AI maps body fat and muscle to predict risk

The study, published May 5 in Radiology and highlighted by the Radiological Society of North America (RSNA), applied a fully automated, open-source deep-learning pipeline to whole-body MRI scans from 66,608 individuals recruited from the UK Biobank and the German National Cohort between April 2014 and May 2022, the authors report (AuntMinnie; News-Medical; RSNA). The cohort had a mean age of 57.7 years, included 34,443 males, and had mean BMI 26.2 (AuntMinnie; News-Medical). The team computed age-, sex-, and height-normalized z-scores for body composition metrics including subcutaneous adipose tissue, visceral adipose tissue, skeletal muscle, skeletal muscle fat fraction, and intramuscular adipose tissue (AuntMinnie; News-Medical).

Per the RSNA release and reporting in AuntMinnie and News-Medical, the body-composition measures were expressed as z-scores (low: z < -1; middle: z = -1 to 1; high: z > 1) and the prognostic value of those z-scores was tested for incident diabetes, major adverse cardiovascular events, and all-cause mortality (AuntMinnie; RSNA). Key reported effect sizes include a 2.26-fold higher risk of future diabetes for individuals with high visceral fat, a 1.54-fold increased risk of future major cardiovascular events associated with high intramuscular fat, and a 1.44-fold higher all-cause mortality for low skeletal muscle (AuntMinnie). The authors made their segmentation framework open-source and describe it as fully automated in the RSNA and News-Medical coverage.

Editorial analysis: Population-scale imaging studies that combine whole-body MRI with automated segmentation increasingly produce normative reference maps that can be converted to individual-level z-scores. Industry observers note that converting pixel-level segmentations into standardized z-scores is a practical route to integrate imaging-derived biomarkers into epidemiology and risk modelling pipelines.

Editorial analysis: For clinicians and ML practitioners focused on medical imaging, the study demonstrates that muscle quantity and quality metrics carry prognostic information distinct from traditional anthropometrics such as BMI. The result underscores a recurring pattern in imaging research where distribution and composition measures outperform crude surrogate metrics for some long-term health outcomes.

Editorial analysis: Observers should look for independent validation in non-European cohorts, availability and documentation of the open-source code and models, and prospective studies that evaluate whether integrating these MRI-derived z-scores into clinical workflows improves decision making or outcomes. Additionally, practitioners deploying similar pipelines will monitor compute cost, segmentation robustness across scanner vendors, and regulatory pathways for imaging biomarkers.

The paper includes direct commentary on BMI limitations. Jakob Weiss, M.D., Ph.D., is quoted saying, "Many risk scores and treatment decisions still rely on BMI or waist circumference because they are simple to obtain. But BMI does not reliably reflect a person's actual body composition," as reported in the RSNA/press materials (AuntMinnie; News-Medical; RSNA). The authors note that reference standards for asymptomatic individuals across age and sex are scarce, and the study supplies a large-scale reference map intended to address that gap (News-Medical; RSNA).