Medical images provide a rich source of quantitative phenotypes that can be analysed in large population cohorts. A major focus of our current research is to develop and apply computational approaches for extracting such phenotypes and combining them with genomic and clinical data. We use both interpretable, biologically defined image features and representations learned by deep-learning models, with a particular focus on retinal and vascular imaging.<\/p>\n\n\n\n
Much of this research has been supported by the SNSF Sinergia project VascX<\/strong><\/a>, which brings together computational biology, ophthalmology, angiology and population research to investigate vascular phenotypes and their relationships with genetics and disease.<\/p>\n\n\n\n In our recent work, accepted for publication in Nature Genetics<\/strong>, we systematically compared traditional, interpretable features extracted from retinal fundus images with representations learned by the retinal foundation model RETFound. We found that the two approaches capture partly complementary information: combining them improves prediction of several diseases and medically relevant traits, while genetic analyses reveal both shared and distinct biological signals.<\/p>\n\n\n\n The results illustrate that deep-learning representations need not replace explicitly defined image phenotypes. Combining the two can provide both improved predictive performance and greater insight into the biological information contained in medical images.<\/p>\n\n\n\n Preprint:<\/strong> Comparing tangible retinal image characteristics with deep learning features reveals their complementarity for gene association and disease prediction<\/em><\/a> (medRxiv Jan. 2026)<\/p>\n\n\n\n More:<\/strong> UNIL article Combining AI and traditional eye image features improves disease and gene detection<\/a><\/em><\/p>\n\n\n\n In work accepted for publication in Nature Communications<\/strong>, we fine-tuned RETFound to predict chronological age from retinal fundus images of more than 70,000 UK Biobank participants. The resulting retinal age gap<\/em> \u2014 the difference between predicted and chronological age \u2014 is associated with cardiometabolic and inflammatory traits, cognitive performance, mortality and several major diseases.<\/p>\n\n\n\n Genetic analyses connect retinal ageing to genes involved in longevity, metabolism, neurodegeneration and age-related eye disease. Analyses performed separately in females and males further reveal differences in the biological signatures underlying retinal ageing, including a stronger contribution of retinal vasculature in females.<\/p>\n\n\n\n Preprint:<\/strong> Deep learning aging marker from retinal images unveils sex-specific clinical and genetic signatures<\/a> <\/em>(medRxiv, Oct. 2025)<\/p>\n\n\n\n In a Nature Communications<\/strong> study published in 2024, we extracted 17 interpretable vascular phenotypes \u2014 including vessel diameter, tortuosity, density, branching patterns and vascular angles \u2014 from more than 130,000 retinal images of nearly 72,000 UK Biobank participants.<\/p>\n\n\n\n Genome-wide analyses showed that these traits are heritable but have largely distinct genetic architectures. Their associations with disease and cardiovascular risk factors, together with Mendelian-randomisation analyses and replication in independent cohorts, demonstrate how large-scale image-derived phenotyping can provide biological and potentially causal insights into vascular health.<\/p>\n\n\n\n Article:<\/strong> Phenotypic and genetic characteristics of retinal vascular parameters and their association with diseases<\/em><\/a> (Nature Communications, 2024)<\/p>\n\n\n\n Related recent work:<\/u><\/strong><\/p>\n\n\n\n Extending our image-based phenotyping beyond the retina, we compared vascular properties derived from imaging of the retina, carotid arteries, aorta and brain in the UK Biobank. We identified phenotypic and genetic relationships between vascular characteristics in different organs, supporting the idea that retinal vascular features contain information about systemic vascular health.<\/p>\n\n\n\n Article:<\/strong> Cross-organ analysis reveals associations between vascular properties of the retina, the carotid and aortic arteries, and the brain<\/em><\/a> <\/em>(Communications Medicine, 2026)<\/p>\n\n\n\n Sex-stratified analyses of retinal vascular phenotypes revealed differences in their relationships with disease and in their genetic architecture. Among other findings, females showed higher heritability and more associated loci for several retinal traits, and we identified the first X-chromosomal genetic association for retinal vascular phenotypes.<\/p>\n\n\n\n Accepted in Communications Biology<\/strong> We developed RetFit<\/strong>, a deep-learning biomarker that estimates cardiorespiratory fitness from retinal fundus images. RetFit is associated with cardiovascular outcomes, mortality and a broad range of disease traits, while genetic and interpretability analyses suggest that it captures information distinct from conventional fitness estimates and is strongly influenced by retinal vascular features.<\/p>\n\n\n\n Preprint:<\/strong> RetFit: A Novel Deep Learning Biomarker based on Cardiorespiratory Fitness derived from the Retina<\/em><\/a> Using deep learning to segment carotid ultrasound images, we developed automated measurements of carotid lumen diameter in nearly 19,000 UK Biobank participants. Genetic analyses revealed substantial heritability and a strong genetic correlation with carotid intima-media thickness, despite a much weaker phenotypic correlation.<\/p>\n\n\n\n Preprint:<\/strong> Genetic Insights from Automated Lumen Diameter Measurements in Carotid Ultrasounds of the UK Biobank<\/em><\/a><\/p>\n\n\n\n Our most recent work examines the interpretability of vision transformers<\/strong> in retinal imaging. By quantitatively benchmarking different attention-based attribution methods for RETFound, we found that the choice of interpretability method substantially affects the resulting explanations and that the best-performing approach identifies anatomically appropriate vascular structures without having been explicitly trained to do so.<\/p>\n\n\n\nCombining interpretable image features and deep learning<\/h2>\n\n\n\n
Retinal age as a marker of biological ageing<\/h2>\n\n\n\n
Genetics of retinal vascular phenotypes<\/h2>\n\n\n\n
Vascular phenotypes across organs<\/h2>\n\n\n\n
Sex-specific retinal vascular biology<\/h2>\n\n\n\n
Preprint:<\/strong> Sex-specific disease association and genetic architecture of retinal vascular traits<\/em><\/a><\/p>\n\n\n\nRetFit: Estimating cardiorespiratory fitness from the retina<\/h2>\n\n\n\n
In review<\/strong><\/p>\n\n\n\nCarotid vascular phenotyping<\/h2>\n\n\n\n
Understanding what retinal foundation models learn<\/h3>\n\n\n\n