Regularization by neural style transfer for MRI field-transfer reconstruction with limited data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning significantly advances MRI reconstruction but requires large, specific datasets.
• Reconstruction in data-limited scenarios remains a critical challenge.
• Existing methods like regularization by denoising (RED) use denoisers as priors.

Purpose of the Study:

• To introduce Regularization by Neural Style Transfer (RNST) for MRI field-transfer reconstruction.
• To address the challenge of MRI reconstruction in data-limited settings.
• To generate high-field-quality MRI images from low-field inputs without paired training data.

Main Methods:

• Developed RNST, integrating a neural style transfer (NST) engine with a denoiser.
• Leveraged style priors from NST to overcome data limitations.
• Tested RNST across diverse anatomical planes (axial, coronal, sagittal) and noise levels.

Main Results:

• RNST successfully reconstructed high-quality MRI images with superior clarity, contrast, and structural fidelity.
• The method demonstrated robustness even with imperfect alignment between style and content images.
• Achieved superior performance compared to lower-field references.

Conclusions:

• RNST offers a scalable, data-efficient solution for MRI field-transfer reconstruction.
• The framework shows significant potential for improving MRI in resource-limited clinical environments.
• RNST effectively utilizes style priors to address data scarcity in MRI reconstruction.