Abstract
Ultra-low-field magnetic resonance imaging (MRI) offers an affordable and portable alternative to conventional high-field scanners, but its lower signal-to-noise ratio and limited availability of annotated data present significant challenges for developing robust learning-based methods. We propose MYRA (Multi-domain hYbrid MRI Representation Architecture), a self-supervised learning framework that jointly exploits complementary information from image space and raw k-space to learn transferable representations for low-field MRI. MYRA consists of three stages: (i) independent masked autoencoder pretraining on image and k-space domains, (ii) cross-domain representation alignment using a negative-free contrastive objective combined with a physics-informed Fourier-bridge consistency loss, and (iii) task-specific fine-tuning for downstream image restoration. Unlike existing approaches that primarily operate in a single domain, MYRA explicitly leverages the inherent Fourier relationship between image space and k-space to regularize representation learning. We evaluate the proposed framework on a real 47~mT portable brain MRI dataset across reconstruction and denoising tasks under limited supervision. Experimental results demonstrate that domain-matched pretraining improves downstream performance, while the proposed Fourier bridge stabilizes cross-domain optimization and enables effective label-efficient learning. These findings highlight the potential of physics-aware self-supervised representation learning for improving low-field MRI analysis and advancing accessible neuroimaging in resource-constrained settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_049.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=xod7pAQf8v
BibTex
@InProceedings{AkiOlu_MYRA_MICCAISAT2026,
author = { Akinmuleya, Oluwatobi Iyanuoluwa AND Koomson, Isaac Kofi AND Shambangu, Lucia Ndapewa AND Meschack-Augustin, Cirubyankabagurhi Njuci},
title = { { MYRA: Multi-domain Hybrid MRI Representation Architecture for Self-Supervised Low-Field MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}