samleoqh / Viola-Unet

A description of the winning solution for the validation dataset in the 2022 Intracranial Hemorrhage Segmentation challenge (INSTANCE 2022)

Viola-Unet

A description of the winning solution: Viola-Unet for the validation dataset in the 2022 Intracranial Hemorrhage Segmentation challenge (INSTANCE 2022)

Test the pre-trained models in docker

1. Download the docker image crainet.tar.gz
2. Prepare your input folder such as /home/yourname/Desktop/input (a folder contains all CT files for testing) will be synchronized with /input in the docker container, and output folder such as /home/yourname/Desktop/predict (an empty folder used to save segmentation file) will be synchronized with /predict in the docker container. Note that, the filename of the segmentation mask file (./predict) is the same as the CT file (./input), the segmentation mask is a 3D zero-one array(0 stands for background, 1 stands for target), and the meta information of the segmentation mask file is consistent with that of original CT file (*.nii.gz).
3. Run the program with commands as follows (assuming on Linux OS) with outputing inference messages to stdout（-e PYTHONUNBUFFERED=1）

docker load < crainet.tar.gz
docker run --gpus "device=0" --name crainetx -e PYTHONUNBUFFERED=1 -v /home/yourname/Desktop/input:/input -v /home/yourname/Desktop/predict:/predict crainet:latest

4. The program will do following: first, read each CT file(*.nii.gz) in folder ./input, then, use pre-trained models (ensemble of viola-Unet and nnU-Net) to segment the CT scans one by one and save the segmented masks to ./predict, and output inference messages to terminal like:

model nnUNet-kf0 loaded successfully!
model nnUNet-kf1 loaded successfully!
model nnUNet-kf2 loaded successfully!
model nnUNet-kf3 loaded successfully!
model nnUNet-kf4 loaded successfully!
model ViolaUNet_l-kf0 loaded successfully!
model ViolaUNet_l-kf1 loaded successfully!
model ViolaUNet_l-kf2 loaded successfully!
model ViolaUNet_l-kf3 loaded successfully!
model ViolaUNet_l-kf4 loaded successfully!

------------------start predicting input volume: 001.nii.gz - 1/130 -------------------
Predicted lesion volume : 0.178 ml
001.nii.gz
--------Cost time: 10.901 sec --------

------------------start predicting input volume: 002.nii.gz - 2/130 -------------------
Predicted lesion volume : 6.677 ml
002.nii.gz
--------Cost time: 6.153 sec --------

------------------start predicting input volume: 003.nii.gz - 3/130 -------------------

5. The structrue of the input and output folder for testing :

├── /home/yourname/Desktop/input
          │   ├── 144.nii.gz
          │   ├── 145.nii.gz
          │   ├── 146.nii.gz

├── /home/yourname/Desktop/predict
          │   ├── 144.nii.gz
          │   ├── 145.nii.gz
          │   ├── 146.nii.gz

6. What if only having CPU and Windows OS, you can Run the inference as following:

docker load -i crainet.tar.gz
docker run --name crainetx -v D:\data\CT\test\input\:/input -v D:\data\CT\test\predict\:/predict crainet:latest

just replace D:\data\CT\test\input\ and D:\data\CT\test\predict\ with your own folder path.

Citation:

Please consider citing our work if you find the code helps you

@inproceedings{liu2023ICH,
  title={Voxels Intersecting along Orthogonal Levels Attention U-Net for Intracerebral Haemorrhage Segmentation in Head CT},
  author={Qinghui Liu and Bradley J MacIntosh and Till Schellhorn and Karoline Skogen and KyrreEeg Emblem and Atle Bjørnerud},
  booktitle={Proceedings of ISBI 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI)},
  year={2023}
}