TY - GEN
T1 - Ultra Low-Field MRI Enhancement via Conditional Diffusion Model
AU - Pang, Haowen
AU - Li, Xueqi
AU - Yan, Tianyi
AU - Ye, Chuyang
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2026
Y1 - 2026
N2 - Ultra-low-field magnetic resonance imaging (ULF-MRI) provides a promising low-cost and portable alternative to conventional high-field MRI (HF-MRI), particularly in resource-limited settings. However, its substantially degraded image quality remains a major obstacle to widespread clinical adoption. In this work, we propose a conditional diffusion model (CDM) for enhancing ULF-MRI image quality. During training, multi-modal ULF-MRI inputs (T1W, T2W, and FLAIR) are concatenated with noisy versions of the corresponding HF-MRI images and fed into a U-Net. The U-Net is trained to iteratively predict and remove the noise, thereby progressively synthesizing the HF-MRI. At inference, the multi-modal ULF-MRI inputs are concatenated with Gaussian noise and passed through the trained U-Net, which performs iterative denoising to generate the corresponding HF-MRI. Our framework is trained on paired 64mT–3T MRI data from the ULF-EnC Challenge. Experimental results show that the proposed method substantially improves structural detail and tissue contrast critical for clinical interpretation. These findings highlight the potential of CDM to bridge the quality gap between ULF-MRI and HF-MRI, enabling more accessible and reliable diagnostic imaging in low-resource environments through advanced generative modeling.
AB - Ultra-low-field magnetic resonance imaging (ULF-MRI) provides a promising low-cost and portable alternative to conventional high-field MRI (HF-MRI), particularly in resource-limited settings. However, its substantially degraded image quality remains a major obstacle to widespread clinical adoption. In this work, we propose a conditional diffusion model (CDM) for enhancing ULF-MRI image quality. During training, multi-modal ULF-MRI inputs (T1W, T2W, and FLAIR) are concatenated with noisy versions of the corresponding HF-MRI images and fed into a U-Net. The U-Net is trained to iteratively predict and remove the noise, thereby progressively synthesizing the HF-MRI. At inference, the multi-modal ULF-MRI inputs are concatenated with Gaussian noise and passed through the trained U-Net, which performs iterative denoising to generate the corresponding HF-MRI. Our framework is trained on paired 64mT–3T MRI data from the ULF-EnC Challenge. Experimental results show that the proposed method substantially improves structural detail and tissue contrast critical for clinical interpretation. These findings highlight the potential of CDM to bridge the quality gap between ULF-MRI and HF-MRI, enabling more accessible and reliable diagnostic imaging in low-resource environments through advanced generative modeling.
KW - Diffusion model
KW - Image synthesis
KW - Low-field MRI
UR - https://www.scopus.com/pages/publications/105043966517
U2 - 10.1007/978-3-032-23344-8_2
DO - 10.1007/978-3-032-23344-8_2
M3 - Conference contribution
AN - SCOPUS:105043966517
SN - 9783032233431
T3 - Lecture Notes in Computer Science
SP - 9
EP - 17
BT - Enhancing Ultra-Low-Field MRI with Paired High-Field MRI Comparisons for Brain Imaging - 1st International Challenge, ULF-EnC 2025, Held in Conjunction with MICCAI 2025, Proceedings
A2 - Chen, Zhaolin
A2 - Dayarathna, Sanuwani
A2 - Islam, Kh Tohidul
A2 - Peiris, Himashi
A2 - Zakavi, Parisa
A2 - Zhong, Shenjun
PB - Springer Science and Business Media Deutschland GmbH
T2 - 1st ULF-EnC 2025 Challenge on Enhancing Ultra-Low-Field MRI with Paired High-Field MRI Comparisons for Brain Imaging, held in conjunction with MICCAI 2025
Y2 - 23 September 2025 through 23 September 2025
ER -