SCISeg: Automatic Segmentation of T2-weighted Intramedullary Lesions in Spinal Cord Injury

This repository contains the code for deep learning-based segmentation of the spinal cord and intramedually lesions in spinal cord injury (SCI). The code is based on the nnUNetv2 framework.

Model Overview

The model was trained on raw T2-weighted images of SCI patients from seven sites comprising traumatic (acute preoperative, intermediate, chronic) and non-traumatic (ischemic SCI and degenerative cervical myelopathy, DCM) SCI lesions. The data included images with heterogenous resolutions (axial/sagittal/isotropic) and scanner strengths (1T/1.5T/3T). To ensure uniformity across sites, all images were initially re-oriented to RPI. Given an input image, the model is able to segment both the lesion and the spinal cord.

Updates

2024-07-24

• We have released SCIsegV2: A Universal Model for Intramedullary Lesion Segmentation in SCI. The new model is trained on a larger cohort covering both traumatic and non-traumatic SCI lesions. SCIsegV2 is available as part of the SCT via the sct_deepseg function; see the installation instructions below.
• The computation of midsagittal tissue bridges is now fully-automated and powered by SCIsegV2. The automatic computation of tissue bridges is available via the sct_analyze_lesion function as part of SCT v6.4 and higher.
• We have moved away from ANIMA metrics and have started to use MetricsReloaded instead. This wrapper script is used to compute metrics and an internal fork of the package is maintained here.
• The computation of midsagittal tissue bridges is now fully-automated and powered by SCIsegV2. The automatic computation of tissue bridges is available via the sct_analyze_lesion function as part of SCT v6.4 and higher.

Using SCIsegV2

Install dependencies

• Spinal Cord Toolbox (SCT) v6.4 or higher -- follow the installation instructions here
• conda
• Python (v3.9)

Once the dependencies are installed, download the latest SCIseg model:

sct_deepseg -install-task seg_sc_lesion_t2w_sci

Getting the lesion and spinal cord segmentation

To segment a single image, run the following command:

sct_deepseg -i <INPUT> -task seg_sc_lesion_t2w_sci

For example:

sct_deepseg -i sub-001_T2w.nii.gz -task seg_sc_lesion_t2w_sci

The outputs will be saved in the same directory as the input image, with the suffix _lesion_seg.nii.gz for the lesion and _sc_seg.nii.gz for the spinal cord.

Automatic measurements of midsagittal tissue bridges

This new functionality is available via SCT's sct_analyze_lesion. The function computes the midsagittal tissue bridges and outputs the ventral and dorsal tissue bridges.

sct_analyze_lesion -m <SUBJECT>_lesion_seg.nii.gz -s <SUBJECT>_sc_seg.nii.gz

Citation Info

If you find this work and/or code useful for your research, please cite our papers:

@article {Naga Karthik2024.01.03.24300794,
	author = {Enamundram Naga Karthik* and Jan Valosek* and Andrew C. Smith and Dario Pfyffer and Simon Schading-Sassenhausen and Lynn Farner and Kenneth A. Weber II and Patrick Freund and Julien Cohen-Adad},
	title = {SCIseg: Automatic Segmentation of T2-weighted Intramedullary Lesions in Spinal Cord Injury},
	elocation-id = {2024.01.03.24300794},
	year = {2024},
	doi = {10.1101/2024.01.03.24300794},
	publisher = {Cold Spring Harbor Laboratory Press},
	URL = {https://www.medrxiv.org/content/early/2024/04/21/2024.01.03.24300794},
	eprint = {https://www.medrxiv.org/content/early/2024/04/21/2024.01.03.24300794.full.pdf},
	journal = {medRxiv},
	note = {*Shared first authorship}
}

@article {karthik2024scisegv2universaltoolsegmentation,
      title={SCIsegV2: A Universal Tool for Segmentation of Intramedullary Lesions in Spinal Cord Injury}, 
      author={Enamundram Naga Karthik* and Jan Valošek* and Lynn Farner and Dario Pfyffer and Simon Schading-Sassenhausen and Anna Lebret and Gergely David and Andrew C. Smith and Kenneth A. Weber II and Maryam Seif and RHSCIR Network Imaging Group and Patrick Freund and Julien Cohen-Adad},
      year={2024},
      eprint={2407.17265},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.17265}, 
      note = {*Shared first authorship}
}