Segment-Any-Muscle: Towards an Open-Source, Contrast-Agnostic Computer-Vision Muscle Segmentation Model for MRI and CT

Motivation: Declines in muscle health are common in multiple conditions, and MRI is the gold standard for non-invasively assessing muscle. Computer-vision models show promise for the automated quantification of muscle health with MRI but lack generalizability across sequences and imaging modalities. Goal(s): To develop a contrast-agnostic computer-vision muscle segmentation model for the lumbar paraspinal muscles that generalizes to multiple MRI contrasts as well as CT. Approach: We trained, tested, and validated the model on over 1,500 MRI and CT images. Results: The model showed high accuracy (Sørensen-Dice≥0.865) on an external, multimodal validation dataset and can be accessed at github.com/MuscleMap/MuscleMap . Impact: This contrast-agnostic computer-vision model can automatically and accurately assess muscle health from both MRI and CT. We are expanding this to all muscles to support multiple clinical and research applications linking muscle health to overall health and disease.