bavaria-quebec: A Shell repository from sct-pipeline

Cooperation between NeuroPoly (Polytechnique Montreal,Quebec) and TUM (Munich, Bavaria)

This project aims to investigate both brain and spine MRI scans for Multiple Sclerosis Research.

Note:

As all studied data are kept privately, this repository does not (and should not) feature patient-related data and/or scans. These are shared in a privacy-preserving manner between the NeuroPoly and TUM labs.

Data collection and organization
Analysis Pipeline

Data collection and organization

Background Information

In this project, we use the Spinal Cord Toolbox to analyze both, brain and spinal cord for multiple sclerosis research questions. The MS brain database has been curated by various PhD students of the lab at TUM and various pipelines, such as CAT12, LST and Freesurfer have been employed to analyze the data. Similarly, this project aims to create a spinal cord mri database to provide the possibility of large-scale analysis of spinal cord lesions, and the joined analysis with the brain.

BIDS compliance

Originally, the datasets and cohorts have not complied with the BIDS standard. The conversion to BIDS is cohort-/dataset specific, i.e. conversion scripts have to be adapted to the pecularities of the data (naming convertions, dataformats etc.). We attempt to gather all of the scripts related to the conversion in BIDS Tools and describe the newly introduced BIDS conventions that are relevant to people working on this project in this repository.

Brain Dataset

The brain database consists of T1w, T2w, Flair Scans of the brain, scanned with isotropic resolution.

Spinal Cord Dataset

In contrast to the brain database, the imaging of spinal cord is much more complex, and requires a more sophisticated protocol.

In a first step, the clinical practitioners perform a low-resolution scout scan to place the patient in the MRI scanner in order to locate the spinal cord. Subsequently, the spinal cord is imaged in two to three sagittal stacks. Taking the sagittaly scanned images into consideration, axial imaging of the spinal cord is planned and conducted. Axial images are obtained in order to verify lesions from a different perspective and to detect (smaller) lesions that are only visible in axial slices (Breckwoldt et al., Weier et al.).

Finally, the dataset consists of a large-scale, multi-scanner patient cohort comprising multiple (longitudinal) sessions each consisting of multiple axially and sagitally aquired 2D MRI scans for each patient. All results are stored as DICOMs to the clinical PACS severs, and are subsequently extracted as one NIFTI+JSON per scan.

BIDS conform Raw Dataset Structure

To establish a basis we convert the database to a BIDS-compliant database facilitating the integration with existing tools. For this we parse the existing dataset, and infer scan orientation, modality and other important properties (image origin and content) from the image data (and not the heterogenous json sidecars). The newly generated, raw (i.e. 'unstitched') BIDS compliant database adheres to the following structure:

├── CHANGES.md
├── code
│   └── sct-v5.6
├── dataset_description.json
├── LICENSE
├── participants.json
├── participants.tsv
├── README.md
└── sub-m123456
    ├── ses-20220101
        ├── anat
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-1_T2w.json
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-1_T2w.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-2_T2w.json
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-2_T2w.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-3_T2w.json
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-3_T2w.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-1_dseg.json
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-1_dseg.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-2_dseg.json
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-2_dseg.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-3_dseg.json
        │   ├── sub-m123456_ses-20220101_acq-ax_chunk-3_dseg.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-sag_chunk-1_T2w.json
        │   ├── sub-m123456_ses-20220101_acq-sag_chunk-1_T2w.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-sag_chunk-2_T2w.json
        │   ├── sub-m123456_ses-20220101_acq-sag_chunk-2_T2w.nii.gz
        │   ├── sub-m123456_ses-20220101_acq-sag_chunk-3_T2w.json
        │   └── sub-m123456_ses-20220101_acq-sag_chunk-3_T2w.nii.gz
    │   └── sub-m123456_ses-20220101_scans.tsv
    └── ses-20220202
        ├── anat
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-1_T2w.json
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-1_T2w.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-2_T2w.json
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-2_T2w.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-3_T2w.json
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-3_T2w.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-1_dseg.json
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-1_dseg.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-2_dseg.json
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-2_dseg.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-3_dseg.json
        │   ├── sub-m123456_ses-20220202_acq-ax_chunk-3_dseg.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-sag_chunk-1_T2w.json
        │   ├── sub-m123456_ses-20220202_acq-sag_chunk-1_T2w.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-sag_chunk-2_T2w.json
        │   ├── sub-m123456_ses-20220202_acq-sag_chunk-2_T2w.nii.gz
        │   ├── sub-m123456_ses-20220202_acq-sag_chunk-3_T2w.json
        │   └── sub-m123456_ses-20220202_acq-sag_chunk-3_T2w.nii.gz
        └── sub-m123456_ses-20220202_scans.tsv

Analysis Pipeline

General Information

Pipeline scripts can be found in directory.

Preprocessing

To use SCT's capabilities w.r.t. statistical analysis and registration to the PAM50 template, it is necessary to stitch the axial and sagittal slices to whole-spine images. For example, we stitch all axial chunks sub-{id}_ses-{id}_chunk-{no}_T2w.nii.gz to a whole spine image sub-{id}_ses-{id}_T2w.nii.gz.

For this, we evaluated multiple stitching algorithms. After thorough examination of the results, we decided to use the approach by Lavdas et al., which is now integrated into SCT v.5.8.

Moreover, in a previous study by Bussas et al., multiple sclerosis lesion segmentation labels were created by two medical doctors on a chunk basis. To process them using SCT, the lesion masks are stitched together as well, and binarized to {0,1}. This is necessary due to stitching's resampling stage.

To obtain the stitched database, please run:

sct_run_batch -script preprocess_raw_dataset.sh -path-data /path/to/db -path-output result

Spinal Cord Processing

To use SCT on the spinal cord database we implement a bash-based multi-subject script for the entire workflow in SCT.

Currently, it consists of the following steps:

Segmentation of spinal cord
Labeling of vertebrae
Computation of CSA in cervical spine
Lesion Segmentation
Computation of Dice scores for manual and DL-segmented lesions.

Link to scripts.

Brain CSA Measurement

To use SCT on a BIDS compliant brain database (i.e. on the cervical spine regions), we implement a bash-based multi-subject script for the entire workflow in SCT.