Deep cardiac phenotyping with applications in imaging genetics

The emergence of large prospective biobanks with paired imaging and genetic data, such as the UK Biobank (UKBB), has enabled the investigation of image-derived phenotypes in genetic association studies, aiming to identify genetic variants that drive phenotypic differences observable in medical images. Traditionally, these studies have focused on handcrafted phenotypes—traits known to be clinically relevant. However, the unprecedented sample sizes now available facilitate data-driven phenotyping, offering the potential to uncover more subtle phenotypic variations that can enhance genetic discovery.

In this work, we focus on cardiovascular cine magnetic resonance (CMR) imaging and apply state-of-the-art image processing techniques to generate 3D meshes of the myocardium from over 50,000 UKBB participants. We then leverage geometric deep learning to learn unsupervised representations of these shapes while explicitly preserving their topology.

First, we phenotype static meshes at end-diastole (ED) in an unsupervised manner using convolutional mesh autoencoders (CoMA), a well-established shape analysis technique. We then introduce a novel methodology to capture dynamic patterns across the full cardiac cycle, inherently incorporating the periodicity of cardiac motion. These learned representations are subsequently analyzed in genome-wide association studies (GWAS) to identify genetic variants linked to both static and dynamic phenotypes. Additionally, we demonstrate the effectiveness of an ensemble-based phenotyping framework in improving discovery power. Finally, we perform an in-depth genetic analysis to interpret our findings in the context of existing biological evidence and identify potential candidate genes underlying these associations.