset confirms the model's ability to handle varying scanning parameters and noise levels across different clinical institutions.
Discussion and Conclusion
By integrating semantic context (neighboring teeth) and geometric priors (SDM), CMS-Net successfully addresses the challenge of interdental adhesion. The use of centroid prompting provides an efficient mechanism for instance segmentation without the overhead of more complex architectures.
Future work will focus on expanding the dataset to include a wider variety of dental pathologies and complex restorations. Additionally, integrating patient metadata (e.g., age, clinical history) could further refine the model's predictive accuracy. Ultimately, this framework provides a robust foundation for integrating automated 3D tooth modeling into real-time clinical workflows.
Summary updated