Related Experiment Video
Updated: Aug 15, 2026
17:16
Registered Bioimaging of Nanomaterials for Diagnostic and Therapeutic Monitoring
Published on: December 9, 2010
Noise-Level Adaptive Diffusion Priors for Zero-Shot Low-Field MRI Quality Enhancement
Jiacai Cai1, Shoujin Huang1, Zihao Wang1
1College of Health Science and Environmental Engineering, Shenzhen Technology University, Shenzhen, China.
NMR in Biomedicine
|August 14, 2026
Summary
High-field diffusion models enhance low-field MRI quality without retraining, improving noise suppression and contrast. This method, Nila, offers a practical solution for accessible medical imaging across diverse systems.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Biophysics
Background:
- Low-field MRI (LF-MRI) offers improved accessibility but suffers from reduced signal-to-noise ratio, weaker contrast, and variable image quality.
- Limited low-field training data and inaccessibility of raw k-space data hinder LF-MRI's practical application.
- Existing enhancement methods struggle with the heterogeneity of LF-MRI systems.
Purpose of the Study:
- To adapt a high-field-trained diffusion model for LF-MRI quality enhancement without requiring low-field specific retraining.
- To develop a method robust to variations in scanner systems and data availability (k-space vs. magnitude-only).
- To evaluate the performance of the proposed method against established baselines and high-field references.
Main Methods:
- Adapted a pretrained high-field diffusion model for LF-MRI enhancement using noise-level adaptive measurement guidance.
- Incorporated phase augmentation to enable processing of magnitude-only DICOM inputs.
- Evaluated the method (Nila) on diverse datasets including healthy volunteers and patient cases across multiple field strengths (0.05 T to 3 T).
Main Results:
- Nila significantly improved visual image quality, noise suppression, and tissue contrast compared to raw LF input and baselines (BM3D, MiDiffusion, NAFNet).
- The method preserved lesion appearance in patient cases and demonstrated superior similarity to high-field references (best LPIPS, highest NMI).
- Explainable AI uncertainty maps highlighted spatial output variability, localized to tissue boundaries and ambiguous regions.
Conclusions:
- High-field diffusion priors can be effectively repurposed for LF-MRI enhancement, offering a practical and reusable solution.
- The proposed framework addresses limitations of LF-MRI data availability and system heterogeneity.
- This approach holds promise for improving the utility and accessibility of low-field MRI in clinical and research settings.
