Segmentation model quantifying bone marrow fat on whole-body Dixon MRI reveals association between vertebral fat and diabetes

Motivation: Robust assessment of the fat distribution in normal and abnormal bone marrow on whole-body Dixon MRI images is difficult. Goal(s): Our goal was to develop a deep learning model to segment and quantify whole-body bone marrow and evaluate its value in a community-based study. Approach: A three-dimensional nnU-Net model was trained with diverse bone marrow changes, tested on internal and external test sets, and applied to images from community-based populations. Results: The model precisely segmented bone marrow and determined fat fractions, thus confirming the correlation between vertebral body BMFF and diabetes. Impact: Our novel three-dimensional nnU-Net model for automated assessment of whole-body bone marrow fat sheds new light on the link between bone marrow adiposity and diabetes.