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.
Abstract
Magnetic resonance imaging (MRI) is a widely used non-invasive medical imaging tool. Recent advancements include whole-body low-field MRI systems (0.55T) to address challenges like high cost and infrastructure demands of high-field MRI. However, low-field images suffer from low signal-to-noise ratio (SNR), which reduces image quality and limits spatial resolution. Although averaging multiple images can achieve the desired SNR while maintaining high spatial resolution, this approach increases total scan time which compromises patient comfort, elevates the risk of motion artifacts, and reduces clinical throughput. Post-processing denoising techniques, such as a patch-based low rank denoiser (PROST) and block match ing 3D filtering (BM3D) have been proposed to address this issue, and more recently, deep learning (DL) methods (both supervised and self-supervised). However, supervised DL requires large datasets, often unavailable for low-field MRI. Most notably, there is no clarify on which of these methods works best, nor a systematic comparison exists for kneeimaging at 0.55T. In this work, we address this by implementing and evaluating two classical methods (PROST, BM3D) and four learning-based approaches (self-supervised: Noisier2Noise,Blind2Unblind and Accelerated Deep Image Prior; supervised: DnCNN) on both public fastMRI (3T/1.5T) and In-house 0.55T knee MRI datasets. Performance is analyzed across configurations to determine optimal denoising strategies and whether self-supervised deep learning-based methods are more effective compared to classical and supervised DL based techniques. Results show that: 1) DnCNN outperforms all methods on the fastMRI database yet it lacks generalization when evaluated on low-field data. 2) On higher simulated levels of noise the self-supervised methods are more effective than classical methods, while on lower levels, classical denoising method are more effective. 3) On 0.55T low-field MRI data, Blind2Unblind is more effective than classical and supervised DL methods. In conclusion, in this study self-supervised methods showed superior performance over supervised and classical methods for low-field knee MR images. However, validation in a larger dataset of knee at 0.55T is needed to further support this conclusion. These findings aim to advance the development of practical denoising strategies that deliver high-quality images while enabling shorter acquisition times and broader accessibility of low-field knee MRI.
OBJECTIVES
To evaluate the diagnostic performance of low-field 0.55 T knee magnetic resonance imaging (MRI) with and without deep learning (DL)-based image denoising techniques compared to conventional 3 T MRI.
METHODS
This retrospective study included 33 knee MRIs from 26 patients (10 women, mean age 52 years; ±16.9 years) with knee pain (mean duration 2.7 years; ±4.8 years) who underwent both 0.55 T and 3 T MRI. DL-based denoising (ImT-MRD) was applied to all 0.55 T datasets. Four radiologists assessed image quality, anatomical landmark visibility (4-point Likert scale), and diagnostic confidence for pathology detection (10-point scale). Diagnostic accuracy, including sensitivity and specificity, was calculated using consensus reading as standard of reference. Agreement was performed using Lin's concordance correlation coefficient (CCC).
RESULTS
0.55 T without DL showed a moderate sensitivity of 0.83, a specificity of 0.82, and an overall diagnostic accuracy of 0.83. For 3 T, sensitivity, specificity, and diagnostic accuracy were 0.95, 0.98, and 0.97, respectively. 0.55 T with DL improved sensitivity to 0.97 and diagnostic accuracy to 0.98. Agreement with 3 T was excellent for 0.55 T with DL (CCC: 0.94) and substantial for 0.55 T without DL (CCC: 0.78). Image quality was significantly higher for 0.55 T with DL compared to 0.55 T without DL (2.92vs.2.55;p = 0.005), yet 3 T outperformed both (3.44;p < 0.001). Diagnostic confidence was significantly higher for 0.55 T with DL compared to 0.55 T without DL (8.21vs.7.66;p < 0.001), with 3 T demonstrating the highest confidence levels overall (8.60; all p < 0.001).
CONCLUSION
DL-based denoising of 0.55 T knee MRI significantly enhances the diagnostic performance, demonstrating results comparable to conventional 3 T, which were not achieved with 0.55 T MRI alone.
S. Ulas, Madeline Hess, Zheren Zhu et al.· European Journal of Radiolog...· 0 citations
Low-field magnetic resonance imaging (MRI) is an affordable medical imaging technique used to assess the structural and functional features of internal and external tissues and organs. However, its clinical effectiveness is often constrained by prolonged scan times and reduced image quality due to compromised signal-to-noise ratios. Deep learning (DL) has emerged as an effective solution for reconstruction of undersampled MRI data. This study presents a comprehensive comparative evaluation of six U-Net variants, including U-Net, Attention U-Net, Residual U-Net, Recurrent-Residual (R2) U-Net, Residual-Attention U-Net (RAU-Net), and U-Net++, for image domain reconstruction of low-field (0.3 Tesla) MRI data from the M4RAW dataset. Root Sum of Squares (RSS) magnitude images were used for training and testing. The models were trained on 1024 MRI volumes comprising 18,432 image slices. Reconstruction performance was evaluated under Cartesian undersampling at acceleration factors of 2, 4, 8, and 16, and estimated different statistical indices including Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Normalized Mean Squared Error (NMSE) and Coefficient of Determination (R2). The results demonstrate that the vanilla U-Net performs well at lower acceleration factors, achieving average SSIM and PSNR values of 0.9303 and 39.18 dB, respectively, at an acceleration factor of 2. In contrast, more complex architectures, particularly U-Net++ and RAU-Net, exhibit superior performance at higher acceleration factors. In particular, U-Net++ achieved average SSIM values of 0.8754 and 0.8731, and PSNR values of 35.21 dB and 35.06 dB at acceleration factors of 8 and 16, respectively. These findings inform the development of efficient and accessible imaging solutions, promote the broader adoption of low-field MRI, and encourage further innovation in cost-effective medical imaging technologies.
Akif Ahmed Nasif Purno, K. M. T. K. Siddiki, S. M. Chapal Hossain· Magnetic Resonance Imaging· 0 citations
The iterative three-operator splitting (TOS) method to solve the composite regularization problem and a composite denoising network (CDNet) that maintains consistent arithmetic structures with it are proposed, demonstrating the effectiveness of deep unrolling in accelerating iterative optimization.
David Muigai, E. Mwangi, Henry Kiragu· International Journal of Adv...· 0 citations
Portable low-field MRI systems are a promising complement to conventional high-field systems, enabling broader access to MRI. However, correspondence in cortical thickness estimates between low- and high-field MRI in young people remains limited despite its importance for neurodevelopment and psychopathology. To evaluate how multiple low-field image processing approaches improve cortical thickness correspondence with high-field MRI in a large sample of young individuals, we collected ultra-low-field (64mT) and high-field (3T) MRI data from a community sample of young people. We applied deep learning–based image processing approaches (SynthSR v1.0, SynthSR v2.0, recon-all-clinical, and recon-any) to low-field data acquired across multiple sequences (T1- and T2-weighted) and orientations (axial, coronal, sagittal, and multi-orientation), with and without resampling and/or co-registration. We assessed global, lobar, and regional cortical thickness correspondence with 3T MRI measures using Pearson and intraclass correlations. We compared pipelines using Steiger’s Z-tests and Fisher’s Z-tests. A total of 150 individuals (mean age, 18.63±5.07; 80 female) were included. We observed the highest global correspondence with recon-all-clinical applied to coronal T1-weighted images (r=0.40, pFDR=2.6e-05). At the lobar and regional levels, multi-orientation T2-weighted images processed with recon-all-clinical showed the highest correspondence across the greatest number of regions (4/12 lobes; 13/68 regions). The highest correspondence and largest improvements were in frontal, cingulate, and temporal regions, including the right pars triangularis (r=0.52, pFDR=4.78e-11; Z=4.78, pFDR=4.25e-06), right caudal anterior cingulate (r=0.47, pFDR=3.83e-09; Z=5.46, pFDR=1.32e-07), and left parahippocampal (r=0.58, pFDR=2.98e-14; Z=5.17, pFDR=6.01e-07). We observed significantly improved cortical thickness correspondence in low-field MRI in young people. The recon-all-clinical pipeline yielded moderate correspondence, particularly in frontal, cingulate, and temporal regions. Our results highlight the potential of low-field MRI as an affordable and scalable approach for assessing cortical thickness in young people.
Sun-Ho Choi, Julia Shaw, Rebecca Cooper et al.· bioRxiv· 0 citations
A deep-learning framework for image normalisation and noise reduction to elevate 0.35T brain MRI toward high-field standards with almost-perfect reader agreement is developed and validated and is a low-cost route to more equitable neuroimaging.
Rachmat Hidayat, Linda Purnama· Sriwijaya Journal of Radiol...· 0 citations
Low-field magnetic resonance imaging (MRI) systems are highly practical in resource-constrained scenarios due to their low cost and portability, which is regarded as an important alternative to conventional high-field MRI. However, they suffer from significant challenges caused by low signal intensity, resulting in the low signal-to-noise ratio (SNR) and bringing severe noise interference within the image. Due to the presence of complex composite noise in low-field MRI, conventional diffusion models, designed to restore image quality via generative processes based on Gaussian denoising, cannot effectively address the complex noise in this setting. To address this problem, in this article, we propose a frequency-informed diffusion network (FIDNet) for low-field MRI enhancement. The key innovation of FIDNet lies in its dual-domain design, in which frequency-domain information from the original k-space measurement signals are incorporated to guide the diffusion-based reconstruction in the image domain. To achieve this goal, we introduce a frequency-domain binary transformation (FDBT) module, which converts high-magnitude components of the K-space measurement data into a binary representation and maps them to a unified distribution that is compatible with image-domain features. By doing so, the incorporation of frequency-domain priors could facilitate the denoising process, which is particularly beneficial for addressing the complex and composite noise present in the image domain. In addition, to further remove the complex noise and preserve significant signals, we propose an adaptive Fourier filtering (AFF) block, which utilizes adaptive filters to dynamically adjust filtering parameters based on the characteristics of the MRI image. Generally speaking, the extensive experiments on various datasets demonstrate the proposed FIDNet achieves remarkable performance with great generalization capacity. Our code is available at https://github.com/zhoutao960906/FIDNet
Tao Zhou, Qi Liu, Ziru Li et al.· IEEE Transactions on Instrum...· 0 citations