Introduction

Cardiovascular diseases (CVDs) remain the leading global cause of death, which underscores the need for improved risk prediction and early detection. Among commonly used imaging modalities, cardiac MRI (CMR) provides a non-invasive way to quantify morphological and functional cardiac parameters—including chamber dimensions, aortic geometry, valvular measurements, myocardial mechanics, and flow dynamics—with many phenotypes clinically validated and supported by reference ranges.

Previous work, in particular studies using UK Biobank imaging, has extracted widely used cardiac traits such as chamber volumes, ejection fractions, and strains from short-axis, long-axis, and aortic cine images [1–3]. Those pipelines often relied on summary measures at specific time points (for example end-diastole and end-systole) and paid less attention to dynamic temporal behaviour and richer geometric descriptors. In addition, the full scope of CMR—including T1 mapping and flow MRI—for micro-structural and haemodynamic features that are difficult to obtain with other modalities has often been underused, which limits insight into CVD risk and disease status.

More recent studies have used further CMR modalities available in UK Biobank [4–6], but they frequently treat a single modality in isolation, so complementary information across modalities is not fully exploited. Foundation models [7–9] represent a major shift: large models may be adapted or fine-tuned for CVD tasks. Many current approaches nevertheless depend on large language models (e.g. Qwen, LLaMA) or self-supervised encoders that yield high-dimensional embeddings. Those strategies raise issues such as hallucination, limited interpretability, and weak alignment with well-defined, biologically meaningful traits—complicating translation to imaging–genetics, risk stratification, or integration into electronic health records.

By building on established methods to extract a broad set of clinically informed phenotypes—grounded in peer-reviewed literature and extending the range of cardiac features across multiple CMR modalities—this pipeline targets gaps in imaging–genetics research. Most of these phenotypes have established reference ranges, which supports large-scale analyses including risk stratification, disease subtyping, and multi-organ integration.

This documentation has four aims: (i) to summarise the multimodal CMR pipeline for phenotype extraction; (ii) to give the cardiological background needed to interpret the features; (iii) to unify style and terminology across diverse clinical measures; and (iv) to motivate downstream analyses such as genome-wide association studies (GWAS) and phenome-wide association studies (PheWAS). Together, these aims support reproducible research and broader translational use of large-scale CMR imaging.

References

  1. Bai, W., et al. (2018). Automated cardiovascular magnetic resonance image analysis with fully convolutional networks. Journal of Cardiovascular Magnetic Resonance, 20(1), 65.

  2. Bai, W., et al. (2018). Recurrent neural networks for aortic image sequence segmentation with sparse annotations. In MICCAI 2018 (pp. 586–594). Springer.

  3. Bai, W., et al. (2020). A population-based phenome-wide association study of cardiac and aortic structure and function. Nature Medicine, 26(10), 1654–1662.

  4. Ferdian, E., et al. (2020). Fully automated myocardial strain estimation from cardiovascular MRI–tagged images using a deep learning framework in the UK Biobank. Radiology: Cardiothoracic Imaging, 2(1), e190032.

  5. Puyol-Antón, E., et al. (2020). Automated quantification of myocardial tissue characteristics from native T1 mapping using neural networks with uncertainty-based quality-control. Journal of Cardiovascular Magnetic Resonance, 22(1), 60.

  6. Beeche, C., et al. (2025). Early vascular aging determined by 3-dimensional aortic geometry: genetic determinants and clinical consequences. Circulation, 152(11), 748–761.

  7. Christensen, M., et al. (2024). Vision–language foundation model for echocardiogram interpretation. Nature Medicine, 30(5), 1481–1488.

  8. Liu, R., et al. (2024). Teach multimodal LLMs to comprehend electrocardiographic images. arXiv:2410.19008.

  9. Zhang, Y., et al. (2025). Towards cardiac MRI foundation models: comprehensive visual–tabular representations for whole-heart assessment and beyond. Medical Image Analysis, 106, 103756.