Cycle-Consistent Zero-Shot Through-Plane Super-Resolution for Anisotropic Head MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Clinical magnetic resonance (MR) images are often anisotropic, with lower through-plane resolution than in-plane resolution.
• This anisotropy hinders the performance of processing pipelines requiring isotropic resolutions.
• Deep learning-based super-resolution (SR) methods risk generating unrealistic 'hallucinations' in high-resolution (HR) images.

Purpose of the Study:

• To develop a super-resolution method for anisotropic MR images that guarantees cycle-consistency with low-resolution observations.
• To address concerns about hallucinations in deep learning-based SR by ensuring fidelity to the original data.
• To construct and apply a specific linear forward map for the denoising diffusion null space model (DDNM) in the context of 2D MR acquisition.

Main Methods:

• Analyzed the forward problem in 2D MR acquisition to define an appropriate linear map ($A$).
• Trained a denoising diffusion probabilistic model on multi-dataset T₁-weighted (T₁-w) head MR images.
• Implemented the DDNM using the derived linear map $A$ for the MR image super-resolution task.

Main Results:

• The developed DDNM approach successfully generated exact cycle-consistent and realistic high-resolution MR images.
• The method demonstrated excellent qualitative and quantitative performance across diverse T₁-w MR datasets, including external and out-of-domain data.
• Evaluated results using both distortion and perceptual metrics, confirming the effectiveness of the SR technique.

Conclusions:

• The proposed DDNM framework with a tailored forward map effectively enhances the resolution of anisotropic MR images while maintaining cycle-consistency.
• This approach offers a reliable solution for super-resolution in clinical MR imaging, mitigating hallucination risks.
• The method shows strong generalizability across different datasets and imaging sites, indicating its clinical potential.