SynPoC: a high-quality generative diffusion model for transforming ultra-low-field point-of-care MRI using high-field MRI representations.
Authors
Affiliations (16)
Affiliations (16)
- Monash Biomedical Imaging, Monash University, Blackburn Road, Clayton, VIC, 3168, Australia.
- Australian National Imaging Facility, Brisbane, QLD, Australia.
- Data Science and AI, Monash University, Exhibition Walk, Clayton, VIC, 3800, Australia.
- Department of Neuroscience, Monash University, Clayton, VIC, Australia.
- Surgery, Monash University, Clayton, VIC, Australia.
- Herston Imaging Research Facility, University of Queensland, Brisbane, QLD, Australia.
- School of Electrical Engineering and Computer Science, University of Queensland, Brisbane, QLD, Australia.
- Department of Neurology, Royal Adelaide Hospital, Adelaide, SA, South Australia.
- South Australian Health and Medical Research Institute, Adelaide, SA, Australia.
- SA Medical Imaging, SA Health, Adelaide, SA, Australia.
- School of Clinical Science, Queensland University of Technology, Brisbane, QLD, Australia.
- David Hartley Chair of Radiology, Royal Perth Hospital, Perth, WA, Australia.
- Medical School, University of Western Australia, Perth, WA, Australia.
- Radiology, Alfred Hospital, Melbourne, VIC, Australia.
- Monash Biomedical Imaging, Monash University, Blackburn Road, Clayton, VIC, 3168, Australia. [email protected].
- Data Science and AI, Monash University, Exhibition Walk, Clayton, VIC, 3800, Australia. [email protected].
Abstract
Ultra-low-field (ULF) point-of-care (PoC) Magnetic Resonance Imaging (MRI) offers a promising pathway to improve accessibility in medical imaging due to its portability and lower cost. However, the diagnostic utility of ULF MRI is currently limited by lower image quality, particularly in signal-to-noise ratio, resolution, and contrast. To address this, we introduce SynPoC, a generative diffusion model designed to enhance ULF MRI by synthesizing high-field MRI-like images. SynPoC employs a conditional adversarial diffusion framework that leverages both noise and contrast-specific features to model inter-field representations. We evaluated SynPoC across a multi-site dataset of 180 participants, including both healthy individuals and patients with a variety of brain conditions. The enhanced images exhibited improved anatomical clarity and structural alignment with corresponding high-field MRI, as supported by quantitative and volumetric analyses. Our model demonstrates promise for image quality enhancement and research applications; however, as with other generative approaches, there is a non-zero risk of hallucinated or misleading features, particularly near low-SNR boundaries and fine structures. We therefore provide synchronized slice-by-slice comparison videos (3T, PoC, SynPoC) to aid reader inspection and emphasize that SynPoC is not intended for diagnostic decision-making without additional safeguards and validation. Further validation is warranted before diagnostic use.