Score-Based Diffusion Models With Self-Supervised Learning for Accelerated 3D Multi-Contrast Cardiac MR Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Long scan times limit the clinical utility of 3D multi-contrast cardiac MRI (3D-MC-CMR).
• Accelerated acquisition techniques are crucial for widespread adoption of advanced cardiac imaging.

Purpose of the Study:

• To develop and validate a novel method for accelerating 3D-MC-CMR acquisition using score-based diffusion models and self-supervised learning.
• To enable accurate and efficient reconstruction of 3D-MC-CMR images without requiring fully sampled training data.

Main Methods:

• Established a mapping between undersampled k-space data and MR images using a self-supervised Bayesian reconstruction network.
• Developed a joint score-based diffusion model to learn the distribution of 3D-MC-CMR images.
• Reconstructed 3D-MC-CMR images via conditioned Langevin Markov Chain Monte Carlo sampling.

Main Results:

• The proposed method achieved accurate reconstruction of 3D-MC-CMR images, even at a high acceleration rate of 14x.
• Reconstructed images yielded high-quality T1 and T1ρ parametric maps comparable to reference maps.
• Outperformed traditional compressed sensing and existing self-supervised deep learning MRI reconstruction techniques.

Conclusions:

• Score-based diffusion models with self-supervised learning offer a powerful approach to accelerate 3D-MC-CMR acquisition.
• This method significantly improves scan efficiency while maintaining diagnostic image quality.
• Enables more accessible and widespread application of advanced cardiac MRI techniques.