Correlation of whole-body fat content and its distribution with lumbar fat content under different age, gender and BMI conditions: quantitative analysis of MRI based on deep learning

Junjie Fang, Jiazhen Zhu, Aie Liu, Chuanli Cheng, 邹超, Aizhu Sheng, Fang Gan, Bin Chen, Guangyu Tang

The correlation between abnormal distribution of systemic fat deposition and fat content in the vertebral body remains unclear, and AI segmentation offers novel approach for large-scale quantitative analysis. To explore the effects of gender, age, and body mass index on the correlation between whole-body fat content/distribution and lumbar spine fat content using deep learning based MRI quantification technology. One hundred thirteen volunteers (63 males/50 females; 47.7 ± 13.3 years) who underwent coronal whole-body and lumbar sagittal MRI examination were enrolled. Through automatic segmentation and quantification, total adipose tissue (TAT), body-fat percentage (BF%), liver fat fraction (LFF), subcutaneous fat fraction (SAT/TAT), internal fat fraction (IAT/TAT), abdominal subcutaneous/visceral fat ratio (ASVFR), erector spinae–multifidus fat fraction (ESFF), psoas major fat fraction (PMFF) and lumbar vertebral fat fraction (VFF) were obtained. Pearson correlation was then used for stratified analysis based on sex, age (≤ 45, 46–59, ≥ 60 years) and BMI (underweight/normal/over-weight/obese). All preset variables were entered into a Lasso model (10-fold cross-validation) to identify independent VFF determinants and to calculate R² and the root-mean-square error (RMSE). VFF correlated moderately with ESFF and PMFF (r_ ESFF =0.53, r_ PMFF =0.39), weakly with TAT and BF% (r_ TAT =0.23, r_ BF% =0.26), and rarely with LFF ( r = 0.08). Fat distribution exhibits “site-specificity”, with a relatively significant correlation between ASVFR and VFF ( r =-0.40), a weak positive correlation between IAT/TAT and VFF ( r = 0.29), and a slightly negative association between SAT/TAT and VFF ( r =-0.29). Correlations between VFF and relevant indicators exhibited stratified differences by gender, age, and BMI, with significant gender-specific variations, insignificant age-specific variations, and ESFF being the sole indicator prominently and positively correlated with VFF across all BMI strata. The multivariate model (R² = 0.63) identified ESFF, PMFF, and ASVFR as the core drivers of VFF, rather than whole-body fat or liver fat. Paravertebral muscle fat infiltration and visceral-fat predominance are associated with higher lumbar marrow fat accumulation, which tends to be more effective in the female, elderly and overweight groups. The constructed multi-variable model may provide a quantitative imaging framework for future studies investigating vertebral marrow fat and bone health.