Score-based diffusion models for accelerated MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Score-based diffusion models are powerful generative tools for image modeling.
• Inverse problems in medical imaging, such as accelerated MRI, require robust reconstruction techniques.
• Existing methods often struggle with limited training data or complex reconstruction tasks.

Purpose of the Study:

• To introduce a novel method for solving inverse problems in imaging, particularly accelerated MRI, using score-based diffusion models.
• To enable reconstruction of complex-valued data from magnitude images only.
• To develop a method that is agnostic to sampling patterns and generalizes to unseen data.

Main Methods:

• Training a continuous time-dependent score function using denoising score matching.
• Employing an iterative inference approach combining a numerical Stochastic Differential Equation (SDE) solver and a data consistency step.
• Utilizing magnitude images for training while reconstructing complex-valued outputs.

Main Results:

• The proposed model successfully reconstructs complex-valued data from magnitude-only training.
• The method demonstrates robustness across various sub-sampling patterns and generalizes to different body parts and imaging sequences.
• Performance surpasses models trained with full supervision, achieving superior quality and practicality.

Conclusions:

• Score-based diffusion models offer a powerful, versatile, and data-efficient approach for accelerated MRI reconstruction.
• The generative nature of the model allows for uncertainty quantification, a significant advantage over traditional regression methods.
• The method's strong performance and generalization capabilities make it highly practical for clinical applications.