Abstract
Cross-field brain MRI translation changes field-dependent
image appearance while preserving anatomy. We study same-contrast
translation across five field strengths and three contrasts with one
shared model. DINOv3-PE-DPT keeps the public DINOv3 ViT-H+/16
weights frozen and adapts them to MRI synthesis through a learned
nine-channel input projection, MRI-specific conditioning, and a Dense
Prediction Transformer (DPT) decoder. The complete MRI adaptation
involves approximately 19.78M learned parameters, about 2.3% of the
861.2M-parameter checkpoint, while the final continuation updates only
215,304 shared low-rank DPT parameters.
Under the local MRIxFields2026 protocol with 180 reassembled 30-
slice central-slab outputs from 15 reporting subjects, DINOv3-PE-DPT
achieves SSIM 0.9023, nRMSE 0.2342, and LPIPS 0.1339, compared
with 0.7397 for the released StarGAN v2 prediction tree, 0.8365 for
source copy, and 0.8418 for the paired-training affine diagnostic. His-
torical branch comparisons show that DPT and LoRA variants already
reach 0.895–0.902 SSIM; the final continuation changes SSIM by about
0.0003 relative to its parent and has slightly worse LPIPS. Router
diagnostics show near-uniform mixtures and score-equivalent uniform
routing, so the supported finding is unified dense prediction using frozen
public DINOv3 weights and MRI-specific adaptation, not expert-router
specialization. The study is limited to an adaptively developed local
benchmark, without evidence of pathology preservation or external-
cohort generalization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MRIxFields2026_014.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=oSyaVphupx
BibTex
@InProceedings{WuYue_DINOv3PEDPT_MICCAISAT2026,
author = { Wu, Yuehan AND Wang, Lu AND Ge, Siyuan AND Liu, Bo AND Hwang, Jenq-Neng AND Zhu, Cheng Cheng},
title = { { DINOv3-PE-DPT: Frozen Public DINOv3 Weights for Cross-Field Brain MRI Translation } },
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}
}