Abstract
Ultra-low-field MRI can improve access, but acquisitions
at tens of millitesla typically exhibit lower signal-to-noise ratio (SNR)
and less stable tissue contrast than clinical high-field scans. A physics-
conditioned latent generative method is presented for same-contrast
cross-field MRI synthesis. For MRIxFields Task 2, a shared model maps
0.1T inputs to 1.5, 3, 5, and 7T under explicit acquisition-parameter
conditioning. Paired slices are embedded by a variational autoencoder
(VAE). A residual predictor conditioned on (B0,TR,TE,TI) estimates
a coarse latent target, and a latent denoising diffusion bridge model
(DDBM) refines the source-to-target transition. The model is pretrained
on public paired 64mT/3T data and adapted on three paired challenge
subjects. On the official Task-2 evaluation, the method ranks first on
SSIM (primary metric; 0.878) and Dice (0.823). Acquisition parameters
enter the mapping, while anatomy is represented in a compact latent
space.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MRIxFields2026_019.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MRIxFields2026_019_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=P6C7dqAhTX
BibTex
@InProceedings{CheSai_PhysicsConditioned_MICCAISAT2026,
author = { Cheng, Sai AND Chen, Linxin AND Zong, Fangrong},
title = { { Physics-Conditioned Latent Diffusion Bridges for Ultra-Low-Field to Multi-Field MRI Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17274},
month = {pending},
page = {pending}
}