Abstract
Multi-contrast MRI reconstruction (MCMR) leverages complementary information across contrasts to improve reconstruction under accelerated acquisition. However, many learning-based MCMR methods rely on fixed contrast ordering, predefined contrast roles, or fully sampled auxiliary references, which limits robustness when contrast availability varies. Such fixed designs can also encourage contrast-position shortcut learning, reducing effective cross-contrast information sharing. We propose a physics-consistent contrast-invariant framework for self-supervised learning. At its core is a contrast-level permutation training strategy, where each contrast is permuted together with its associated k-space data, masks, and sensitivity maps to preserve data-consistency alignment while discouraging contrast-position shortcuts. To enable content-driven fusion, we introduce Contrast Fusion-Invariant UNet (CFI-UNet), which computes cross-contrast interaction scores from local spatial and global frequency features and refines representations using spatial–frequency modules. We further incorporate prompt and curriculum learning to train a single model across acceleration factors and sampling trajectories while stabilizing high-acceleration self-supervised training. Experiments on M4Raw and BraTS demonstrate consistent improvements over self-supervised MCMR baselines, with improved reconstruction under 1D/2D undersampling and greater robustness to missing or unseen contrasts within a protocol. Our source code with trained weights is available at: https://github.com/chiew-group/PermuteMRI. Magnetic Resonance Imaging Multi-Contrast MRI MRI Reconstruction Deep Learning
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RIME_007.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=nYzJPIQteV
BibTex
@InProceedings{AnvKia_SelfSupervised_MICCAISAT2026,
author = { Anvari Hamedani, Kian AND Kadota, Brenden AND Chiew, Mark},
title = { { Self-Supervised Multi-Contrast MRI Reconstruction via Permutation-Driven Contrast Invariance } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}