Abstract
Low-field MRI has the potential to improve access to medical imaging, but its clinical utility is limited by reduced tissue contrast and spatial resolution compared with high-field MRI. Image quality transfer (IQT) offers a route to enhancing low-field images, but supervised approaches require paired low- and high-field data, which are difficult to acquire and may generalise poorly across scanners and protocols. In this work, we formulate low-field MRI enhancement as a zero-shot inverse problem using an unconditional diffusion prior trained only on high-quality MR images. This eliminates the need for paired low-/high-field training data and enables the same prior to be reused across MRI sequences by modifying only the measurement model. To improve measurement consistency, we introduce a clean-image-space data-consistency update using a nonlinear conjugate-gradient solver, enabling more accurate reconstruction under the nonlinear contrast transformation in the low-field measurement model. We evaluate the method on simulated low-field observations generated from Human Connectome Project T1-weighted brain MRI, in both 2D slice-wise and 3D volume-wise settings. The 3D formulation additionally enables enhancement of through-plane anatomical detail. Compared with baseline methods, the proposed method achieves higher PSNR and SSIM, with particularly clear improvements in the more challenging 3D setting. Qualitative evaluation on real 0.36T brain MRI further suggests improved tissue contrast and resolution. These preliminary results indicate that stronger clean-image-space data consistency can reduce hallucinations and improve zero-shot low-field MRI IQT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_063.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=DkWAHp94lS
BibTex
@InProceedings{HuaYul_Diffusionbased_MICCAISAT2026,
author = { Huang, Yuliang AND Tregidgo, Henry F. J. AND Figini, Matteo AND Ogbole, Godwin AND Alexander, Daniel C.},
title = { { Diffusion-based Zero-shot Image Quality Transfer for Low-Field MRI Enhancement with Nonlinear Conjugate Gradient } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}