Introduction
Accurate 3D tooth segmentation is a fundamental prerequisite for modern computer-aided dental diagnosis, orthodontic planning, and guided surgical interventions. Despite its importance, the clinical workflow remains bottlenecked by the manual delineation of tooth structures in Cone Beam Computed Tomography (CBCT) scans. This process is not only prohibitively labor-intensive—often requiring hours of expert labor per patient—but also suffers from significant inter-observer variability.
While deep learning has demonstrated remarkable success in medical image analysis, its application in dentistry has been hindered by a lack of high-quality, publicly available 3D datasets. Most existing open-source dental benchmarks rely on 2D X-rays, which lack the spatial depth required for complex 3D reconstructions. Furthermore, the few studies addressing 3D segmentation often utilize private datasets, preventing reproducible research and fair performance comparisons. To address these challenges, we introduce CTooth, the first open-source, fully annotated 3D dataset for tooth volume segmentation, alongside a robust benchmark framework.
The CTooth Dataset
The primary contribution of this work is the CTooth dataset, designed to serve as a "gold standard" for the dental research community.
Composition and Diversity
The dataset consists of 22 high-resolution CBCT volumes, totaling 7,363 slices. To ensure the benchmark reflects real-world clinical challenges, the dataset includes a variety of "edge cases," such as:
- Missing teeth and varied dental anatomy.
- Dental implants and prosthetic crowns.
- Metal appliances (e.g., braces), which typically introduce significant imaging artifacts that confound standard segmentation algorithms.
Annotation Process
To ensure ground-truth accuracy, four dental experts performed manual voxel-wise annotations. Each volume required approximately seven hours of meticulous delineation. Of the total slices, 5,504 are fully annotated, providing a dense supervision signal for training deep learning models.
Methodology: The BM-Unet Framework
Alongside the dataset, we propose BM-Unet (Benchmark U-Net), a 3D attention-based framework optimized for the unique geometry of dental structures. The architecture incorporates several key technical innovations:
- Residual Encoder Blocks (REBs): By utilizing shortcut connections, the network preserves high-frequency spatial details, ensuring that "tiny" features of tooth roots and narrow inter-dental spaces are not lost during downsampling.
- Attention Bottleneck: An attention branch is integrated at the model's bottleneck to help the network selectively focus on tooth structures while suppressing background noise and artifacts caused by metal fillings.
- Deep Supervision: To account for the significant variation in tooth size and shape, the model employs deep supervision at multiple scales, refining the segmentation response throughout the training process.
- Weighted Dice Loss: Given that teeth occupy a relatively small volume within a CBCT scan, a weighted loss function is used to mitigate class imbalance and prevent the model from becoming biased toward the background voxels.
Key Insights and Results
Our experiments demonstrate that the integration of attention mechanisms is critical for 3D dental segmentation. Specifically, the use of SKNet (Selective Kernel) attention proved most effective, as it allows the network to adaptively adjust its receptive field to capture both large molars and delicate root structures.
Performance Metrics
The proposed framework achieved state-of-the-art results on the CTooth benchmark, significantly outperforming established medical segmentation baselines:
- Dice Score: 88.04%
- Intersection over Union (IoU): 78.71%
- Sensitivity: 94.71%
In comparative testing, the attention-based 3D U-Net achieved a gain of over 25% in Dice score compared to DenseVoxelNet and consistently outperformed other standard architectures like V-Net and 3D HighResNet. These results suggest that the model is highly effective at identifying tooth voxels even in the presence of complex artifacts.
Conclusion and Future Work
CTooth provides a much-needed standardized benchmark for the dental imaging community, moving the field beyond the limitations of 2D panoramic views. While the current dataset of 22 volumes is a significant step forward, we acknowledge the need for larger, more diverse cohorts. Future work will focus on expanding CTooth into a multi-organization dataset, incorporating scans from different CT manufacturers and more diverse patient populations to further enhance model generalization and clinical utility.
Summary updated