Objectives:
Develop uncertainty-aware vision transformer models for automated VAT/SAT/muscle segmentation from UK Biobank abdominal and/or full body MRI data, establishing precision body composition biomarkers that outperform traditional anthropometrics for cardiovascular and metabolic risk prediction.
Primary Outcomes:
Incident myocardial infarction, stroke, type 2 diabetes, cardiovascular mortality, and all-cause mortality over 10+ years follow-up.
Technical Innovation:
Novel two-stage approach using masked auto-encoding on Swin-UNETR models for foundation weights, followed by supervised segmentation with confidence-aware losses. This creates robust uncertainty-quantified body composition phenotyping – critical for clinical deployment where model confidence guides clinical decision- making.
Risk Factor Analysis & Research:
Comprehensive risk prediction analysis using three distinct approaches: (1) traditional risk factors only (HbA1c, lipids, CRP, blood pressure), (2) MRI-derived body composition metrics only, and (3) ensemble methods combining both modalities. Novel MRI-based risk scoring schema will be developed using gradient boosted trees trained on outcome data, establishing first-of-its-kind imaging-derived cardiometabolic and mortality risk stratification.
Clinical Impact:
Body composition-derived risk scores will enable precision cardiovascular risk stratification beyond BMI, particularly for metabolically-normal obesity and lean diabetes phenotypes where traditional measures fail. Uncertainty-aware predictions allow graduated clinical responses based on model confidence.
Population Health Applications:
Large-scale phenotyping revealing hidden body composition patterns across demographics, enabling precision public health interventions and redefining metabolic risk assessment paradigms.
