This is the implementation for the paper: Learning site-specific styles for multi-institutional unsupervised cross-modality domain adaptation
If you find our code/paper helpful for your research, please kindly consider citing our work:
@article{liu2023learning,
title={Learning Site-specific Styles for Multi-institutional Unsupervised Cross-modality Domain Adaptation},
author={Liu, Han and Fan, Yubo and Xu, Zhoubing and Dawant, Benoit M and Oguz, Ipek},
journal={arXiv preprint arXiv:2311.12437},
year={2023}
}
If you have any questions, feel free to contact [email protected] or open an Issue in this repo.
In this playground, you can download our pre-trained sythesis model and some preprocessed data. You will explore how to generate synthetic T2 images with controllable styles!
Participants are provided with the labeled source domain images (T1 contrast-enhanced MRIs), and the unlabeled target domain images (high-resolution T2 MRIs). The goal of this challenge is to obtain a segmentation model for the target domain (T2) with no manual labeling. Particularly, the datasets were collected from multiple institutions, and such heterogeous data lead to extra challenges to the UDA problem.
Our UDA framework is an image-level domain alignment approach and consists of three steps, as shown below. To tackle the data heterogeneity, we aim to make the segmentation model robust to any T2 styles by training the model on images with diverse T2 stlyes.
Instead of removing site-specific styles by data harmonization, we generate them! 🤓
We performed unpaired image translation to translate T1CE MRIs to T2. We extended the QS-Attn to 3D, and modified the generator to a dynamic network. The dynamic network can generate controllable output styles by conditioning on a one-hot site code. Details and demos are provided in the 'synthesis' folder.
We used nnU-Netv2 for segmentation tasks. We created a customized trainer and designed two augmentation techniques to augment the local intensities of the structures-of-interest. Our customized trainer and the augmentation script are provided in the 'segmentation' folder.
Once you download the nnU-Net repo, you need to
- add the 'challenge' folder to the nnUNet/nnunetv2/training/nnUNetTrainer/variants
- replace the 'masking.py' in nnUNet/nnunetv2/training/data_augmentation/custom_transforms.
To further reduce the domain gap, we include real target domain images for training. We use the model obtained by step 2 to generate pseudo labels on the unlabeled target domain images. Then we re-train the nnU-Net using the combined data, i.e., synthetic images (w/ real labels) and real images (w/ pseudo labels). Of course, the quality of pseudo labels matters! In our solution, we simply filter out the poor pseudo labels via connected component analysis. Feel free to explore more advanced pseudo label filtering methods!
Each column belongs to the same site/style. From each column on the right panel, we can see that our synthesized images (top 3 rows) match the style of the real image (bottom row) from the same site.
