MartinRovang/WMH_segmentation

WMH segmentation using FLAIR.
Research article »


Report Bug · Request Feature

Weights can be recived on request.

Run on your own computer

1. Clone repo.

2. Change directory to repo.

3. Add datafolder in current format;

-data
--imagesTr
---ID1.nii.gz
   ...
   
--labelsTr
---ID1.nii.gz
   ...

--imagesvalTr
---ID5.nii.gz
   ...
   
--labelsvalTr
---ID5.nii.gz
   ...

4. Run command docker build -t WMHSEG .

5. Run command docker run -it --gpus all --shm-size=8g --ulimit memlock=-1 --ulimit stack=67108864 --rm -v ${PWD}/:/workspace wmh:latest /bin/bash.

6. Run command cd /workspace.

For training (Needs scanner datafile)

1. Run command cd /2_5DWMHSEG_TRAIN.

2. Change the conf/config.yaml file.

3. Run command python main.py.

For prediction (Needs weight file)

1. Run command cd /2_5DSEG_PRED.

2. Change config.yaml file.

1. Run command python main_prediction.py.

Test results

DSC:  0.6808092008717108 +- 0.1749727617385947
Hausdorff:  9.751399615092273 +- 9.279947874717204
Recall:  0.7914909337086632 +- 0.16611485702407042
F1:  0.5269392661792544 +- 0.1548622287067465
AVD:  55.08561295745984 +- 54.99577215339091

The other models used:

nnU-Net -> https://catalog.ngc.nvidia.com/orgs/nvidia/resources/nnunet_for_pytorch (Achieved best results)

Deep Bayesian Networks (Hypermapp3r) -> https://hypermapp3r.readthedocs.io/en/latest/