Time‐Conditioned Zero‐Shot Self‐Supervised Reconstruction for Accelerated 3D Ultra‐Low‐Field MRI
Open Access
- 28 April 2026
- journal article
- Published by Wiley in Magnetic Resonance in Medicine
- Vol. 96 (3) , 1303-1312
Abstract
Purpose: Ultra‐low‐field (ULF) MRI provides a cost‐effective, portable imaging option but has relatively low SNR and long acquisition times compared to standard clinical scans. This study presents a time‐conditioned zero‐shot self‐supervised learning image reconstruction framework (ULF‐ZS‐SSL) to accelerate 3D‐acquired single‐coil ULF MRI without relying on external training data. In addition, for faster computation, a transfer‐learning (TL) variant (ULF‐ZS‐SSL‐TL) was implemented by pretraining on a small fully‐sampled ULF brain dataset and fine‐tuning on the target subject in a zero‐shot manner. Methods: This image reconstruction method combines a physics‐based data‐consistency step with a 3D residual network prior and sinusoidal time‐step embeddings to improve convergence speed. Data were acquired on a 47 mT Halbach‐based scanner using 3D turbo spin‐echo sequences with T 1 ‐, T 2 ‐, and inversion‐recovery–T 1 ‐weighted contrasts. Additional T 1 ‐weighted wrist scans were acquired to evaluate cross‐anatomy generalization. Both true and retrospectively undersampled data were compared with total variation (TV) and model‐based deep learning (MoDL). Results: The ULF‐ZS‐SSL method produced high‐quality reconstructions across all tested contrasts, outperforming zero‐filled and TV reconstructions, particularly at higher acceleration factors. Time‐step conditioning improved convergence speed, while ULF‐ZS‐SSL‐TL further accelerated the image reconstruction three‐fold, enabling full 3D reconstructions in about 3 min. Pretraining on brain data also worked well for wrist reconstructions, indicating cross‐anatomy generalization. Conclusion: The ULF‐ZS‐SSL framework enables accurate, training‐free reconstruction of undersampled single‐coil ULF MRI data, as does the ULF‐ZS‐SSL‐TL approach using minimal training data. The combination of physics‐based unrolling, time‐step conditioning, and transfer‐learning supports rapid and robust application in portable or resource‐limited ULF MRI systems.Keywords
Funding Information
- European Research Council (101021218)
- Nederlandse Organisatie voor Wetenschappelijk Onderzoek (18981)
This publication has 22 references indexed in Scilit:
- Pushing the limits of low‐cost ultra‐low‐field MRI by dual‐acquisition deep learning 3D superresolutionMagnetic Resonance in Medicine, 2023
- Low‐field MRI: Clinical promise and challengesJournal of Magnetic Resonance Imaging, 2022
- In vivo 3D brain and extremity MRI at 50 mT using a permanent magnet Halbach arrayMagnetic Resonance in Medicine, 2020
- Self‐supervised learning of physics‐guided reconstruction neural networks without fully sampled reference dataMagnetic Resonance in Medicine, 2020
- Low‐cost and portable MRIJournal of Magnetic Resonance Imaging, 2019
- Low‐field MRI: An MR physics perspectiveJournal of Magnetic Resonance Imaging, 2019
- Learning a variational network for reconstruction of accelerated MRI dataMagnetic Resonance in Medicine, 2017
- Sparse MRI: The application of compressed sensing for rapid MR imagingMagnetic Resonance in Medicine, 2007
- Generalized autocalibrating partially parallel acquisitions (GRAPPA)Magnetic Resonance in Medicine, 2002
- Advances in sensitivity encoding with arbitrary k‐space trajectoriesMagnetic Resonance in Medicine, 2001