Stable Deep MRI Reconstruction Using Generative Priors

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Neuroscience

Background:

• Data-driven methods show promise in magnetic resonance imaging (MRI) reconstruction.
• Clinical integration is hindered by poor generalizability and interpretability of current deep learning models.
• Existing approaches often struggle with varying data distributions and lack uncertainty estimation.

Purpose of the Study:

• To develop a unified framework for generalizable and interpretable MRI reconstruction.
• To address the limitations of current data-driven approaches in clinical settings.
• To enable uncertainty quantification in MRI reconstruction.

Main Methods:

• Proposed a novel deep neural network regularizer trained generatively on reference magnitude images.
• Integrated the trained regularizer into a classical variational framework for reconstruction.
• Developed a fast algorithm for joint image and sensitivity map estimation in parallel MRI.

Main Results:

• Achieved high-quality MRI reconstructions independent of undersampling patterns.
• Demonstrated robust performance with out-of-distribution data (contrast variation).
• Enabled uncertainty quantification through a probabilistic interpretation of reconstructions.
• Showcased competitive performance against state-of-the-art methods in parallel MRI reconstruction.

Conclusions:

• The proposed generative prior-based framework enhances generalizability and interpretability in MRI reconstruction.
• The method provides flexible and robust reconstructions with reliable uncertainty quantification.
• This approach offers a promising alternative to end-to-end deep learning for clinical MRI applications.