MRI

MRIQT — Image-Quality Transfer for Ultra-Low-Field MRI

Slide between a portable ultra-low-field (0.064 T) neonatal MRI scan and its MRIQT enhancement — a physics-aware 3D diffusion model that transfers high-field image quality to bedside scans.

Code repository ↗
Ultra-low-field input, MRIQT and the high-field reference for one case (Al Abed et al., ISBI 2026).

Partners

University of BonnUniversity of BonnUniversity Hospital BonnUniversity Hospital BonnTechnical University of Munich

Research themes

Medical ImagingResponsible AI

Projects

MRIQT: Image-quality …

People

Malek Al AbedSebiha DemirAnne GroteklaesElodie GermaniShahrooz FaghihroohiHemmen SabirShadi Albarqouni

Slide to compare the portable ultra-low-field scan (uLF, right) with the MRIQT enhancement (left). Use the buttons to switch between two patients.

uLF (0.064 T) MRIQT MRIQT uLF (0.064 T)
Neonatal brain MRI from the University Hospital Bonn cohort. Ethical approval: Ethics Committee of the University of Bonn (Nr. 167/22).

arXiv · PDF · Code · Accepted at IEEE ISBI 2026, London

Why it matters

Infants change fast in the first weeks of life: their brains differ from adult brains in both anatomy and tissue contrast. In neonatal intensive care units, many infants have neurological conditions that need repeated MRI, yet standard high-field MRI means disconnecting them from life support and moving them to a scanner room — slow, often requiring sedation and exposure to loud noise.

Portable ultra-low-field MRI (0.064 T) brings the scanner to the bedside, but with low signal-to-noise and reduced diagnostic quality. MRIQT closes that gap: it performs 3D image-quality transfer (uLF → high-field) with a conditional diffusion model guided by physics-aware k-space simulation and an anatomy-preserving 3D perceptual objective. The cohort comprises neonates with an average age of two months and nine clinically relevant neurological pathologies, often occurring together.

Ultra-low-field input, MRIQT output and high-field reference for one case

How it works

  • Physics-aware degradation: realistic ultra-low-field scans are simulated from high-field ones in k-space.
  • 3D conditional diffusion: the model enhances the scan while preserving anatomy.
  • Guidance and an SNR-weighted 3D perceptual loss reduce artefacts and hallucinations.

Training — MRIQT learns from physics-based uLF simulations to reconstruct anatomically faithful images from noisy inputs:

MRIQT training

Inference — given a real uLF scan, MRIQT iteratively denoises a noisy initialisation while conditioning on the uLF measurement:

MRIQT inference

Results

Qualitative comparison of MRIQT with competing methods

Compared with competing methods on PSNR, LPIPS, MAE, MS-SSIM, SSIM and Pearson correlation:

Quantitative comparison

Related publications

How to cite
@inproceedings{alabed2026mriqt,
  title     = {MRIQT: Physics-Aware Diffusion Model for Image Quality Transfer in Neonatal Ultra-Low-Field MRI},
  author    = {Al Abed, Malek and Demir, Sebiha and Groteklaes, Anne and Germani, Elodie and Faghihroohi, Shahrooz and Sabir, Hemmen and Albarqouni, Shadi},
  booktitle = {IEEE International Symposium on Biomedical Imaging (ISBI)},
  year      = {2026},
  eprint    = {2511.13232},
  archivePrefix = {arXiv}
}