MODEL-BASED FREE-BREATHING CARDIAC MRI RECONSTRUCTION USING DEEP LEARNED & STORM PRIORS: MODL-STORM

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Cardiac MRI is crucial for diagnosing heart conditions but often requires long scan times.
• Undersampling k-space data can accelerate MRI acquisition but leads to image artifacts and reduced quality.
• Current deep learning methods struggle to integrate diverse prior information effectively.

Purpose of the Study:

• To develop a novel model-based reconstruction framework for accelerated free-breathing and ungated (FBU) cardiac MRI.
• To integrate deep learned (DL) priors and smoothness regularization on manifolds (STORM) priors for enhanced image recovery.
• To overcome limitations of current deep learning algorithms in incorporating prior information.

Main Methods:

• A model-based reconstruction framework was developed, integrating DL and STORM priors.
• DL priors were utilized to capture local correlations within the data.
• STORM priors were employed to leverage subject-dependent non-local similarities.
• A novel formulation allowed seamless integration of DL with prior information.

Main Results:

• The framework successfully reconstructed FBU cardiac MRI from highly undersampled measurements.
• Experimental results demonstrated the potential of the DL and STORM priors.
• The proposed method showed preliminary success in accelerating FBU cardiac MRI acquisition.

Conclusions:

• The developed model-based framework effectively integrates DL and STORM priors for cardiac MRI reconstruction.
• This approach offers a promising direction for accelerating FBU cardiac MRI.
• The method has the potential to improve patient comfort and diagnostic efficiency.