Deep Learning to Estimate Biological Age on Whole-body MRI and Investigate its Association with Frailty and Long-term Mortality

• Applicant: Dr. Matthias Jung
• Subject Area: Radiology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 518480401

Objective: The proposed study aims to assess whether deep learning estimates of biological age based on 3D segmentations of body composition (BC) derived from whole-body MRI (BC-age model) can predict frailty and all-cause mortality beyond chronological age. Background: Europe’s population is getting older. The concept of healthy aging defined by the World Health Organization (WHO) is becoming increasingly important. In contrast to healthy aging, frailty is a state of increased vulnerability to stressors that increases a person’s risk for adverse health outcomes, such as falls, hospitalization rate, and death. Frailty is strongly linked to an individual’s body composition and is thought to reflect biological age rather than chronological age. A non-invasive and opportunistic biological age measure may improve personalized medicine and management of (pre-)frailty. Whether deep learning can estimate biological age based on 3D segmentation masks of muscle and adipose tissue compartments derived from whole-body MRI remains unclear. Methods: The BC-age model will be developed on 30.000 whole-body MRIs of the German National Cohort (GNC). The deep learning model will take 3D-BC segmentation masks as an input and will output an estimate of chronological age (assumed to be equal to biological age in the asymptomatic GNC). This BC-age model will be independently tested on the UK Biobank (n= 55.345) to predict frailty (primary outcome) and all-cause mortality (secondary outcome). Preliminary work: In preparation for a successful completion of the proposed project, we used n=200 whole-body MRIs of the GNC to develop a patch-based stack of convolutional neural networks (CNNs) for automated 3D-BC quantification defined as subcutaneous fat (SAT), visceral fat (VAT), intramuscular fat (IMAT), and musculature. The 3D-CNN’s performance was validated on independent internal and external datasets not seen at any point during model training. Segmentations performed by the 3D-CNN resulted in averaged Dice Scores of ≥ 0.96 on the independent internal validation set (GNC, n=50) and ≥ 0.95 on the external validation set (UK Biobank, n=50) for SAT, VAT, IMAT, and musculature. Relevance: Body composition-based biological age estimates may serve as a new measure to identify individuals at risk in geroscience and as an opportunistic imaging biomarker on cross-sectional imaging studies performed in clinical practice to identify (pre-)frailty.

• DFG Programme: WBP Fellowship
• International Connection: USA
• Host: Professor Dr. Michael Lu