Lightunetr Models

--- 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.