Automated deep-learning quantification of intramuscular fat in lumbar spine muscles on Dixon MRI: validation and normative reference values from 173 healthy adults

Abstract Background Intramuscular fat (IMF) accumulation in lumbar‑spine muscles is associated with sarcopenia, low‑back pain and poorer surgical outcomes. Dixon MRI allows the measurement of IMF with voxel-wise fat fraction (FF) mapping, but quantitative use in research or clinics is limited by labour-intensive 3D muscle segmentation and by the absence of healthy reference data. We present a lightweight pipeline that requires a small number of labelled images yet delivers accurate segmentation and Dixon-based reference ranges for lumbar spine IMF in asymptomatic adults with diverse physical activity levels. Methods Twenty-six Dixon scans (44.8 ± 14.3 years old; 12 males / 14 females) were manually annotated for psoas, iliacus, quadratus lumborum and the erector spinae + multifidus complex. A 3D U‑Net was trained with a realistic data augmentation scheme that combines small rigid rotations and non-linear B-spline deformations mimicking real anatomical variability, enhancing generalisation. Mean fat fraction (FF) of each muscle was extracted from Dixon FF images. Segmentation (Dice, relative volume difference, Hausdorff distance) and FF estimation accuracy was assessed with a five‑fold Monte‑Carlo cross‑validation and was compared to a multi-atlas approach. FF agreement with manual measurements was examined using linear regression and Bland–Altman analysis. The trained model then processed the complete study cohort of 173 healthy adults (20–70 years old; sedentary to highly active) to derive healthy subjects FF percentiles. Results Automated FF strongly correlated with manual derived values (R² = 0.96), showing no systematic bias. The 3D U-Net achieved Dice coefficients of up to 0.936 (erector spinae + multifidus) and demonstrated superior performance over multi-atlas segmentation across all metrics (p < 0.01). The model was also able to capture different characteristic FF values for each muscle, with median (IQR) normative FF values of 21.8% (17.4–26.9%) for erector spinae + multifidus, 18.0% (14.9–22.2%) for quadratus lumborum, 13.4% (10.5–16.3%) for psoas and 11.6% (9.5–13.8%) for iliacus, with significant sex and activity differences (p < 0.01). Conclusion With fewer than 30 manual annotations, a 3D U-Net model enhanced by anatomically realistic data augmentation, provides a reliable and accurate solution for automated IMF quantification from Dixon MRI. We deliver novel Dixon MRI reference values of lumbar spine muscle composition in healthy adults. Both the automated estimation of FF, and the reference FF values reported here, facilitates the study of sarcopenia, spinal pathologies, and muscle degeneration.