This repository contains scripts to train and run inference with LHU-Net using the AbdomenAtlas1.0 dataset.
-
Dataset (AbdomenAtlas1.0Mini): available on Hugging Face
Download -
Pretrained weights (used for Touchstone benchmark):
Download
- Update the dataset/output paths in the training config files.
- Run:
cd train
./train.sh- Update the dataset/output paths in the inference config files.
- Run:
cd inference
./inference.shIf you only want predictions (no metric calculation), comment out the last line in alex.sh.
- The LHU-Net version used in the Touchstone benchmark is a weaker variant than the one in the main LHU-Net repository.
- This repository is implemented with MONAI, not nnUNetv2.
@article{bassi2024touchstone,
title={Touchstone benchmark: Are we on the right way for evaluating ai algorithms for medical segmentation?},
author={Bassi, Pedro RAS and Li, Wenxuan and Tang, Yucheng and Isensee, Fabian and Wang, Zifu and Chen, Jieneng and Chou, Yu-Cheng and Kirchhoff, Yannick and Rokuss, Maximilian R and Huang, Ziyan and others},
journal={Advances in Neural Information Processing Systems},
volume={37},
pages={15184--15201},
year={2024}
}LHU-Net:
@inproceedings{sadegheih2025lhu,
title={LHU-Net: A lean hybrid u-net for cost-efficient, high-performance volumetric segmentation},
author={Sadegheih, Yousef and Bozorgpour, Afshin and Kumari, Pratibha and Azad, Reza and Merhof, Dorit},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={326--336},
year={2025},
organization={Springer}
}