VQ-SToRM: Vector-Quantized Smoothness Regularization on Manifolds for Free-Breathing, Ungated Real-Time Cardiac MRI...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Real-time, free-breathing, ungated cardiac MRI (CMR) is crucial for patients unable to perform breath-holds.
• Conventional methods face challenges with image quality due to aggressive undersampling and lack of fully sampled references.
• Existing unsupervised methods struggle to represent distinct cardiac and respiratory motions simultaneously.

Purpose of the Study:

• To introduce VQ-SToRM, an unsupervised framework for real-time CMR reconstruction.
• To address limitations of continuous latent spaces in representing complex physiological motions.
• To enable accurate motion-resolved imaging in challenging patient populations.

Main Methods:

• Developed VQ-SToRM, adapting Vector-Quantized Variational Autoencoder for real-time CMR.
• Replaced continuous latent manifolds with a learned discrete codebook for joint training on undersampled k-t space data.
• Utilized subject-specific, unsupervised learning on non-Cartesian spiral acquisitions.

Main Results:

• VQ-SToRM accurately resolved cardiac and respiratory motion in free-breathing, ungated CMR.
• A compact codebook (5 embeddings, dimension 10) proved optimal for representing motion content.
• VQ-SToRM demonstrated smoother frame-to-frame transitions and superior image quality metrics compared to existing methods.

Conclusions:

• VQ-SToRM offers a promising unsupervised approach for clinically practical real-time CMR.
• The discrete codebook effectively captures dominant cardiac and respiratory motion.
• This method enhances image quality and stability for non-breath-holding patients.