Incorporating the distance Transform maps of image segmentation labels into CNNs-based segmentation tasks has received significant attention in 2019. These methods can be classified into two main classes in terms of the main usage of distance transform maps.

• Designing new loss functions
• Adding an auxiliary task, e.g. distance map regression

However, with these new methods on the one hand and the diversity of the specific implementations and dataset-related challenges on the other, it's hard to figure out which design can generalize well beyond the experiments in the original papers. In this repository, we want to re-implement these methods (published in 2019) and evaluate them on the same 3D segmentation tasks (heart and liver tumor segmentation).

Here is the code, and trained modles can be downloaded from Baidu Disk (pw:mgn0).

The authors would like to thank the organization team of MICCAI 2017 liver tumor segmentation challenge MICCAI 2018 and left atrial segmentation challenge for the publicly available dataset. We also thank the reviewers for their valuable comments and suggestions. We appreciate Cheng Chen, Feng Cheng, Mengzhang Li, Chengwei Su, Chengfeng Zhou and Yaliang Zhao to help us finish some experiments. Last but not least, we thank Lequan Yu for his great PyTorch implementation of V-Net and Fabian Isensee for his great PyTorch implementation of nnU-Net.

@inproceedings{ma-MIDL2020-SegWithDist,
  title={How Distance Transform Maps Boost Segmentation CNNs: An Empirical Study},
  author={Ma, Jun and Wei, Zhan and Zhang, Yiwen and Wang, Yixin and Lv, Rongfei and Zhu, Cheng and Chen, Gaoxiang and Liu, Jianan and Peng, Chao and Wang, Lei and Wang, Yunpeng and Chen, Jianan},
  booktitle={Medical Imaging with Deep Learning},
  pages = {479--492},
  volume = {121},
  month = {06--08 Jul},
  year={2020},
  series = {Proceedings of Machine Learning Research},
  editor = {Tal Arbel and Ismail Ben Ayed and Marleen de Bruijne and Maxime Descoteaux and Herve Lombaert and Christopher Pal},
  publisher = {PMLR},
  url = {https://proceedings.mlr.press/v121/ma20b.html}
}