Generalizable spinal cord multiple sclerosis lesion segmentation across MRI contrasts, protocols, and centers

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate characterization of spinal cord lesions in multiple sclerosis (MS) is vital for clinical management.
• Existing automated MRI segmentation tools lack generalizability due to variations in imaging protocols and contrasts.
• This limits their reliable application in diverse real-world clinical settings.

Purpose of the Study:

• To develop a robust, multi-site, and multi-contrast segmentation framework for spinal cord MS lesions.
• To improve the generalizability and clinical translation of automated lesion detection and segmentation.

Main Methods:

• A segmentation model was trained on a large dataset of 4428 annotated spinal cord MRI images from 1849 individuals with MS.
• The dataset included six MRI contrasts (T1w, T2w, T2*w, PSIR, STIR, UNIT1) from 23 imaging centers across 1.5T, 3T, and 7T scanners.
• Model performance was evaluated using neuroradiologist assessments and quantitative metrics on external datasets.

Main Results:

• The proposed model demonstrated superior generalization capabilities compared to existing contrast-specific segmentation pipelines (p < 0.01).
• Robustness was confirmed across various spinal levels, acquisition resolutions, and binarization thresholds.
• Quantitative evaluations on external datasets further validated the model's performance.

Conclusions:

• The developed framework enables accurate and reliable segmentation of spinal cord MS lesions, even with heterogeneous MRI data.
• This addresses a significant challenge hindering the clinical translation of automated segmentation tools.
• The model is integrated into the Spinal Cord Toolbox (v7.2+) and its code is publicly available.