Skip to content

Latest commit

 

History

History
144 lines (100 loc) · 4.59 KB

README.md

File metadata and controls

144 lines (100 loc) · 4.59 KB
Raw

The nnUNet pipeline

Here, we detail all the necessary steps we made to train and use the nnUNetv2 framework with the CanProCo dataset. We show how to:

  • set-up the environment
  • preprocess the data
  • train the model
  • perform inference
  • evaluate the predictions
  • compare the performances of different models

Use this steps if you want to retrain the entire model on the CanProCo dataset and evaluate its performance.

Installation instructions

Installation of Anima metrics

Installation of

cd ~
mkdir anima/
cd anima/
wget -q https://github.com/Inria-Empenn/Anima-Public/releases/download/v4.2/Anima-macOS-4.2.zip # for MACOS
unzip Anima-macOS-4.2.zip
rm Anima-macOS-4.2.zip
git lfs install
git clone --depth 1 https://github.com/Inria-Visages/Anima-Scripts-Public.git
git clone --depth 1 https://github.com/Inria-Visages/Anima-Scripts-Data-Public.git

Configure directories

cd ~
mkdir .anima/
touch .anima/config.txt

echo "[anima-scripts]" >> .anima/config.txt
echo "anima = ${HOME}/anima/Anima-Binaries-4.2/" >> .anima/config.txt
echo "anima-scripts-public-root = ${HOME}/anima/Anima-Scripts-Public/" >> .anima/config.txt
echo "extra-data-root = ${HOME}/anima/Anima-Scripts-Data-Public/" >> .anima/config.txt

Installation of required libraries

Create a virtual invironment:

conda create -n venv_nnunet python=3.9

Activate the environment with the following command:

conda activate venv_nnunet

To install required libraries to train an nnUNet v2:

pip install -r requirements_nnunet.txt

Install SpinalCordToolbox 6.0 :

Installation link : https://spinalcordtoolbox.com/user_section/installation.html

Data preparation

Create the following folders:

mkdir nnUNet_raw
mkdir nnUNet_preprocessed
mkdir nnUNet_results

We are training a region-based nnUNet taking an image from contrasts PSIR or STIR and creating a mask with 0=background, 1=spinal cord and 2=MS lesion.

Convert the data to the nnUNet format :

python convert_BIDS_to_nnunet.py --path-data /path/to/BIDS/dataset --path-out /path/to/nnUNet_raw --taskname TASK-NAME --tasknumber DATASET-ID  --contrasts PSIR,STIR --test-ratio XX --time-point ses-XX --type training --exclude-file /path/to/exclude_file.yml

Note

The test ratio is 0.2 for 20% (train ratio is therefore 80%). For M0 images, the time point is ses-M0.

To mutliply PSIR images by -1 before training and convert the data to the nnUNet format :

python convert_BIDS_to_nnunet_with_mul_PSIR.py --path-data /path/to/BIDS/dataset --path-out /path/to/nnUNet_raw --taskname TASK-NAME --tasknumber DATASET-ID  --contrasts PSIR,STIR --test-ratio XX --time-point ses-XX --type training --exclude-file /path/to/exclude_file.yml

Model training

Before training the model, nnU-Net performs data preprocessing and checks the integrity of the dataset:

export nnUNet_raw="/path/to/nnUNet_raw"
export nnUNet_preprocessed="/path/to/nnUNet_preprocessed"
export nnUNet_results="/path/to/nnUNet_results"

nnUNetv2_plan_and_preprocess -d DATASET-ID --verify_dataset_integrity

You will get the configuration plan for all four configurations (2d, 3d_fullres, 3d_lowres, 3d_cascade_fullres).

To train the model, use the following command:

CUDA_VISIBLE_DEVICES=XXX nnUNetv2_train DATASET-ID CONFIG FOLD --npz

Note

Example for Dataset 101, on 2d config on fold 0: CUDA_VISIBLE_DEVICES=2 nnUNetv2_train 101 2d 0 --npz

Model inference

Convert data to nnUNet format for inference using convert_BIDS_to_nnunet.py or convert_BIDS_to_nnunet_with_mul_PSIR.py with --type=inference.

Then perform inference:

CUDA_VISIBLE_DEVICES=XXX nnUNetv2_predict -i /path/to/image/folder -o /path/to/predictions -d DATASET_ID -c CONFIG --save_probabilities -chk checkpoint_best.pth -f FOLD

Inference evaluation

First, convert the predictions back the BIDS format ; this only keeps the lesion segmentation and discards spinal cord segmentation :

python convert_predictions_to_BIDS.py --pred-folder /path/to/predictions --out-folder /path/to/output/folder --conversion-dict /path/to/conversion/dict

If you are converting predictions which are not the evaluations set from nnUNet, use the flag --not-imageTs.

Then, you can evaluate the lesion prediction with Anima metrics

python evaluate_lesion_seg_prediction.py --pred-folder path/to/predictions --dataset path/to/dataset --animaPath path/to/animaSegPerfAnalyzer --output-folder path/to/output_folder

Evaluation analysis

The following Notebook nnUNet_inference_analysis.ipynb was used to perform analysis of the nnUNet segmentations. To use it, change the path to the csv files.