Artificial Intelligence–based Quantification of Pleural Plaque Volume and Association With Lung Function in Asbestos-exposed Patients 💻 🔎

Open sourcing segmentation models for pleural plaques in thoracic CT scans following our publication

Managed by Kevin Groot Lipman (k.groot.lipman[at]nki[dot]nl, k.b.w.grootlipman[at]gmail[dot]com)

1. Setting up the environment 🌳

Install miniconda3: https://docs.conda.io/en/latest/miniconda.html or

curl -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

Create and activate a conda environment with Python

conda create --name torch python=3.11
conda activate torch

Install latest PyTorch version with conda and cuda support: https://pytorch.org/get-started/locally/

Only then install nnUNet

pip install nnunet

Please read the updated nnUNet repo for instructions: https://github.com/MIC-DKFZ/nnUNet

2. Preprocess your DICOM files

2.1 Convert DICOM to Nifti

nnUNet models do not accept DICOM files. Therefore, we must first convert our selected CT scans to the Nifti format. Install conversion software:

conda activate torchv2
git clone https://github.com/JoostJM/nrrdify.git
cd nrrdify
pip install .

Run the conversion script:

nrrdify DICOM_FOLDER --out OUTPUT_FOLDER --format nii.gz

If DICOM scans are not being converted, it is most likely due to inconsistent slice spacing. This is intended behavior since we don't process CT scans with inconsistent 'gaps'.

3. How to run the segmentation model 🚀

3.1 Download the weights

Download models/TaskXXX_PP.zip Replace XXX with the task number you want to run (200-204)

3.2 Set your environment variables

See: https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/set_environment_variables.md

3.3 Running inference

• Unzip TaskXXX_PP.zip in your nnUNet_results folder defined in 3.2
• Run nnUNet_predict TaskXXX_PP -i INPUT_FOLDER -o OUTPUT_FOLDER -f all -c 3d_fullres

The script will run and the AI model will output binary segmentation masks in 'OUTPUT_FOLDER'.

4. Inspection and correction output

4.1 Install Slicer

• Download Slicer 5.3 at https://download.slicer.org/
• Start Slicer, go to 'File' -> 'Add data' -> 'Choose File(s) to Add'. Select the Nifti CT scan (with _0000.nii.gz at the end) and the corresponding segmentation (same name without _0000).
• Before you click 'OK', ensure that the CT scan is loaded as 'Volume' (in column 'Description') and the segmentation is loaded as 'Segmentation'.
• The CT scan and corresponding segmentation is now visible
• Click on 'Welcome to Slicer', then 'Segment Editor'.
• Here are several tools to correct the segmentation; the most common ones are the 'Paint' and the 'Erase' tool on the left.

5. Troubleshooting 🔨

• If nnUNet cannot find the CT scans in your folder, you probably forgot to append the _0000 before .nii.gz
• Other problems, please don't hesitate to reach out

6. Acknowledgments

• This project heavily relied on nnUNet; big shoutout to them: https://github.com/MIC-DKFZ/nnUNet