This repository contains all the code necessary to automatically extract Perfusion Map Descriptors (PMDs) from DSC-MRI images, as described in our paper:

Discriminating Distal Ischemic Stroke from Seizure-Induced Stroke Mimics Using MR Perfusion Imaging

The repository also includes scripts used for the statistical and machine learning analysis presented in the paper.

The PMD extraction pipeline is fully automated. Running extract_PMDs.ps1 will process an entire dataset, extracting perfusion features in a reproducible manner.

• Operating system: Windows
• CPU: All steps can run on standard CPUs (e.g., Intel Core i5). Though if a GPU is available you can speed up the skull stripping.

For each patient, the following data is required:

1. DSC-MRI e.g. 2D-EPI with gadolinium-based contrast

Also needed:

1. Brain Atlas: e.g. The Harvard-Oxford cortical and subcortical atlas from FSL.
2. Atlas Template: A corresponding template for the brain atlas.

Optional:

1. Other image types e.g. DWI (T1 B1000) scan.

All images must be in .nii.gz format.

The storage location of the brain atlas and its corresponding template is flexible. You can specify their paths at the top of the extract_PMDs.ps1 script, along with the dataset location.

Ensure your data is organized as follows:

dataset/
├── group_1/
│   ├── patient_1/
│   │   └── study_1/
│   │       ├── perf_nii/
│   │       │   └── perf.nii.gz
│   │       └── dwi_nii/
│   │           └── b1000.nii.gz
│   ├── patient_2/
│   │   └── study_1/
│   │       ├── perf_nii/
│   │       │   └── perf.nii.gz
│   │       └── dwi_nii/
│   │           └── b1000.nii.gz
│   └── patient_3/
│       └── ...
├── group_2/
│   └── patient_1/
│       └── study_1/
│           └── ...
└── group_3/
    └── ...

Each level in the hierarchy is described as follows:

• group_X/: A group of patients (e.g., treatment cohort, control group).
• patient_X/: A unique patient identifier within the group.
• study_X/: A study session for a patient (only add 1 study per patient).
• perf_nii/: Folder containing perfusion imaging data.
  • perf.nii.gz: Perfusion NIfTI file.
• dwi_nii/: Folder containing diffusion-weighted imaging data. (This could be any image type. If we can register an atlas to it, we can get PMDs for it.)
  • b1000.nii.gz: DWI NIfTI file.

This pipeline uses two separate Conda environments:

We use a Windows-compatible fork of HD-BET by sh-shahrokhi. This installs hd-bet as a callable command in powershell/cmd, given the correct venv is activated.

Installation:

1. Create and activate a venv
   conda create --name hdbet python==3.11
   conda activate hdbet

2. Clone the forked repo
   git clone https://github.com/MIC-DKFZ/HD-BET

3. Install requirements
   cd HD-BET (or to wherever setup.py lives)
   pip3 install -e .

4. Test installation
   hd-bet --help

Installation:

1. Create and activate a venv
   conda create --name PMD python==3.11
   conda activate PMD

2. Clone repo and install requirements
   pip3 install git+https://github.com/Marijn311/PMD_extraction_and_analysis.git

Once installation is complete and data is prepared:

1. Open PowerShell

2. Run the extract_PMDS.ps1 script
   cd "location of this repo where the extract_PMDs.ps1 lives"
   ./extract_PMDs.ps1

All patients will be processed automatically. For each patient a PMDs.xlsx file will be created.

Warnings are logged automatically:
Motion artifacts or unclear bolus → Logged to perfusion_tool/error_logs.txt
Near zero reference value or many NaN/Inf values → Logged to error_logs.txt

Once the PMDs.xlsx excel file has been created for every patients in the dataset, you can run analysis/create_pmd_dataset.py.
This script loads all individual PMD .xlsx files and merges them into a single .nc file using xarray.

You can generate the distribution and asymmetry analysis by running analysis/statistical_analysis.py on the .nc file.
You can train the logistic regression model using the files in analysis/machine_learning.