Gray matter segmentation algorithm: sct_deepseg graymatterΒΆ

For segmenting the gray matter, SCT features the graymatter task of sct_deepseg. This model uses a 2D nnU-Net architecture and outputs a binary segmentation.

  • Algorithm: 2D nnU-Net (nnUNetV2)

  • Pros: Agnostic to MRI contrast and spinal cord region; trained on diverse pathologies

  • Cons: 2D model (processes axial slices independently)

The model was trained on datasets from >20 sites, covering:

  • 3 magnetic field strengths: 1.5T, 3T, 7T

  • 9 MRI sequences: T2*w, MTR, T1w (axial), PSIR, rAMIRA, PDw, MP2RAGE (UNI/T1map), QSM, SWI

  • Spinal cord regions: Cervical, thoracic, and lumbar

  • 1367 subjects from healthy controls, pediatrics, multiple sclerosis, spinal muscular atrophy, cervical degenerative myelopathy, spinal cord injury, amyotrophic lateral sclerosis, post-polio syndrome, and stroke

More details about the model and its training data are available at github.com/ivadomed/model-gm-contrast-region-agnostic.