A novel deep learning method for large-scale analysis of bone marrow adiposity using UK Biobank Dixon MRI data

Abstract

OBJECTIVES: Bone marrow adipose tissue (BMAT) represents >10% of total fat mass in healthy humans and further increases in diverse clinical conditions, but the impact of BMAT on human health and disease remains poorly understood. Magnetic resonance imaging (MRI) allows non-invasive measurement of the bone marrow fat fraction (BMFF), and human MRI studies have begun identifying associations between BMFF and skeletal or metabolic diseases. However, such studies have so far been limited to smaller cohorts: analysis of BMFF on a larger, population scale therefore has huge potential to reveal fundamental new knowledge of BMAT's formation and pathophysiological functions. The UK Biobank (UKBB) is undertaking whole-body MRI of 100,000 participants, providing the ideal opportunity for such advances. MATERIALS AND METHODS: Herein, we developed a deep learning pipeline for high-throughput BMFF analysis of these UKBB MRI data. Automatic bone marrow segmentation was achieved by designing new lightweight attention-based 3D U-Net convolutional neural networks that allowed more-accurate segmentation of small structures from large volumetric data. Using manual segmentations from 61-64 subjects, the models were trained against four bone marrow regions of interest: the spine, femoral head, total hip and femoral diaphysis. Models were validated using a further 10-12 datasets for each region and then used to segment datasets from a further 729 UKBB participants. BMFF was then determined and assessed for expected and new pathophysiological characteristics. RESULTS: Dice scores confirmed the accuracy of the models, which matched or exceeded that for conventional U-Net models. The BMFF measurements from the 729-subject cohort confirmed previously reported relationships between BMFF and age, sex and bone mineral density, while also identifying new site- and sex-specific BMFF characteristics. CONCLUSIONS: We have established a new deep learning method for accurate segmentation of small structures from large volumetric data. This method works well for accurate, large-scale BMFF analysis from UKBB MRI data and has the potential to reveal novel clinical insights. The application of our method across the full UKBB imaging cohort will therefore allow identification of the genetic and pathophysiological factors associated with altered BMAT. Together, our findings establish the utility of deep learning for population-level BMFF analysis and promise to help elucidate the full impact of BMAT on human health and disease.