Learning a preconditioner to accelerate compressed sensing reconstructions in MRI.
Ontology highlight
ABSTRACT: Purpose
To learn a preconditioner that accelerates parallel imaging (PI) and compressed sensing (CS) reconstructions.Methods
A convolutional neural network (CNN) with residual connections was used to train a preconditioning operator. Training and validation data were simulated using 50% brain images and 50% white Gaussian noise images. Each multichannel training example contains a simulated sampling mask, complex coil sensitivity maps, and two regularization parameter maps. The trained model was integrated in the preconditioned conjugate gradient (PCG) method as part of the split Bregman CS method. The acceleration performance was compared with that of a circulant PI-CS preconditioner for varying undersampling factors, number of coil elements and anatomies.Results
The learned preconditioner reduces the number of PCG iterations by a factor of 4, yielding a similar acceleration as an efficient circulant preconditioner. The method generalizes well to different sampling schemes, coil configurations and anatomies.Conclusion
It is possible to learn adaptable preconditioners for PI and CS reconstructions that meet the performance of state-of-the-art preconditioners. Further acceleration could be achieved by optimizing the network architecture and the training set. Such a preconditioner could also be integrated in fully learned reconstruction methods to accelerate the training process of unrolled networks.
SUBMITTER: Koolstra K
PROVIDER: S-EPMC9299023 | biostudies-literature |
REPOSITORIES: biostudies-literature
Similar Datasets
Project description:PurposeDesign of a preconditioner for fast and efficient parallel imaging (PI) and compressed sensing (CS) reconstructions for Cartesian trajectories.TheoryPI and CS reconstructions become time consuming when the problem size or the number of coils is large, due to the large linear system of equations that has to be solved in ℓ1 and ℓ2 -norm based reconstruction algorithms. Such linear systems can be solved efficiently using effective preconditioning techniques.MethodsIn this article we construct such a preconditioner by approximating the system matrix of the linear system, which comprises the data fidelity and includes total variation and wavelet regularization, by a matrix that is block circulant with circulant blocks. Due to this structure, the preconditioner can be constructed quickly and its inverse can be evaluated fast using only two fast Fourier transformations. We test the performance of the preconditioner for the conjugate gradient method as the linear solver, integrated into the well-established Split Bregman algorithm.ResultsThe designed circulant preconditioner reduces the number of iterations required in the conjugate gradient method by almost a factor of 5. The speed up results in a total acceleration factor of approximately 2.5 for the entire reconstruction algorithm when implemented in MATLAB, while the initialization time of the preconditioner is negligible.ConclusionThe proposed preconditioner reduces the reconstruction time for PI and CS in a Split Bregman implementation without compromising reconstruction stability and can easily handle large systems since it is Fourier-based, allowing for efficient computations.
| S-EPMC6283050 | biostudies-literature
Project description:Compressed sensing accelerates magnetic resonance imaging (MRI) acquisition by undersampling of the k-space. Yet, excessive undersampling impairs image quality when using conventional reconstruction techniques. Deep-learning-based reconstruction methods might allow for stronger undersampling and thus faster MRI scans without loss of crucial image quality. We compared imaging approaches using parallel imaging (SENSE), a combination of parallel imaging and compressed sensing (COMPRESSED SENSE, CS), and a combination of CS and a deep-learning-based reconstruction (CS AI) on raw k-space data acquired at different undersampling factors. 3D T2-weighted images of the lumbar spine were obtained from 20 volunteers, including a 3D sequence (standard SENSE), as provided by the manufacturer, as well as accelerated 3D sequences (undersampling factors 4.5, 8, and 11) reconstructed with CS and CS AI. Subjective rating was performed using a 5-point Likert scale to evaluate anatomical structures and overall image impression. Objective rating was performed using apparent signal-to-noise and contrast-to-noise ratio (aSNR and aCNR) as well as root mean square error (RMSE) and structural-similarity index (SSIM). The CS AI 4.5 sequence was subjectively rated better than the standard in several categories and deep-learning-based reconstructions were subjectively rated better than conventional reconstructions in several categories for acceleration factors 8 and 11. In the objective rating, only aSNR of the bone showed a significant tendency towards better results of the deep-learning-based reconstructions. We conclude that CS in combination with deep-learning-based image reconstruction allows for stronger undersampling of k-space data without loss of image quality, and thus has potential for further scan time reduction.
| S-EPMC9914543 | biostudies-literature
Project description:Following their success in numerous imaging and computer vision applications, deep-learning (DL) techniques have emerged as one of the most prominent strategies for accelerated MRI reconstruction. These methods have been shown to outperform conventional regularized methods based on compressed sensing (CS). However, in most comparisons, CS is implemented with two or three hand-tuned parameters, while DL methods enjoy a plethora of advanced data science tools. In this work, we revisit [Formula: see text]-wavelet CS reconstruction using these modern tools. Using ideas such as algorithm unrolling and advanced optimization methods over large databases that DL algorithms utilize, along with conventional insights from wavelet representations and CS theory, we show that [Formula: see text]-wavelet CS can be fine-tuned to a level close to DL reconstruction for accelerated MRI. The optimized [Formula: see text]-wavelet CS method uses only 128 parameters compared to >500,000 for DL, employs a convex reconstruction at inference time, and performs within <1% of a DL approach that has been used in multiple studies in terms of quantitative quality metrics.
| S-EPMC9388129 | biostudies-literature
Project description:PurposeTo propose a novel compressed sensing (CS) high spatial resolution functional MRI (fMRI) method and demonstrate the advantages and limitations of using CS for high spatial resolution fMRI.MethodsA randomly undersampled variable density spiral trajectory enabling an acceleration factor of 5.3 was designed with a balanced steady state free precession sequence to achieve high spatial resolution data acquisition. A modified k-t SPARSE method was then implemented and applied with a strategy to optimize regularization parameters for consistent, high quality CS reconstruction.ResultsThe proposed method improves spatial resolution by six-fold with 12 to 47% contrast-to-noise ratio (CNR), 33 to 117% F-value improvement and maintains the same temporal resolution. It also achieves high sensitivity of 69 to 99% compared the original ground-truth, small false positive rate of less than 0.05 and low hemodynamic response function distortion across a wide range of CNRs. The proposed method is robust to physiological noise and enables detection of layer-specific activities in vivo, which cannot be resolved using the highest spatial resolution Nyquist acquisition.ConclusionThe proposed method enables high spatial resolution fMRI that can resolve layer-specific brain activity and demonstrates the significant improvement that CS can bring to high spatial resolution fMRI. Magn Reson Med 76:440-455, 2016. © 2015 The Authors. Magnetic Resonance in Medicine published by Wiley Periodicals, Inc. on behalf of International Society for Magnetic Resonance in Medicine. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
| S-EPMC5253235 | biostudies-literature
Project description:The AI community has been paying attention to submodular functions due to their various applications (e.g., target search and 3D mapping). Learning submodular functions is a challenge since the number of a function's outcomes of N sets is 2 N . The state-of-the-art approach is based on compressed sensing techniques, which are to learn submodular functions in the Fourier domain and then recover the submodular functions in the spatial domain. However, the number of Fourier bases is relevant to the number of sets' sensing overlapping. To overcome this issue, this research proposed a submodular deep compressed sensing (SDCS) approach to learning submodular functions. The algorithm consists of learning autoencoder networks and Fourier coefficients. The learned networks can be applied to predict 2 N values of submodular functions. Experiments conducted with this approach demonstrate that the algorithm is more efficient than the benchmark approach.
| S-EPMC7249116 | biostudies-literature
Project description:PURPOSE:Conventional non-Cartesian compressed sensing requires multiple nonuniform Fourier transforms every iteration, which is computationally expensive. Accordingly, time-consuming reconstructions have slowed the adoption of undersampled 3D non-Cartesian acquisitions into clinical protocols. In this work we investigate several approaches to minimize reconstruction times without sacrificing accuracy. METHODS:The reconstruction problem can be reformatted to exploit the Toeplitz structure of matrices that are evaluated every iteration, but it requires larger oversampling than what is strictly required by nonuniform Fourier transforms. Accordingly, we investigate relative speeds of the two approaches for various nonuniform Fourier transform kernel sizes and oversampling for both GPU and CPU implementations. Second, we introduce a method to minimize matrix sizes by estimating the image support. Finally, density compensation weights have been used as a preconditioning matrix to improve convergence, but this increases noise. We propose a more general approach to preconditioning that allows a trade-off between accuracy and convergence speed. RESULTS:When using a GPU, the Toeplitz approach was faster for all practical parameters. Second, it was found that properly accounting for image support can prevent aliasing errors with minimal impact on reconstruction time. Third, the proposed preconditioning scheme improved convergence rates by an order of magnitude with negligible impact on noise. CONCLUSION:With the proposed methods, 3D non-Cartesian compressed sensing with clinically relevant reconstruction times (<2 min) is feasible using practical computer resources. Magn Reson Med 79:2685-2692, 2018. © 2017 International Society for Magnetic Resonance in Medicine.
| S-EPMC6755916 | biostudies-literature
Project description:PurposeHigh-resolution three-dimensional (3D) structural MRI is useful for delineating complex or small structures of the body. However, it requires long acquisition times and high SAR, limiting its clinical use. The purpose of this work is to accelerate the acquisition of high-resolution images by combining compressed sensing and parallel imaging (CSPI) on a 3D-GRASE sequence and to compare it with a (CS)PI 3D-FSE sequence. Several sampling patterns were investigated to assess their influence on image quality.MethodsThe proposed k-space sampling patterns are based on two undersampled k-space grids, variable density (VD) Poisson-disc, and VD pseudo-random Gaussian, and five different trajectories described in the literature. Bloch simulations are performed to obtain the transform point spread function and evaluate the coherence of each sampling pattern. Image resolution was assessed by the full-width at half-maximum (FWHM). Prospective CSPI 3D-GRASE phantom and in vivo experiments in knee and brain are carried out to assess image quality, SNR, SAR, and acquisition time compared to PI 3D-GRASE, PI 3D-FSE, and CSPI 3D-FSE acquisitions.ResultsSampling patterns with VD Poisson-disc obtain the lowest coherence for both PD-weighted and T2 -weighted acquisitions. VD pseudo-random Gaussian obtains lower FWHM, but higher sidelobes than VD Poisson-disc. CSPI 3D-GRASE reduces acquisition time (43% for PD-weighted and 40% for T2 -weighted) and SAR (∼45% for PD-weighted and T2 -weighted) compared to CSPI 3D-FSE.ConclusionsCSPI 3D-GRASE reduces acquisition time compared to a CSPI 3DFSE acquisition, preserving image quality. The design of the sampling pattern is crucial for image quality in CSPI 3D-GRASE image acquisitions.
| S-EPMC6619236 | biostudies-literature
Project description:Background4D flow magnetic resonance imaging (MRI) of CSF can make an important contribution to the understanding of hydrodynamic changes in various neurological diseases but remains limited in clinical application due to long acquisition times. The aim of this study was to evaluate the accuracy of compressed SENSE accelerated MRI measurements of the spinal CSF flow.MethodsIn 20 healthy subjects 4D flow MRI of the CSF in the cervical spine was acquired using compressed sensitivity encoding [CSE, a combination of compressed sensing and parallel imaging (SENSE) provided by the manufacturer] with acceleration factors between 4 and 10. A conventional scan using SENSE was used as reference. Extracted parameters were peak velocity, absolute net flow, forward flow and backward flow. Bland-Altman analysis was performed to determine the scan-rescan reproducibility and the agreement between SENSE and compressed SENSE. Additionally, a time accumulated flow error was calculated. In one additional subject flow of the spinal canal at the level of the entire spinal cord was assessed.ResultsAveraged acquisition times were 10:21 min (SENSE), 9:31 min (CSE4), 6:25 min (CSE6), 4:53 min (CSE8) and 3:51 min (CSE10). Acquisition of the CSF flow surrounding the entire spinal cord took 14:40 min. Bland-Altman analysis showed good agreement for peak velocity, but slight overestimations for absolute net flow, forward flow and backward flow (<?1 ml/min) in CSE4-8. Results of the accumulated flow error were similar for CSE4 to CSE8.ConclusionA quantitative analysis of acceleration factors CSE4-10 showed that CSE with an acceleration factor up to 6 is feasible. This allows a scan time reduction of 40% and enables the acquisition and analysis of the CSF flow dynamics surrounding the entire spinal cord within a clinically acceptable scan time.
| S-EPMC7364783 | biostudies-literature
Project description:PurposeTo explore the feasibility of MR Fingerprinting (MRF) to rapidly quantify relaxation times in the human eye at 7T, and to provide a data acquisition and processing framework for future tissue characterization in eye tumor patients.MethodsIn this single-element receive coil MRF approach with Cartesian sampling, undersampling is used to shorten scan time and, therefore, to reduce the degree of motion artifacts. For reconstruction, approaches based on compressed sensing (CS) and matrix completion (MC) were used, while their effects on the quality of the MRF parameter maps were studied in simulations and experiments. Average relaxation times in the eye were measured in 6 healthy volunteers. One uveal melanoma patient was included to show the feasibility of MRF in a clinical context.ResultsSimulation results showed that an MC-based reconstruction enables large undersampling factors and also results in more accurate parameter maps compared with using CS. Experiments in 6 healthy volunteers used a reduction in scan time from 7:02 to 1:16 min, producing images without visible loss of detail in the parameter maps when using the MC-based reconstruction. Relaxation times from 6 healthy volunteers are in agreement with values obtained from fully sampled scans and values in literature, and parameter maps in a uveal melanoma patient show clear difference in relaxation times between tumor and healthy tissue.ConclusionCartesian-based MRF is feasible in the eye at 7T. High undersampling factors can be achieved by means of MC, significantly shortening scan time and increasing patient comfort, while also mitigating the risk of motion artifacts.
| S-EPMC6519255 | biostudies-literature
Project description:PurposeChemical exchange saturation transfer is a novel and promising MRI contrast method, but it can be time-consuming. Common parallel imaging methods, like SENSE, can lead to reduced quality of CEST. Here, parallel blind compressed sensing (PBCS), combining blind compressed sensing (BCS) and parallel imaging, is evaluated for the acceleration of CEST in brain and breast.MethodsThe CEST data were collected in phantoms, brain (N = 3), and breast (N = 2). Retrospective Cartesian undersampling was implemented and the reconstruction results of PBCS-CEST were compared with BCS-CEST and k-t sparse-SENSE CEST. The normalized RMSE and the high-frequency error norm were used for quantitative comparison.ResultsIn phantom and in vivo brain experiments, the acceleration factor of R = 10 (24 k-space lines) was achieved and in breast R = 5 (30 k-space lines), without compromising the quality of the PBCS-reconstructed magnetization transfer rate asymmetry maps and Z-spectra. Parallel BCS provides better reconstruction quality when compared with BCS, k-t sparse-SENSE, and SENSE methods using the same number of samples. Parallel BCS overperforms BCS, indicating that the inclusion of coil sensitivity improves the reconstruction of the CEST data.ConclusionThe PBCS method accelerates CEST without compromising its quality. Compressed sensing in combination with parallel imaging can provide a valuable alternative to parallel imaging alone for accelerating CEST experiments.
| S-EPMC6497066 | biostudies-literature