Brain-spinal cord MS lesion segmentation on MP2RAGE images

[Nilser3]

Description

Given that we have multicentre MS 3T MP2RAGE data that includes the brain and spinal cord, and also that some models have been developed to segment MS lesions in the brain (robust even with UNIT1 images), this issue is to generate GTs of MS lesions in the brain from these models and then train a model to segment MS lesions in the brain-spinal cord MP2RAGE images.

Data

• basel-mp2rage
• marseille-3T-mp2rage
• nih-ms-mp2rage

Testing brain MS lesion segmentation models

1. samseg model

• Version dev available on freesurfer-linux-ubuntu22_x86_64-dev-20240112-88246c0
• Command-line: run_samseg -i image_UNIT1.nii.gz --lesion --save-probabilities -o brain_samseg/

2. WMH-SynthSeg model

• Version dev available on freesurfer-linux-ubuntu22_x86_64-dev-20240112-88246c0 using inference.py
• Command-line: python inference.py --i UNIT1_images/ --save_lesion_probabilities --o brain_ms_lesions/

Results

[Nilser3]

Lesions in the brainstem and cerebellum

MS lesion in the pons

MS lesion in the cerebellar peduncle

MS lesions in the cerebellum

MS lesions in the brainstem and cerebellum are not well segmented by either method,
which means that in order to train a MS brain-spinal cord model it will be necessary to manually segment/correct these lesions.