[Paper Review] Low-Field Magnetic Resonance Image Enhancement using Undersampled k-Space

Daniel Tweneboah Anyimadu, Mohammed Abdalla+2 more|arXiv (Cornell University)|Mar 14, 2026

Advanced MRI Techniques and Applications0 citations

TL;DR

Proposes a k-space–driven deep learning framework (kSURF) that jointly reconstructs and super-resolves undersampled low-field MRI data with voxel-wise uncertainty, achieving HF-like quality from LF k-space.

ABSTRACT

Low-field magnetic resonance imaging (MRI) offers a cost-effective alternative for medical imaging in resource-limited settings. However, its widespread adoption is hindered by two key challenges: prolonged scan times and reduced image quality. Accelerated acquisition can be achieved using k-space undersampling, while image enhancement traditionally relies on spatial-domain postprocessing. In this work, we propose a novel deep learning framework based on a U-Net variant that operates directly in k-space to super-resolve low-field MR images directly using undersampled data while quantifying the impact of reduced k-space sampling. Unlike conventional approaches that treat image super-resolution as a postprocessing step following image reconstruction from undersampled k-space, our unified model integrates both processes, leveraging k-space information to achieve superior image fidelity. Extensive experiments on synthetic and real low-field brain MRI datasets demonstrate that k-space-driven image super-resolution outperforms conventional spatial-domain counterparts. Furthermore, our results show that undersampled k-space reconstructions achieve comparable quality to full k-space acquisitions, enabling substantial scan-time acceleration without compromising diagnostic utility.

Motivation & Objective

• Motivate faster LF-MRI via undersampling while preserving diagnostic detail.
• Develop a unified end-to-end model that reconstructs and enhances LF-MRI directly in k-space.
• Incorporate voxel-wise uncertainty to quantify reconstruction reliability.
• Compare k-space–driven approach to spatial-domain SR/IQT methods and assess robustness under distribution shifts.
• Demonstrate applicability to synthetic and real LF-MRI brain data with various undersampling patterns.

Proposed method

• Introduce a k-space dual-channel U-Net (kSURF) that processes real/imaginary k-space components to produce HF-like k-space outputs.
• Train in paired HF LF data using synthetic LF simulations from HF-HCP datasets with multiple undersampling patterns (Cartesian, pseudo-radial, 2D random).
• Optimize with a combined loss (MSE + MAE + L2 regularization) and employ cross-validated ensembles to estimate voxel-wise uncertainty as Var(predictions).
• Process data in patches (32×32×32) after FFT-based pre-processing to align LF and HF volumes for training.
• Evaluate both in spatial-domain SR/IQT (sIQT) baselines and in k-space, comparing PSNR/SSIM across InD and OOD test sets.
• Provide qualitative and quantitative results including uncertainty maps to assess reliability.

Experimental results

Research questions

• RQ1 Can a unified k-space–driven model reconstruct HF-like images from undersampled LF k-space with fidelity comparable to fully sampled HF acquisitions?
• RQ2 Does operating in the frequency domain preserve more critical frequency information than conventional spatial-domain SR methods under aggressive undersampling?
• RQ3 What is the impact of undersampling pattern and rate on reconstruction quality and uncertainty estimates?
• RQ4 How does the proposed kSURF method compare to existing spatial-domain SR/IQT and reconstruction baselines in both in-distribution and out-of-distribution scenarios?
• RQ5 Is there demonstrable improvement in real LF-MRI data (including pathological cases) when using kSURF with uncertainty quantification?

Key findings

• kSURF consistently achieves higher PSNR and SSIM than spatial-domain baselines across undersampling patterns and rates on synthetic InD data.
• kSURF attains robust performance under domain shift (OOD) and generally outperforms sIQT in both quantitative metrics and uncertainty reliability.
• Pseudo-radial undersampling yields the best quantitative results among tested patterns for kSURF.
• Uncertainty maps (voxel-wise variance) provide interpretable confidence about reconstructions and correlate with residual error maps, indicating reliable uncertainty quantification.
• Real LF-MRI data (0.3T) with healthy and Parkinsonian subjects show that LF–zero-filling degrades image quality, while kSURF improves reconstruction and provides meaningful uncertainty estimates.
• Across evaluations, kSURF outperforms competing methods such as LRTV, ESPCN, ISO U-Net, and 3D SR U-Net in both InD and, to a large extent, OOD scenarios.