lesion_ms

MS lesion segmentation on spinal cord MRI images

This segmentation model for spinal cord MS lesion segmentation uses a 3D U-Net architecture. It outputs a binary segmentation of MS lesions. The model was trained and evaluated on a large-scale dataset comprising 4,428 annotated images from 1,849 persons with MS across 23 imaging centers. The dataset included images acquired on GE, Siemens or Philips MRI systems, at 1.5T, 3T or 7T, using six distinct MRI contrasts: T2w (n=3,060), T2*w (n=548), PSIR (n=363), UNIT1 (reconstructed uniform image from MP2RAGE sequence, n=343), STIR (n=92), and T1w (n=22), and spans 2D axial (n=2,895), 2D sagittal (n=1,160), and 3D (n=373) acquisition planes. The field-of-view coverage varied across sites (brain and upper SC, or SC only). Image resolution exhibited high variability, with an average (± standard deviation) of 1.10±1.13 x 0.51±0.24 x 3.27±1.95 mm³ reported in “RPI-” orientation

Reference

@article{doi:10.1177/13524585261427333,
   author = {Pierre-Louis Benveniste and Laurent Létourneau-Guillon and David Araujo and Lydia Chougar and Dumitru Fetco and Masaaki Hori and Kouhei Kamiya and Steven Messina and Charidimos Tsagkas and Bertrand Audoin and Rohit Bakshi and Elise Bannier and Daniel Blezek and Jean-Christophe Brisset and Virginie Callot and Erik Charlson and Michelle Chen and Olga Ciccarelli and Sarah Demortière and Gilles Edan and Massimo Filippi and Tobias Granberg and Cristina Granziera and Christopher C. Hemond and B. Mark Keegan and Anne Kerbrat and Jan Kirschke and Shannon Kolind and Pierre Labauge and Lisa Eunyoung Lee and Yaou Liu and Caterina Mainero and Julian McGinnis and Nilser Laines Medina and Mark Mühlau and Govind Nair and Kristin P. O’Grady and Jiwon Oh and Russell Ouellette and Alexandre Prat and Daniel S. Reich and Maria A. Rocca and Timothy M. Shepherd and Seth A. Smith and Leszek Stawiarz and Jason Talbott and Roger Tam and Shahamat Tauhid and Anthony Traboulsee and Constantina Andrada Treaba and Paola Valsasina and Zachary Vavasour and Marios Yiannakas and Hervé Lombaert and Julien Cohen-Adad},
   title ={Generalizable spinal cord multiple sclerosis lesion segmentation across MRI contrasts, protocols, and centers},
   journal = {Multiple Sclerosis Journal},
   volume = {0},
   number = {0},
   pages = {13524585261427333},
   year = {2026},
   doi = {10.1177/13524585261427333},
   note ={PMID: 42028790},
   URL = {https://doi.org/10.1177/13524585261427333},
   eprint = {https://doi.org/10.1177/13524585261427333},
   abstract = { Background/Objectives: Characterizing spinal cord multiple sclerosis (MS) lesions in MRI is critical for diagnosis, monitoring, and treatment evaluation. However, current automated approaches for lesion detection and segmentation are typically designed for specific MRI contrasts or acquisition sites, limiting their generalizability in real-world clinical settings where imaging protocols vary widely. This work proposes a robust multi-site, multi-contrast segmentation framework for spinal cord lesions.Methods: The segmentation model was trained and evaluated on a large-scale dataset comprising 4428 annotated images from 1849 persons with MS across 23 imaging centers, encompassing six MRI contrasts (T1w, T2w, T2*w, PSIR, STIR, and UNIT1) acquired at 1.5 tesla (T), 3 T, and 7 T.Results: Likert-type assessment performed by neuroradiologist ratings demonstrated superior generalization of the model compared to existing contrast-specific pipelines (p < 0.01). Additional experiments evaluated robustness across spinal levels, acquisition resolutions, binarization thresholds, and quantitative evaluation on external labeled datasets.Conclusions: The proposed model can achieve accurate and reliable spinal cord MS lesion segmentation across heterogeneous MRI data, addressing a key barrier to clinical translation. The model is available in the Spinal Cord Toolbox v7.2 and higher.Code repository: https://github.com/ivadomed/seg-sc-ms-lesion-multicontrast }
}

Project URL: https://github.com/ivadomed/ms-lesion-agnostic

usage: sct_deepseg lesion_ms [-i <file> [<file> ...]] [-o <str>] [-install]
                             [-custom-url CUSTOM_URL [CUSTOM_URL ...]]
                             [-largest {0,1}] [-fill-holes {0,1}]
                             [-remove-small REMOVE_SMALL [REMOVE_SMALL ...]]
                             [-qc <folder>] [-qc-dataset <str>]
                             [-qc-subject <str>] [-qc-plane <str>]
                             [-qc-seg <file>] [-h] [-v <int>] [-r {0,1}]
                             [-test-time-aug] [-soft-ms-lesion] [-single-fold]

INPUT/OUTPUT

-i

Image filename(s) to segment. If segmenting multiple files, separate filenames with a space.

-o

Output file name. The chosen filename will be used as a base name, and model-specific suffixes will be added to the end depending on the type of output (e.g. ‘_cord.nii.gz’, ‘_gm.nii.gz’, etc.).

TASKS

-install

Install models that are required for specified task.

Default: False

-custom-url

URL(s) pointing to the .zip asset for a model release. This option can be used with -install to install a specific version of a model. To use this option, navigate to the ‘Releases’ page of the model, find release you wish to install, and right-click + copy the URL of the .zip listed under ‘Assets’. Example: sct_deepseg lesion_ms -install -custom-url CUSTOM_URL sct_deepseg lesion_ms -i t2.nii.gz

PARAMETERS

-largest

Possible choices: 0, 1

Keep the largest connected object from each output segmentation; if not set, all objects are kept.

Default: 0

-fill-holes

Possible choices: 0, 1

If set, small holes in the segmentation will be filled in automatically.

Default: 0

-remove-small

Minimal object size to keep with unit (mm3 or vox). A single value can be provided or one value per prediction class. Single value example: 1mm3, 5vox. Multiple values example: 10 20 10vox (remove objects smaller than 10 voxels for class 1 and 3, and smaller than 20 voxels for class 2).

-test-time-aug

Perform test-time augmentation (TTA) by flipping the input image along all axes and averaging the resulting predictions. Note: The time it takes to run the model will increase due to the additional predictions.

Default: False

-soft-ms-lesion

If set, the model will output a soft segmentation (i.e. probability map) instead of a binary segmentation.

Default: False

-single-fold

If set, only 1 fold will be used for inference instead of the full 5-fold ensemble. This will speed up inference, but may reduce segmentation quality.

Default: False

MISC ARGUMENTS

-qc

The path where the quality control generated content will be saved.

-qc-dataset

If provided, this string will be mentioned in the QC report as the dataset the process was run on.

-qc-subject

If provided, this string will be mentioned in the QC report as the subject the process was run on.

-qc-plane

Possible choices: Axial, Sagittal

Plane of the output QC. If Sagittal, it is highly recommended to provide the -qc-seg option, as it will ensure the output QC is cropped to a reasonable field of view.

Default: 'Axial'

-qc-seg

Segmentation file to use for cropping the QC. This option is useful when you want to QC a region that is different from the output segmentation. For example, it might be useful to provide a dilated cord segmentation to expand the QC field of view.

If -qc-seg is not provided, the default behavior will depend on the value of -qc-plane:

• ‘Axial’: Without ‘-qc-seg’, a sensible crop radius between 15-40 vox will be automatically used, depending on the resolution and segmentation type.

• ‘Sagittal’: Without ‘-qc-seg’, the full image will be displayed by default. (For very large images, this may cause a crash, so using -qc-seg is highly recommended.)

-v

Possible choices: 0, 1, 2

Verbosity. 0: Display only errors/warnings, 1: Errors/warnings + info messages, 2: Debug mode.

Default: 1

-r

Possible choices: 0, 1

Remove temporary files.

Default: 1