Congenital Heart Anomaly Representation and Machine-learning for Screening

CHARMS aims to transform prenatal screening by developing novel normative representation learning (NRL) algorithms for ultrasound imaging and introducing them into clinical workflows, to enhance early detection of congenital heart anomalies. Addressing disparities in maternity outcomes and global shortages in sonography expertise, the project targets improving the detection of congenital heart disease (CHD), a common and often fatal congenital malformation with low antenatal detection rates. Building on the ERC MIA-NORMAL project, we will create robust out-of-distribution detectors specifically for ultrasound imaging and CHD, utilising comprehensive data, including videos, images, and patient characteristics. CHARMS tools will identify subtle abnormalities in motion patterns or anatomical changes, significantly enhancing CHD detection in practice.We will develop normality-focused models to confidently confirm normal development and malformations using ultrasound video data, design self-supervised anomaly detectors to identify 'unknown unknowns’ and equip models with generative capabilities to ensure trustworthiness and robustness against covariate shifts such as poor image quality or varying ultrasound scanners. Through generative techniques we will also validate and refine 3D+time representations of fetal cardiac anatomy, facilitating a comparative analysis with MRI data, further enhancing model precision.Additionally, CHARMS will explore correlations between non-imaging and imaging phenotypes, utilising knowledge distillation to align features with physiological textbook knowledge. This will generate synthetic scan variants that match clinical phenotypes, enabling a deeper understanding of disease mechanisms and the development of new image-based biomarkers.A Quality Assurance framework will be established to ensure model safety and robustness, complying with the EU AI Act, including protocols for rigorous validation against known physiological facts.