Abstract

Intramuscular fat (IMF) accumulation in lumbar spine muscles is linked to sarcopenia, low-back pain, and poorer surgical outcomes. This study presents an automated Dixon MRI pipeline for IMF quantification that requires a relatively small number of manually labeled scans while producing accurate segmentation and reference values in asymptomatic adults with diverse physical activity levels.

Using 26 manually annotated Dixon scans for training and validation, a 3D U-Net with anatomically realistic augmentation was evaluated against a multi-atlas approach and then applied to 173 healthy adults (20-70 years). Automated fat-fraction estimates strongly correlated with manual measurements (R2 = 0.96), with no systematic bias, and segmentation performance outperformed multi-atlas methods across all assessed metrics. The work also reports normative lumbar-muscle fat-fraction ranges, supporting future studies on sarcopenia, spinal pathology, and muscle degeneration.