A generalizable deep learning system for cardiac MRI Cardiac MRI provides a detailed assessment of myocardial structure, function, and tissue properties. This study introduces a foundational vision system for cardiac MRI capable of representing the full spectrum of human cardiovascular disease and health. The deep-learning model is trained using self-supervised contrastive learning, where visual concepts from cine-sequence cardiac MRI scans are derived from the raw text of accompanying radiology reports. The model is trained and evaluated on data from four major U.S. academic clinical institutions and further tested on the UK BioBank and two additional public datasets. The system demonstrates strong performance across diverse tasks, including left-ventricular ejection fraction regression and the diagnosis of 39 conditions such as cardiac amyloidosis and hypertrophic cardiomyopathy. It achieves clinical-grade diagnostic accuracy with significantly less training data than traditional methods. Traditional deep-learning approaches for cardiovascular disease diagnosis rely on supervised training with curated datasets of predefined conditions. However, these systems often struggle with real-world clinical data, which is heterogeneous and includes multiple concurrent abnormalities. For example, patients with inherited cardiomyopathies may also have severe valvular disease, while those with ventricular thrombus may exhibit signs of heart failure from past ischemic events. Supervised models trained for one task rarely generalize to others, requiring retraining from scratch for each new clinical problem. This process demands thousands of labeled examples and lacks the contextual understanding that human clinicians inherently possess.#stanford #ucsf #medstar #acdc_dataset #uk_biobank