This repository is pytorch's implement of paper: WSSNet: Aortic wall shear stress estimation using deep learning on 4D Flow MRI
Official TensorFolw implementation are here: https://github.com/EdwardFerdian/WSSNet . You can download datasets here and place them in the same level directory as the master like :
├── dataset
| ├── train
| └── val
| └── test
├── master
In this implementation, it not only includes CNN, but also the implementation of Swin Transformer.
I tried, but the effect was not particularly good.
Unlike the official implementation, I did not include a regular loss term. (Just SSIM + MAE)
Here are some results of SwinT:
| Case | MAE | rel (%) | Pearson |
|---|---|---|---|
| Val #1 (normal) | 0.59 | 11.67 | 0.57 |
| Val #2 (normal) | 0.42 | 11.67 | 0.71 |
| Val #3 (normal) | 0.48 | 15.16 | 0.64 |
| Test #1 (normal) | 0.74 | 13.49 | 0.57 |
| Test #2 (LVH) | 1.07 | 16.54 | 0.70 |
| Test #3 (normal) | 0.58 | 12.01 | 0.64 |
| Overall | 0.65 | 13.42 | 0.63 |
Here are some visualizations:
In the visualization file, I only modified the file separately: plot_ tawss_ osi_ flatmap.py. Therefore, the work of modifying other visualization files is entrusted to later parties.
You can train by:
python main.py
You can test some case by:
python test.py