Segmentation

We used nnUnet v1, which at that moment was the latest one. Currently, there is an improved version nnUnet v2

Installation nnUNet

Install nnUNet from link. Install it as integrative framework (-e) because we will need to modify some files later:

git clone https://github.com/MIC-DKFZ/nnUNet.git
cd nnUNet
pip install -e .

Data conversion

We need to comply with the requirements of the data provided in link

Important notes:

• Please make your custom task ids start at 500 to ensure that there will be no conflicts with downloaded pretrained models!!! (IDs cannot exceed 999)

• The image files can have any scalar pixel type. The label files must contain segmentation maps that contain consecutive integers, starting with 0: 0, 1, 2, 3, ... num_labels. 0 is considered background.

Task120_Massachusetts_RoadSegm.py example (link) is used as a template to convert the data from 2d to pseudo-3d.

Note: Task120_Massachusetts_RoadSegm.py deals the situation when the output(labels) are only available for some of the images by doing a join (Line 63/64)

Procedure

In order to use the Task120_Massachusetts_RoadSegm.py. First, change the structure of your data to and convert labelled image to one channel:

data/
├── training/
│   ├── input/
│   │   ├── img1
│   │   ├── img2
│   │   ├── img3
│   │       (...)
│   └── output/
│       ├── img1
│       ├── img2
│       ├── img3
│           (...)
└── testing/
    ├── input/
    │   ├── img1
    │   ├── img2
    │   ├── img3
    │       (...)
    └── output/
        ├── img1
        ├── img2
        ├── img3
            (...)

Note: the label and image names should be the same in input/ and output/ directories.

*convert_label_1_channel.py: it converts the label images to 1 channel

*structure_data.py: it convert the data to the structure necesesary to be used by Task120_Massachusetts_RoadSegm.py

python convert_label_1_channel.py -j convert_label_1_channel_5000_v3.json

python stucture_data.py -j structure_data_5000_v3.json

Task500_BloodVessel.py and Task501_BloodVessel_extra_test.py will make the tranformation from png to nifti files to data. Thse files are based on Task120_Massachusetts_RoadSegm.py.

These files can be found in /scripts/conversion_to_nifti.

In order to be able to run Task500_BloodVessel.py, the following variables need to be defined, according to link.

export nnUNet_raw_data_base="/home/pcallec/nnUNet_blood-vessel/results/nnUNet_raw_data_base"
export nnUNet_preprocessed="/home/pcallec/nnUNet_blood-vessel/results/nnUNet_preprocessed"
export RESULTS_FOLDER="/home/pcallec/nnUNet_blood-vessel/results/nnUNet_trained_models"

Experiment planning and preprocessing

nnUNet_plan_and_preprocess -t 500 --verify_dataset_integrity -pl3d None

According to library repository:

Running nnUNet_plan_and_preprocess will populate your folder with preprocessed data. You will find the output in nnUNet_preprocessed/TaskXXX_MYTASK. nnUNet_plan_and_preprocess creates subfolders with preprocessed data for the 2D U-Net as well as all applicable 3D U-Nets. It will also create 'plans' files (with the ending.pkl) for the 2D and 3D configurations. These files contain the generated segmentation pipeline configuration and will be read by the nnUNetTrainer (see below). Note that the preprocessed data folder only contains the training cases. The test images are not preprocessed (they are not looked at at all!). Their preprocessing happens on the fly during inference.

--verify_dataset_integrity should be run at least for the first time the command is run on a given dataset. This will execute some checks on the dataset to ensure that it is compatible with nnU-Net. If this check has passed once, it can be omitted in future runs. If you adhere to the dataset conversion guide (see above) then this should pass without

Run training

nnUNet_train 2d nnUNetTrainerV2 500 FOLD

Scripts can be seen in: /scripts/scripts_nnunet/commands_all.sh

Note: there was error nnU-Net training: Error: mmap length is greater than file size and EOFError. According to (Link)[https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/common_problems_and_solutions.md#nnu-net-training-error-mmap-length-is-greater-than-file-size-and-eoferror], the files .npy should be deleted.

Run inference on test data

From documentation:

Note that per default, inference will be done with all available folds. We very strongly recommend you use all 5 folds. Thus, all 5 folds must have been trained prior to running inference. The list of available folds nnU-Net found will be printed at the start of the inference.

Therefore prediction is using ensemble of the 5 folds:

nnUNet_find_best_configuration -m 2d -t 500

CUDA_VISIBLE_DEVICES=1 nnUNet_predict -i /home/pcallec/nnUNet_blood-vessel/results/nnUNet_raw_data_base/nnUNet_raw_data/Task500_BloodVessel/imagesTs -o /home/pcallec/nnUNet_blood-vessel/results/test_pred/v3 -t 500 -m 2d 2>&1 | tee output_pred_v3.txt

(time CUDA_VISIBLE_DEVICES=1 nnUNet_predict -i /home/pcallec/nnUNet_blood-vessel/results/nnUNet_raw_data_base/nnUNet_raw_data/Task500_BloodVessel/imagesTs -o /home/pcallec/nnUNet_blood-vessel/results/test_pred/v3_rep -t 500 -m 2d ) 2>&1 | tee output_pred_v3_rep.txt

Extra test

python convert_label_1_channel_extra_test.py -j convert_label_1_channel_extra_test.json
python structure_data_extra_test.py -j structure_data_extra_test.json

python Task501_BloodVessel_extra_test.py

nnUNet_plan_and_preprocess -t 501 --verify_dataset_integrity -pl3d None

CUDA_VISIBLE_DEVICES=1 
nnUNet_predict -i /home/pcallec/nnUNet_blood-vessel/results/nnUNet_raw_data_base/nnUNet_raw_data/Task501_BloodVessel/imagesTr/ -o /home/pcallec/nnUNet_blood-vessel/results/extra_test_pred/ -t 501 -m 2d 2>&1 | tee output_pred_extra_test.txt

Analysis of results

• Processing_results_blood-vessel_metrics_v3.ipynb: Process metrics for validation and test
• Processing_results_blood-vessel_v3.ipynb: Generation superimposed images for validation and test
• Processing_results_nn-UNet_blood-vessel_extraTest.ipynb: Generates superimosed images and calculate metrics for extra test.