Minimum sample size to detect spinal cord atrophy with automatic soft segmentation

Motivation: Spinal cord atrophy, a key biomarker in neurodegenerative diseases like multiple sclerosis, is challenging to measure reliably across varying MRI contrasts and protocols in multi-center studies. Goal(s): To evaluate a novel, contrast-agnostic "soft" segmentation method's sensitivity for atrophy detection by calculating spinal cord cross-sectional area (CSA) at specific vertebral levels. Approach: We applied rescaling and transformations on MRI scans to simulate atrophy and used our segmentation model to calculate CSA, comparing results with previous methods through a scan-rescan experiment. Results: The contrast-agnostic model, significantly reduced sample sizes needed for detecting spinal cord atrophy, especially with soft segmentations. Impact: Reducing the required sample size will allow for early spinal cord atrophy detection, especially in multi-center and multi-contrast studies.