S3R: Shape and Semantics-based Selective Regularization for Explainable Continual Segmentation across Multiple Sites
Here is the implementation for paper:
Comprehensive Importance-based Selective Regularization for Continual Segmentation across Multiple Sites. (https://link.springer.com/chapter/10.1007/978-3-030-87193-2_37) (MICCAI)
S3R: Shape and Semantics-based Selective Regularization for Explainable Continual Segmentation across Multiple Sites (TMI 10.1109/TMI.2023.3260974)
In clinical practice, a segmentation network is often required to continually learn on a sequential data stream from multiple sites rather than a consolidated set. However, during the continual learning process, existing methods are usually restricted by the poor network memorizability on previous sites when learning from images from a new site. In this project, we tackle this interesting problem setting of continual segmentation across multiple sites, by proposing a shape and semantics-based selective regularization (S3R) to enhance both shape and semantic preserving capabilities.
Multi-site prostate MR images can be found HERE, collected from six different data sources out of three public datasets. The preprocessing pipeline in dataloder follows Quande's work, such as SAML and FedDG.
Run train.py
, where training configurations are specified.
This calls regularization_CISR.py
, where network structure and training function are defined:
python run train.py --agent_type regularization_CISR --agent_name Reg_CISR --model_type seg_regression --model_name Unet_regression
Note that in each learning round, the segmentation embedding module need to be pre-trained as suggested by ACNN.
If you make use of the code, please cite:
@article{zhang2021CISR,
title={Comprehensive Importance-based Selective Regularization for Continual Segmentation across Multiple Sites},
author={Zhang, Jingyang and Gu, Ran and Wang, Guotai and Gu, Lixu},
journal={International Conference on Medical Image Computing and Computer Assisted Intervention},
year={2021}
}
@article{zhang2023TMI,
title={S3R: Shape and Semantics-based Selective Regularization for Explainable Continual Segmentation across Multiple Sites},
author={Zhang, Jingyang and Gu, Ran and Peng, Xue and Liu, Mianxin, and Zheng, Hao and Zheng, Yefeng and Ma, Lei and Wang, Guotai and Gu, Lixu},
journal={IEEE Transactions on Medical Imaging },
year={2023}
}
For further question about the code or dataset, please contact '[email protected]'