U-Mamba2: Scaling State Space Models for Dental Anatomy Segmentation in CBCT

Zhi Qin Tan, Xiatian Zhu, Owen Addison, Yunpeng Li

Introduction

Accurate segmentation of dental anatomy—including jaws, teeth, and nerves—from Cone-Beam Computed Tomography (CBCT) scans is a cornerstone of modern dental diagnostics and surgical planning. However, the high resolution and 3D nature of these volumes present significant computational challenges. Traditional Convolutional Neural Networks (CNNs) often fail to capture the necessary global context, while Transformer-based architectures suffer from quadratic complexity, leading to prohibitive memory and time costs. This paper presents U-Mamba2, a novel architecture that leverages State Space Models (SSMs) to achieve high-accuracy, efficient 3D segmentation.

The U-Mamba2 Architecture

U-Mamba2 evolves the previous U-Mamba design by integrating the Mamba2 Structured State-Space Duality (SSD) framework. The model follows a symmetric U-Net-like structure, utilizing standard residual blocks for local feature extraction in the encoder and decoder stages. The core innovation lies in the "bottleneck" of the network, where a specialized U-Mamba2 block is placed to capture long-range global dependencies with linear scaling complexity.

The U-Mamba2 architecture overview, showing the U-Net structure with the specialized U-Mamba2 block at the bottleneck and the optional interactive click data branch.

Unlike the original Mamba's selective scan mechanism, Mamba2 utilizes matrix multiplication-based sequences. This shift allows for significantly better hardware utilization and parallelism, enabling the model to scale effectively to the massive voxel counts inherent in high-resolution medical imaging. Furthermore, for scenarios requiring human-in-the-loop refinement, the architecture includes an interactive branch. This branch uses a point encoder and cross-attention blocks to incorporate user clicks, allowing the model to dynamically adjust segmentations based on expert feedback.

Domain-Specific Enhancements

Beyond architectural improvements, U-Mamba2 incorporates several dental-specific priors to boost performance:

  • Label Smoothing: Applied to help the model distinguish between highly similar adjacent teeth.
  • Weighted Loss: A specialized loss function assigns 10x more importance to tiny, critical structures like the incisive nerves.
  • Symmetry-Aware Augmentation: A mirroring technique that swaps left and right labels during training, exploiting anatomical symmetry without confusing the model's spatial orientation.

Experimental Results

The model was evaluated using data from the ToothFairy3 Challenge, which includes 532 labeled CBCT scans covering 46 anatomical classes. U-Mamba2 achieved first place in both competition tasks. In Task 1 (multi-anatomy segmentation), it reached a mean Dice score of 0.84. In Task 2 (interactive nerve segmentation), it achieved a mean Dice of 0.87 with a remarkably low 95th-percentile Hausdorff Distance (HD95) of 2.15.

Qualitative segmentation results of U-Mamba2 on CBCT scans, demonstrating precise multi-class identification of teeth, jaws, and nerves across different anatomical views.

Efficiency was a key highlight; U-Mamba2 maintained an average inference time of 6-7 seconds per scan on standard hardware, significantly outperforming transformer-based alternatives like SwinUNETR. By tuning the sliding window inference overlap, the authors further reduced inference time by nearly 13% with negligible loss in accuracy.

Conclusion

U-Mamba2 demonstrates that scaling State Space Models, combined with domain-specific anatomical knowledge, provides a powerful solution for complex 3D medical segmentation tasks. While metallic artifacts from dental implants remain a challenge for future robustness, the current framework sets a new benchmark for efficiency and accuracy in dental AI.

Summary updated

BibTeX
@inproceedings{tan2025u,
  author = {Zhi Qin Tan and Xiatian Zhu and Owen Addison and Yunpeng Li},
  title = {{U-Mamba2: Scaling State Space Models for Dental Anatomy Segmentation in CBCT}},
  booktitle = {ODIN @ MICCAI},
  year = {2025},
  url = {https://openreview.net/forum?id=XrGUrxXfXj},
}
Figure