A generalizable deep learning system for cardiac MRI.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Deep Learning for Medical Diagnosis

Background:

• Cardiac Magnetic Resonance Imaging (CMR) is crucial for assessing myocardial structure, function, and tissue characteristics.
• A comprehensive system is needed to represent diverse cardiovascular diseases and health states.
• Existing methods may require extensive labeled data for training.

Purpose of the Study:

• To develop a foundational deep-learning vision system for cardiac MRI.
• To learn visual concepts directly from the text of radiology reports using self-supervised contrastive learning.
• To achieve high diagnostic accuracy across a wide range of cardiovascular conditions.

Main Methods:

• A deep-learning model trained using self-supervised contrastive learning on cine-sequence cardiac MRI scans.
• Learning visual representations from the raw text of accompanying radiology reports.
• Model training and evaluation on multi-institutional US data, UK BioBank, and two external datasets.

Main Results:

• The system demonstrated remarkable performance on various tasks, including left-ventricular ejection fraction regression.
• Accurate diagnosis of 39 different cardiovascular conditions, including cardiac amyloidosis and hypertrophic cardiomyopathy.
• Achieved clinical-grade diagnostic accuracy with significantly reduced training data requirements.

Conclusions:

• The developed deep-learning system effectively contextualizes complex cardiovascular disease from cardiac MRI.
• The system shows emergent capabilities and can be directed towards specific clinical problems.
• This approach offers a promising pathway for efficient and accurate cardiac MRI analysis.