Abstract
Synthesizing ultra-high-field magnetic resonance images from
routinely acquired or low-field scans is challenging because the appear-
ance gap depends jointly on pulse sequence and source field strength,
while paired 7T data are scarce. We present our submission to Task 1 of
the MRIxFields 2026 challenge, which treats T1-weighted (T1W), T2-
weighted (T2W), and T2-FLAIR images with complementary learning
strategies. T1W translation uses field-specific contrastive unpaired trans-
lation followed by structure-aware paired fine-tuning and late-checkpoint
averaging. T2W uses one field-conditioned supervised residual generator.
For T2-FLAIR, a supervised prediction is blended with a field-specific con-
trastive prediction. A low-capacity teacher-anchored stage further adapts
only the last residual blocks and decoder while constraining predictions
on unpaired retrospective scans. On the challenge validation server the
submitted system obtains mean SSIM 0.9095, nRMSE 0.2955, LPIPS
0.0776, and matched Dice 0.8809. A controlled comparison supports
the modality-adaptive design: one uniform field-conditioned supervised
generator applied to T1W reaches SSIM 0.9594 against 0.9812 for our
contrastive branch, losing at every source field. Leave-one-out experi-
ments show teacher anchoring improves all twelve held-out subject–field
combinations, though only by order 10−4, whereas augmentation and
a stand-alone 2.5D model do not generalize. We further quantify that
background masking does not inflate our metrics, and report containerized
inference at 15.6s per case.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MRIxFields2026_006.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=NieTvnVcGu
BibTex
@InProceedings{MiJun_ModalityAdaptive_MICCAISAT2026,
author = { Mi, Junsheng},
title = { { Modality-Adaptive Contrastive and Supervised Learning for Multi-Field MRI-to-7T 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}
}