Image-Derived Phenotype Extraction for Genetic Discovery via Unsupervised Deep Learning in CMR Images

Prospective studies with linked image and genetic data, such as the UK Biobank (UKB), provide an unprecedented opportunity to extract knowledge on basis of image-derived phenotypes. However, extent phenotypes tested within so-called genome-wide association (GWAS) is usually limited handcrafted features, where main limitation proceed otherwise high dimensionality both imaging data. Here, we propose approach phenotyping performed in unsupervised manner, via autoencoders that operate 3D meshes. Therefore, latent variables produced by encoder condense information related geometry biologic structure interest. The network’s training proceeds two steps: first genotype-agnostic second enforces a set markers selected GWAS intermediate representation. This genotype-dependent optimisation procedure allows refinement autoencoder better understand effect encountered. We validated our proposed method left-ventricular meshes derived from cardiovascular magnetic resonance images UKB, leading discovery novel associations that, best knowledge, had not been yet reported literature cardiac