Awesome

HM_DataAug (Team name: opossum)

Histogram Matching for Domain Adaptation: Solution to M&Ms 2020

The authors of this paper declare that the segmentation method they implemented for participation in the M&Ms challenge has not used any pre-trained models nor additional MRI datasets other than those provided by the organizers.

Prepare data

• Clone this repo. and put testing cases in mnms

• Copy and rename the the end-diastole (ED) and end-systole (ES) phases data to a single folder test_data, by running

python prepare_data.py

Prepare Trained Models and code

• Download (also extract) trained models and put them in V2_nnUNet/nnUNet/nnunet

• Install nnUNet

  cd V2_nnUNet/nnUNet

  pip install -e .

  cd nnunet

Please use our modified nnUNet in this repo. rather than the official nnUNet.

All the following commands should be run in V2_nnUNet/nnUNet/nnunet

Solution 1: 3D Best Model

Run

nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/solution1_output -m 3d_fullres -t Task601_BestHMAug --save_npz

Segmentation Results will be in HM_DataAug/mnms/solution1_output.

Solution 2: 3D Final Model

nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/solution2_output -m 3d_fullres -t Task602_HMAugMMS --save_npz

Segmentation Results will be in HM_DataAug/mnms/solution2_output.

Solution 3: 2D-3D Best Model Ensemble

nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/temp_solution3 -m 2d -t Task601_BestHMAug --save_npz

nnUNet_ensemble -f ../../../mnms/solution1_output ../../../mnms/temp_solution3 -o ../../../mnms/solution3_output

Segmentation Results will be in HM_DataAug/mnms/solution3_output.

Solution 4: 2D-3D Final Model Ensemble

nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/temp_solution4 -m 2d -t Task602_HMAugMMS --save_npz

nnUNet_ensemble -f ../../../mnms/solution2_output ../../../mnms/temp_solution4 -o ../../../mnms/solution4_output

Segmentation Results will be in HM_DataAug/mnms/solution4_output.

Solution 5: 2D-3D All Model Ensemble

nnUNet_ensemble -f ../../../mnms/solution1_output ../../../mnms/solution2_output ../../../mnms/temp_solution3 ../../../mnms/temp_solution4 -o ../../../mnms/solution5_output

Segmentation Results will be in HM_DataAug/mnms/solution5_output.

A combo for the 5 solutions

Obtaining the segmentation results of the 5 solutions in a container rather than creating 5 containers.

# generate softmax predictions
nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/solution1_output -m 3d_fullres -t Task601_BestHMAug --save_npz
nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/solution2_output -m 3d_fullres -t Task602_HMAugMMS  --save_npz
nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/temp_solution3 -m 2d -t Task601_BestHMAug --save_npz
nnUNet_predict -i ../../../mnms/test_data -o ../../../mnms/temp_solution4 -m 2d -t Task602_HMAugMMS --save_npz

# ensemble
nnUNet_ensemble -f ../../../mnms/solution1_output ../../../mnms/temp_solution3 -o ../../../mnms/solution3_output
nnUNet_ensemble -f ../../../mnms/solution2_output ../../../mnms/temp_solution4 -o ../../../mnms/solution4_output
nnUNet_ensemble -f ../../../mnms/solution1_output ../../../mnms/solution2_output ../../../mnms/temp_solution3 ../../../mnms/temp_solution4 -o ../../../mnms/solution5_output

Clean Results

• cd ../../../mnms
• rm -rf temp*
• rm solution1_output/*.npz
• rm solution1_output/*.pkl
• rm solution2_output/*.npz
• rm solution2_output/*.pkl