Jul 2026· Magnetic Resonance in Medicine· Vol 96, pp. 2458 - 2473· 0 citations· 101 references
Medicine
TL;DR
A DL framework trained on real ULF and HF‐MRIs to address limitations and improve ULF‐derived brain volume analysis is presented.
Abstract
Ultra‐low‐field (ULF) MRI offers a promising path to accessible neuroimaging, with potential to address global healthcare disparities and advance population‐level brain health research. However, the inherently low signal‐to‐noise ratio (SNR), reduced spatial resolution, and altered tissue contrasts relative to conventional high‐field (HF) scans are significant barriers to ULF analysis and interpretation. While deep learning (DL) approaches have been proposed to enhance ULF image quality, many rely on synthetic training data due to the lack of available subject‐matched ULF and HF scans, introducing potential “domain shift” errors when applied to real acquisitions. Here, we present a DL framework trained on real ULF and HF‐MRIs to address these limitations and improve ULF‐derived brain volume analysis.
Low‐field MRI (LF‐MRI) improves imaging accessibility, but its practical utility is often limited by reduced signal‐to‐noise ratio, weaker tissue contrast, and variable image quality. These challenges are further compounded by the limited availability of low‐field training data across scanners and the fact that raw k‐space data are not always accessible in research and clinical workflows. In this study, we adapted a high‐field‐trained diffusion model for LF‐MRI quality enhancement without low‐field retraining or fine‐tuning. During inference, the pretrained diffusion prior was combined with noise‐level adaptive measurement guidance to suppress low‐field noise while maintaining consistency with measurement‐supported anatomical structures. With phase augmentation, this framework was further extended to magnitude‐only DICOM‐exported inputs when measured phase information was unavailable. The method was evaluated on 10 locally recruited healthy volunteers scanned at 0.05, 0.3, and 3 T, a public 0.064‐T healthy‐subject dataset with 10 paired cases, and patient cases acquired at 0.05 and 0.35 T. Compared with raw LF input and representative baselines including BM3D, MiDiffusion, and NAFNet, the proposed method (Nila) improved visual image quality, noise suppression, and tissue contrast, while preserving lesion‐region appearance in patient cases. On the 0.3 and 0.064 T datasets with coregistered 3‐T references, Nila achieved the best LPIPS and the highest or tied‐highest NMI across contrasts, indicating improved similarity to the high‐field reference. Multisample posterior inference further produced uncertainty maps that provide an explainable‐AI view of spatial output variability, with elevated variance mainly localized to tissue boundaries and ambiguous regions. These results suggest that high‐field diffusion priors can serve as practical and reusable tools for LF‐MRI enhancement across heterogeneous systems.
Jiacai Cai, Shoujin Huang, Zihao Wang et al.· NMR in Biomedicine· 0 citations
Recent updates to the diagnostic criteria of multiple sclerosis (MS) require whole‐brain T2*‐weighted (T2*w) imaging with submillimeter resolution to detect novel diagnostic biomarkers such as the central vein sign. However, to achieve the needed submillimeter spatial resolution, conventional T2*w 3D gradient‐echo scans sequences are limited by prohibitively long scan times for clinical use. Here, we evaluated a different approach based on a segmented 3D echo planar imaging (3D‐EPI) sequence, accelerated with 2D Controlled Aliasing in Parallel Imaging Results in Higher Acceleration (CAIPIRINHA) undersampling and denoised with a deep learning‐based network.
Sreekanth Madhusoodhanan Nair, Bryan Quah, Jaemin Song et al.· Magnetic Resonance in Medici...· 0 citations
Self-supervised methods showed superior performance over supervised and classical methods for low-field knee MR images, and validation in a larger dataset of knee at 0.55T is needed to further support this conclusion.
While slice-to-volume registration and super-resolution reconstruction laid the foundation for motion-corrected 3D T2-weighted fetal brain magnetic resonance imaging (MRI) more than two decades ago, advances in deep learning are now enabling automation across acquisition planning, segmentation, biometry, and image quality control. In this narrative review, we highlight these emerging techniques and analyse their strengths and limitations in the context of clinical translation. We examine the major barriers to widespread clinical adoption of artificial intelligence tools and outline future directions at the clinical interface that may further transform the diagnostic role of fetal MRI. Together, these developments underscore a shifting landscape towards more comprehensive, quantitative in-utero assessment, with the potential to enhance diagnostic accuracy and workflow efficiency, and broaden the clinical applications of fetal MRI in prenatal care.
A. Luis, A. Uus, L. Story et al.· Developmental Medicine & Chi...· 0 citations
Abstract Background Accurate attenuation correction (AC) is essential for quantitative positron emission tomography (PET). Conventional CT‐based AC provides reliable attenuation (μ‐) maps but adds radiation, introduces PET/CT misalignment artifacts, and is unavailable on stand‐alone PET systems. Existing deep learning (DL) methods use non‐attenuation corrected (NAC) PET data to predict CT‐like or directly AC‐PET images, but their dependence on emission characteristics limits generalizability across tracers and anatomical coverage, especially for long axial field‐of‐view PET. Purpose We propose incorporating tissue‐density information from dual‐view scout radiographs as supplementary input to NAC PET, acquired with a clinically significant radiation reduction relative to standard low‐dose CT. Methods Two DL methods were evaluated for generating transmission (Tr) images as surrogate μ‐maps for PET AC: (1) (NAC)‐to‐(Tr), using coronal and sagittal NAC PET slices, and (2) (NAC + Scout)‐to‐(Tr), combining NAC PET with anterior‐posterior and lateral scout views. 53 research scans across four tracers from the PennPET Explorer were used for training and testing, with three additional tracers included for testing. Results (NAC + Scout)‐to‐(Tr) improved quantitative accuracy, reducing NRMSE% to < 10% (versus up to 18% for (NAC)‐to‐(Tr)) and SUV biases to within ‐10%, with significant reductions in brain, liver, and muscle compared to (NAC)‐to‐(Tr). Out‐of‐distribution evaluation confirmed generalizability, maintaining SUV biases below 5%‐8% versus 10%‐13% for (NAC)‐to‐(Tr). In a longitudinal biodistribution study, the scout‐guided model showed consistent performance across repeat scans. Conclusions By leveraging routinely acquired scout scans, this approach enables accurate quantitative PET reconstruction without a full CT, substantially reducing radiation dose and misalignment artifacts.
Florence M. Muller, M. Daube-Witherspoon, M. J. Parma et al.· Medical Physics (Lancaster)· 0 citations
While image quality metrics suggested superior overall correction with MOS, an in vivo framework provided a more detailed characterization of in vivo performance differences, Notably, the framework detected a subtle improvement in FatNav performance with neck masking, an effect uncaptured by conventional image quality metrics.
Zakaria Zariry, Frank Lamberton, Robert Frost et al.· NMR in Biomedicine· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
