1 of 12

DIRECT LOW-FIELD MRI SUPER-RESOLUTION USING UNDERSAMPLED K-SPACE

Daniel Tweneboah Anyimadu, Mohammed M. Abdelsamea, and Ahmed Karam Eldaly

​

Department of Computer Science, University of Exeter, Exeter, UK

​

April 11, 2026

2 of 12

Contents

1

2

3

4

5

6

7

Motivation

Literature Review & Research Gaps

Aims & Objectives

Methodology

Experiments & Results

Conclusion & Future Works

References

3 of 12

Motivation

​

  • High-Field (HF) MRI (≥ 1.5T) is the gold standard for advanced medical imaging but comes with high costs (≥ $2 million), bulky infrastructure, and limited accessibility [1].

​

​

  • Low-Field (LF) MRI (<1T) offers affordable and portable imaging (Fig. 1) but suffers from longer scan times and reduced image quality (Fig. 1) compared to HF-MRI [2].

​

​

  • Moreover, LF-MRI’s reliability is hindered by the paucity of Uncertainty Quantification (UQ), affecting diagnostic accuracy and risk reduction [3].

Fig. 1: (a) Portable MRI system; (b) E.g., of images acquired on LF-MRI compared with conventional HF-MRI.

4 of 12

Literature review & Research gaps

Challenges of LF-MRI have prompted several techniques to emerge:

​

​

  • Longer Scan Time: Undersampling
  • MRI acquires data in k-space; undersampling accelerates scans but introduces artifacts [4] (Fig. 2); compressed sensing & deep learning (DL) help restore images but still miss fine details.

​

​

  • Reduced Image Quality: Image Quality Transfer (IQT)
  • IQT maps low → high-quality images [5] (Fig. 2) but is limited by its focus on the spatial domain.

​

​

  • UQ: Confidence Maps
  • UQ estimates confidence, vital for clinical trust, but is mainly explored in classification and segmentation [3] (Fig. 2).

​

​

  • Key gap: Despite advances, current LF-MRI solutions work in isolation (Fig. 3), limiting overall performance.

Fig. 2: (b) Undersampling Strats.; (c) IQT; (d) UQ.

Fig. 3: Reconstruction (k-space): Recovers missing frequency data → reduces aliasing artifacts.

​

Super-resolution (image): Enhances visual detail → improves resolution & contrast.

5 of 12

Aims & Objectives

These gaps highlight the need for continued innovation, motivating the following aims and objectives:

​

​

  • Aim: Develop a unified, uncertainty-aware DL framework integrating reconstruction, IQT, and UQ to improve LF-MRI image quality and diagnostic reliability.

​

​

Objectives:

  • Reconstruct HF-quality MRI from undersampled k-space while accelerating scan time.

​

  • Enhance image quality using frequency-domain IQT to improve anatomical detail.

​

  • Integrate UQ to generate confidence maps for reliable clinical interpretation.

​

  • Compare direct k-space reconstruction-enhancement with spatial-domain methods.

6 of 12

Methodology

 

 

7 of 12

Methodology

At the core of our unified framework lies a dual-channel U-Net comprising of encoder–decoder with skip connections that operates directly in k-space (Fig. 4).

Fig. 4: K-space dual channel U-Net for LF-MRI joint reconstruction, SR/IQT, and UQ from undersampled k-space.

Key modifications: (i) domain shift to k-space reconstruction, and (ii) joint real–imaginary convolutional processing that mimics complex convolution to exploit complementary MRI signal information.

8 of 12

Experiments & Results

The Human Connectome Project (HCP) dataset [6] was preprocessed for training as follows (Fig. 5):

​

​

  • Low-Field Simulation & Alignment:
  • HF (3T T1w volumes) were converted to realistic LF data (e.g., 0.3T) using a stochastic simulator (tissue-specific SNR) [5]. LF (depth - axial axis) were interpolated to ensure spatial and slice-wise alignment with HF references.

​

​

  • K-space Processing & Undersampling:
  • Data were transformed to complex k-space (FFT). Undersampling (Cartesian, pseudo-radial) was applied to LF data to simulate accelerated acquisition.

​

​

  • Model Input Preparation:
  • Paired undersampled LF and fully sampled HF data were converted into two-channel inputs (real + imaginary) and normalised for DL.

Fig. 5: (a) Data preprocessing pipeline; (b) Overall framework.

9 of 12

Experiments & Results

Performance is assessed quantitatively and qualitatively (Fig. 6) on the HCP dataset.

​

​

  • Quantitative Results: The k-space approach outperformed the spatial-domain and interpolation baseline in SSIM & PSNR (Table 1).

Fig. 6: Comparison of MRI recon., error map, and UQ across spatial and k-space InD domain.

10 of 12

Conclusion & Future Works

  • Research Hypothesis: The k-space domain enables unified DL for joint MRI recon., IQT, and UQ, outperforming the spatial domain.

​

​

  • Experimental Results: Improved anatomical detail and UQ-aware outputs, while maintaining image quality under aggressive undersampling comparable to full k-space reconstruction.

​

​

  • Potential Impact: Supports reliable, low-cost MRI in resource-limited settings while enhancing clinical trust.

​

Future work intends to optimise adaptive k-space sampling using uncertainty maps for efficient and high-quality reconstructions.

11 of 12

References

[1] P. Marzola, F. Osculati, and A. Sbarbati, “High field MRI in preclinical research,” Eur. J. Radiol., vol. 48, no. 2, pp. 165–170, Nov. 2003, doi: 10.1016/J.EJRAD.2003.08.007.

[2] T. C. Arnold, C. W. Freeman, B. Litt, and J. M. Stein, “Low-field MRI: Clinical promise and challenges,” J. Magn. Reson. Imaging, vol. 57, no. 1, pp. 25–44, Jan. 2023, doi: 10.1002/JMRI.28408;CTYPE:STRING:JOURNAL.

[3] Zou, K., Chen, Z., Yuan, X., Shen, X., Wang, M., Fu, H.: A review of uncertainty estimation and its application in medical imaging. Meta-Radiology 1(1), 100003 (2023).

[4] Lustig, M., Donoho, D., Pauly, J.M.: Sparse mri: The application of compressed sensing for rapid mr imaging. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 58(6), 1182–1195 (2007).

[5] Lin, H., Figini, M., D’Arco, F., Ogbole, G., Tanno, R., Blumberg, S.B., Ronan, L., Brown, B.J., Carmichael, D.W., Lagunju, I., et al.: Low-field magnetic resonance image enhancement via stochastic image quality transfer. Medical Image Analysis 87, 102807 (2023).

[6] Sotiropoulos, S.N., Jbabdi, S., Xu, J., Andersson, J.L.R., Moeller, S., Auerbach, E.J., Glasser, M.F., Hernández, M., Sapiro, G., Jenkinson, M., Feinberg, D.A., Yacoub, E., Lenglet, C., Essen, D.C.V., Uğurbil, K., Behrens, T.E.J.: Advances in diffusion mri acquisition and processing in the human connectome project. NeuroImage 80, 125–143 (2013), https://api.semanticscholar.org/CorpusID: 207187214.

[7] Souza, R., Frayne, R.: A hybrid frequency-domain/image-domain deep network for magnetic resonance image reconstruction. In: 2019 32nd SIBGRAPI conference on graphics, patterns and images (SIBGRAPI). pp. 257–264. IEEE (2019).

[8] Browne, M.W.: Cross-validation methods. Journal of mathematical psychology 44(1), 108–132 (2000).

12 of 12

Thank You & Discussion

Questions, feedback, or collaboration ideas are welcome!

​

Let’s advance accessible MRI together!