| --- |
| license: mit |
| pipeline_tag: image-segmentation |
| tags: |
| - medical |
| - 3d |
| - transformer |
| --- |
| |
| # Light-UNETR |
|
|
| Light-UNETR is a lightweight transformer architecture designed for efficient 3D medical image segmentation, introduced in the paper [Harnessing Lightweight Transformer with Contextual Synergic Enhancement for Efficient 3D Medical Image Segmentation](https://huggingface.co/papers/2603.23390). |
|
|
| The model addresses computational efficiency through a Lightweight Dimension Reductive Attention (LIDR) module and a Compact Gated Linear Unit (CGLU). To improve data efficiency, the authors propose a Contextual Synergic Enhancement (CSE) learning strategy. |
|
|
| ## Resources |
|
|
| - **Paper:** [Harnessing Lightweight Transformer with Contextual Synergic Enhancement for Efficient 3D Medical Image Segmentation](https://huggingface.co/papers/2603.23390) |
| - **Code:** [Official GitHub Repository](https://github.com/CUHK-AIM-Group/Light-UNETR) |
|
|
| ## Performance |
|
|
| Light-UNETR significantly reduces computational costs compared to standard architectures. For instance, on the Left Atrial (LA) Segmentation dataset, it reduces FLOPs by 90.8% and parameters by 85.8% compared to state-of-the-art methods while achieving superior performance even with limited (10%) labeled data. |
|
|
| ## Sample Usage |
|
|
| To test a pre-trained Light-UNETR model using the official implementation, you can use the following command structure: |
|
|
| ```bash |
| # Example: Test BraTS model |
| python test_cse.py --dataset brats --model lightunetr --checkpoint lightunetr_best_model_brats_25lab.pth --gpu 0 |
| ``` |
|
|
| ## Citation |
|
|
| If you find this work useful, please cite: |
|
|
| ```bibtex |
| @article{liu2025harnessing, |
| title={Harnessing Lightweight Transformer with Contextual Synergic Enhancement for Efficient 3D Medical Image Segmentation}, |
| author={Liu, Xinyu and Chen, Zhen and Li, Wuyang and Li, Chenxin and Yuan, Yixuan}, |
| year={2025} |
| } |
| ``` |
|
|
| ## Acknowledgement |
|
|
| The authors appreciate the contributions of [SSL4MIS](https://github.com/HiLab-git/SSL4MIS), [Slim UNETR](https://github.com/aigzhusmart/Slim-UNETR), [BCP](https://github.com/DeepMed-Lab-ECNU/BCP), and other referenced codebases. |
