MRIFlow: Magnetic resonance image super-resolution based on normalizing flow and frequency prior

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Super-resolution (SR) is crucial for medical diagnosis, aiming to reconstruct high-resolution (HR) images from low-resolution (LR) ones.
• Traditional SR methods using L1/L2 loss yield high quantitative metrics but lack perceptual quality, potentially leading to misdiagnosis due to the ill-posed nature of SR.
• The need for perceptually superior and accurate HR medical images is critical for reliable clinical decision-making.

Purpose of the Study:

• To introduce MRIFlow, a novel normalizing flow-based method for generating high-resolution (HR) magnetic resonance (MR) images from low-resolution (LR) inputs.
• To enhance the perceptual quality and diagnostic accuracy of super-resolved MR images compared to existing methods.
• To address the ill-posed nature of SR in medical imaging by ensuring accurate and plausible HR image reconstruction.

Main Methods:

• MRIFlow utilizes a normalizing flow architecture for transforming LR MR images to HR MR images.
• The method incorporates frequency affine injectors, leveraging wavelet transforms via ScatterNet to embed crucial frequency information.
• The normalizing flow design allows for inverse operations, ensuring a well-defined mapping from LR to HR domains.

Main Results:

• MRIFlow demonstrated superior quantitative and qualitative results in super-resolution tasks compared to other normalizing flow-based SR methods.
• The method was evaluated on diverse MR image datasets, including IXI, NYU fastMRI, and LGG datasets.
• The frequency affine injectors effectively incorporated frequency information, contributing to improved image reconstruction.

Conclusions:

• MRIFlow offers a significant advancement in super-resolution for medical MR imaging, providing enhanced perceptual quality.
• The proposed method addresses key challenges in SR, such as the ill-posed problem and the need for accurate diagnostic information.
• MRIFlow shows promise for improving the accuracy and reliability of medical image analysis and diagnosis.