T-Mamba: A Unified Framework with Long-Range Dependency in Dual-Domain for 2D & 3D Tooth Segmentation 🦷✨

This repository is the official implementation of the T-Mamba: A unified framework with Long-Range Dependency in dual-domain for 2D & 3D Tooth Segmentation.

Overview

🔥🔥 The code, pre-trained weights, and datasets are fully available.

Currently, T-Mamba supports both 2D and 3D vision tasks. We welcome you to try it out to improve your model's performance.

Dataset 📦

The proposed TED dataset is available at: Hugging Face.

Contact

If you have any questions, please feel free to reach out to me at isjinghao@gmail.com.

Install

conda create -n tmamba python=3.9
conda activate tmamba
pip install -r requirements.txt

cd Tim/causal-conv1d
python setup.py install
cd ../mamba
python setup.py install

=============================
Requirement specific version:
mamba_ssm==1.0.1
causal_conv1d==1.0.0
=============================

Training

sh train_3d.sh # for 3D
sh train_2d.sh # for 2D

Testing (for evaluations)

sh test_3d.sh # for 3D
sh test_2d.sh # for 2D

Inference

sh infer_3d.sh # for 3D
sh infer_2d.sh # for 2D

Citing T-Mamba

If you use TED3 dataset or the T-Mamba network in your research, please use the following BibTeX entry.

@article{hao2026t,
  title={T-Mamba: a unified framework with long-range dependency in dual-domain for 2D \& 3D tooth segmentation},
  author={Hao, Jing and Zhu, Yonghui and He, Lei and Liu, Moyun and Tsoi, James Kit Hon and Hung, Kuo Feng},
  journal={IEEE Transactions on Multimedia},
  year={2026},
  publisher={IEEE}
}