A Deep Learning System That Helps Detect Heart Disease from MRI Scans

by Micaela Harris-Kim
March 31, 2026

Cardiac magnetic resonance imaging (CMR) is the reference standard for evaluating cardiac anatomy and function. It captures video of cardiac and valvular motion, quantifies scar tissue, and identifies areas of poor perfusion - all without exposing patients to radiation. With the emergence of deep learning, clinicians may soon be able to leverage AI systems to more precisely interpret these scans and develop better-targeted therapies for patients with cardiovascular disease.

Deep learning has shown remarkable promise in cardiovascular diagnosis, but meaningful limitations remain in clinical deployment. Patients with complex or overlapping heart conditions present layered diagnoses that are difficult for AI models to parse, often requiring the model to be retrained from scratch for each new setting. Investigators in a joint study at Stanford University and the University of Pennsylvania, led by Rohan Shad, MD, Mrudang Mathur, PhD, and William Hiesinger, MD, built a transformer-based system that learns complex patterns of heart disease from a collection of over 19,000 cardiac MRI scans gathered across multiple hospitals.

The model was trained using radiology reports as guidance, learning to connect visual patterns within scans to the language clinicians use when describing what they see. 

A large language model helped bridge this visual-textual alignment, producing clinically relevant descriptors. The result was a system that began grouping patients with similar cardiovascular disease patterns together - without being explicitly instructed to do so.

Critically, the model can rapidly adapt to new clinical environments using only a small amount of additional training data. The investigators also applied the model to screen more than 40,000 MRI scans from the UK Biobank, identifying a subset with elevated probability of cardiovascular disease. Clinician review confirmed that more than half of those flagged patients had hypertrophic cardiomyopathy, demonstrating the model's potential to surface meaningful disease patterns within large datasets.

Looking ahead, this system could support accurate diagnosis of cardiovascular disease, serve as a population-level screening tool, and potentially reduce the need for invasive procedures such as cardiac biopsies. The researchers also aim to extend the approach to other types of clinical data and reporting beyond MRI.

Additional Stanford Cardiovascular Institute investigators include Dhamanpreet Kaur, Robyn Fong, Joseph Cho, Matthew Leipzig, Euan Ashley, and Curtis Langlotz.

Rohan Shad, MD

Mrudang Mathur, PhD

William Hiesinger, MD