Lesion segmentation in multiple sclerosis (MS)

SCT provides several deep learning-based algorithms to segment lesions in multiple sclerosis (MS). Depending on the image contrast, you can use the following algorithms:

T2w and T2star

The algorithm sct_deepseg_lesion was trained on T2w and T2star images. Details: Gros, C., et al. NeuroImage (2019).

https://raw.githubusercontent.com/spinalcordtoolbox/doc-figures/master/spinalcord-segmentation/sct_deepseg_sc_steps.png

Run the following command to segment the lesion using sct_deepseg_lesion from the input image:

sct_deepseg_lesion -i t2.nii.gz -c t2
Input arguments:

  • -i : Input T2w image with fake lesion

  • -c : Contrast of the input image

Output files/folders:

  • t2_lesionseg.nii.gz : 3D binary mask of the segmented lesion

  • t2_res_RPI_seg.nii.gz : intermediate segmentation file – you can ignore this file

  • t2_RPI_seg.nii.gz : intermediate segmentation file – you can ignore this file

STIR and PSIR

The algorithm seg_sc_ms_lesion_stir_psir was trained on sagittal STIR and PSIR images. It is a region-based model, outputting a single segmentation image containing 2 classes representing the spinal cord and MS lesions. Details: https://github.com/ivadomed/canproco.

https://raw.githubusercontent.com/spinalcordtoolbox/doc-figures/master/lesion-analysis/seg_sc_ms_lesion_stir_psir.gif

You can try seg_sc_ms_lesion_stir_psir on your own STIR or PSIR image using the following command:

sct_deepseg -i psir.nii.gz -task seg_sc_ms_lesion_stir_psir -qc ./qc
Input arguments:

  • -i : Input PSIR (or STIR) image

  • -task : Task to perform. In this case, we use the seg_sc_ms_lesion_stir_psir model

  • -qc : Directory for Quality Control reporting. QC reports allow us to evaluate the segmentation slice-by-slice

MP2RAGE-UNIT1

The algorithm seg_ms_lesion_mp2rage was trained on cropped MP2RAGE-UNIT1 images. Details: Cohen-Adad, J., et al. Zenodo release (2023).

https://raw.githubusercontent.com/spinalcordtoolbox/doc-figures/master/lesion-analysis/model_seg_ms_mp2rage.png

You can try seg_ms_lesion_mp2rage on your own MP2RAGE UNIT1 image using the following commands. As the model was trained on cropped images, we recommend cropping the input image before running the segmentation.

sct_deepseg -i IMAGE_UNIT1 -task seg_sc_contrast_agnostic -o IMAGE_seg
sct_crop_image -i IMAGE_UNIT1 -m IMAGE_seg -dilate 30x30x5
sct_deepseg -i IMAGE_UNIT1 -task seg_ms_lesion_mp2rage -qc ./qc
Input arguments:

  • -i : Input MP2RAGE UNIT1 image

  • -task : Task to perform. In this case, we use the seg_ms_lesion_mp2rage model

  • -qc : Directory for Quality Control reporting. QC reports allow us to evaluate the segmentation slice-by-slice