Summary:
UCLA researchers from the Department of Radiological Sciences have developed a deep learning-based segmentation framework that enables fully automated and reproducible analysis of left ventricular (LV) function in Displacement Encoding with Stimulated Echoes (DENSE) MRI.
Background:
Myocardial strain is a key metric for analyzing how much the heart muscle stretches and contracts during each beat and —is a powerful early indicator of heart dysfunction. It can detect subtle structural and functional changes before symptoms appear, aiding in the diagnosis and management of conditions like heart failure, chemotherapy-related heart damage, and inherited cardiomyopathies. Advances in cardiac imaging, particularly magnetic resonance imaging (MRI), have enabled precise visualization of heart function and medical intervention to improve patient outcome. A specialized MRI technique called DENSE (Displacement Encoding with Stimulated Echoes) can measure these small but critical muscle movements, offering an accurate and reliable way to assess strain and identify early signs of cardiac disease.
Despite its advantages, DENSE MRI is difficult to use in clinical practice because the heart must be manually traced frame-by-frame to extract strain data. This process is slow, labor-intensive, and subject to variability across users. While AI-based solutions have been proposed to automate segmentation, they typically require large datasets of manually labeled DENSE images—something that is rarely available due to the technique’s complexity and limited adoption. There is a clear need for an accurate, automated segmentation approach that can work with minimal training data to make DENSE MRI more accessible for routine clinical use.
Innovation:
UCLA researchers have developed a novel deep learning framework that enables automated and accurate analysis of DENSE MRI without requiring large, manually labeled datasets. This new method, called MaskNet, integrates artificial intelligence with cutting-edge tools in image recognition to automatically identify the heart muscle and measure myocardial strain.
Unlike traditional AI approaches that rely on thousands of labeled training images, MaskNet leverages existing cine cardiac MRI data (which are more widely available and already annotated) and uses the Segment Anything Model (SAM) to help recognize heart structures in DENSE images. Through a process known as domain adaptation, the system transfers what it learns from one type of image to another, enabling it to accurately trace the heart muscle in DENSE MRI scans, even with limited training data. In testing, this integrated approach outperformed both traditional supervised models and each of its individual components, demonstrating strong agreement with manual segmentation on both a global and regional level. By significantly reducing the time, expertise, and data needed to process DENSE MRI, this innovation opens the door to broader clinical and research use of advanced cardiac strain imaging.
Potential Applications:
• Automated cardiac function analysis in clinical MRI workflows
• Early detection and monitoring of heart failure or cardiotoxicity in cancer patients
• Cardiac imaging tools for pharmaceutical trials and drug safety monitoring
• AI-assisted post-processing software for cardiac MRI vendors
• Research studies on myocardial strain in inherited or rare cardiac diseases
• Cloud-based or integrated hospital systems for efficient cardiac image analysis
Advantages:
• Eliminates the need for extensive image annotation
• Leverages widely available cine datasets
• Integrates state-of-the-art foundational model like Segment Anything Model (SAM)
• High segmentation accuracy and reproducibility
• Enables automated strain analysis in DENSE MRI
• Generalizable to other cardiac MRI segmentation tasks
State of Development:
The current prototype has been tested on in vivo cardiac datasets, including 24 subjects. This work has been presented to 2025 Annual Meeting of the International Society for Magnetic Resonance in Medicine (ISMRM) in May 2025.
Related Publications:
Li S, Finn JP, Ruan D, Nguyen KL, Zhong X. SAM-driven MaskNet for left ventricle segmentation on cine DENSE with unsupervised domain adaptation. Proc Intl Soc Mag Reson Med 33. 2025
Reference:
UCLA CASE No. 2025-208