High-dimensional embedding network derived prior for compressive sensing MRI reconstruction

Zhang, Minghui; Li, Mengting; Zhou, Jiazhen; Zhu, Yanjie; Wang, Shanshan; Liang, Dong; Chen, Yang; Liu, Qiegen

doi:10.1016/j.media.2020.101717

Cited by 18 publications

(16 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The DeepRED method uses a denoising filter to regularize the DIP, which can avoid getting into local solutions and make the network learning stable. This is similar to the deep mean-shift prior (DMSP), which is used in imaging reconstruction recently [58]. Figure 10 shows the PSNR and SSIM curves varying with iteration numbers for DIPr, SGLDr, DeepREDr, DIPs, SGLDs and DeepREDs methods.…”

Section: Overfittingmentioning

confidence: 78%

Dynamic PET Image Denoising Using Deep Image Prior Combined With Regularization by Denoising

et al. 2021

View full text Add to dashboard Cite

The quantitative accuracy of positron emission tomography (PET) is affected by several factors, including the intrinsic resolution of the imaging system and inherently noisy data, which result in a low signal-to-noise ratio (SNR) of PET image. To address this problem, in this paper, we proposed a novel deep learning denoising framework aiming to enhance the quantitative accuracy of dynamic PET images via introduction of deep image prior (DIP) combined with Regularization by Denoising (RED), as such the method is labeled as DeepRED denoising. The network structure is based on encoder-decoder architecture and uses skip connections to combine hierarchical features to generate the estimated image. The network input can be random noise or other prior images (such as the patient's own static PET image), avoiding the need of high quality noiseless images, which is limited in PET clinical practice due to high radiation dose. Based on simulated data and real patient data, the quantitative performance of the proposed method was compared with conventional Gaussian filtering (GF), non-local mean (NLM), block-matching and 3D filtering (BM3D), DIP and stochastic gradient Langevin dynamics (SGLD) method. Overall, the proposed method can outperform other conventional methods in substantial visual as well as quantitative accuracy improvements (in terms of noise versus bias performance) with and without prior images.INDEX TERMS Positron emission tomography, deep neural networks, deep image prior, regularization by denoising.

show abstract

Section: Overfittingmentioning

confidence: 78%

Dynamic PET Image Denoising Using Deep Image Prior Combined With Regularization by Denoising

et al. 2021

View full text Add to dashboard Cite

show abstract

“…CS and its applications are reviewed in detail in [2]. CS is commonly used in such areas as super-resolution imaging [3][4][5], MRI fast image acquisition [6][7][8], and seismic data compression [9][10][11]. A signal possesses in these fields natural s-sparse representation corresponding to a specific domain, and the resultant regularized solutions of (1) with the minimal 1 norm leads to sufficiently correct outcomes.…”

Section: Many Kinds Of Modeling Problems Can Be Represented Asmentioning

confidence: 99%

“…i.e., output y t and control u t are bounded. Denote by ȳ ∈ l ∞ the desired output of the control plant described by Equation (7). The meaningful formulation of the problem is to build the polynomials α(•) and β(•), which are guaranteeing Inequality (10)-stabilizing of output and input-and asymptotic bound lim…”

Section: Controller For Discrete Non-minimum Phase Plant Under Unknown-but-bounded Disturbancementioning

confidence: 99%

Design of ℓ1 New Suboptimal Fractional Delays Controller for Discrete Non-Minimum Phase System under Unknown-but-Bounded Disturbance

et al. 2021

View full text Add to dashboard Cite

ℓ1-regularization methodologies have appeared recently in many pattern recognition and image processing tasks frequently connected to ℓ1-optimization in the control theory. We consider the problem of optimal stabilizing controller synthesis for a discrete non-minimum phase dynamic plant described by a linear difference equation with an additive unknown-but-bounded noise. Under considering the “worst” case of noise, the solving of these optimization problem has a combinatorial complexity. The choosing of an appropriate sufficiently high sampling rate allows to achieve an arbitrarily small level of suboptimality using a noncombinatorial algorithm. In this paper, we suggest to use fractional delays to achieve a small level of suboptimality without increasing the sampling rate so much. We propose an approximation of the fractional lag with a combination of rounding and a first-order fractional lag filter. The suggested approximation of the fractional delay is illustrated via a simulation example with a non-minimum phase second-order plant. The proposed methodology appears to be suitable to be used in various pattern recognition approaches.

show abstract

“…They utilized it in a gradient descent approach to perform Bayes risk minimization. DMSP is formulated as follows: By extending the naive DMSP with integration of multi-model aggregation and multi-channel network learning, Zhang et al [30] proposed a high-dimensional embedding network derived prior, and applied the learned prior to single-channel MRI reconstruction via variable augmentation technique.…”

Section: B From Dae To Dsmmentioning

confidence: 99%

“…Liu et al [29] employed enhanced denoising autoencoders (DAE) as prior to MRI reconstruction (DAEPRec), which leverages DAE network as an explicit prior and maximized the likelihood using gradient descent by backpropagating the autoencoder error for MRI reconstruction. Zhang et al [30] further exploited the gradient of the data density prior to reconstruction, termed EDMSPRec. Though these two works provided promising results, unlike VAE in [28], they are not directly related to generative models, lacking intuitive interpretation in the teeminology of machine learning.…”

Section: Introductionmentioning

confidence: 99%

Homotopic Gradients of Generative Density Priors for MR Image Reconstruction

Quan

Zhou

Zhu

et al. 2021

IEEE Trans. Med. Imaging

Self Cite

View full text Add to dashboard Cite

Deep learning, particularly the generative model, has demonstrated tremendous potential to significantly speed up image reconstruction with reduced measurements recently. Rather than the existing generative models that often optimize the density priors, in this work, by taking advantage of the denoising score matching, homotopic gradients of generative density priors (HGGDP) are proposed for magnetic resonance imaging (MRI) reconstruction. More precisely, to tackle the low-dimensional manifold and low data density region issues in generative density prior, we estimate the target gradients in higher-dimensional space. We train a more powerful noise conditional score network by forming high-dimensional tensor as the network input at the training phase. More artificial noise is also injected in the embedding space. At the reconstruction stage, a homotopy method is employed to pursue the density prior, such as to boost the reconstruction performance. Experiment results imply the remarkable performance of HGGDP in terms of high reconstruction accuracy; only 10% of the k-space data can still generate images of high quality as effectively as standard MRI reconstruction with the fully sampled data.

show abstract

High-dimensional embedding network derived prior for compressive sensing MRI reconstruction

Cited by 18 publications

References 38 publications

Dynamic PET Image Denoising Using Deep Image Prior Combined With Regularization by Denoising

Dynamic PET Image Denoising Using Deep Image Prior Combined With Regularization by Denoising

Design of ℓ1 New Suboptimal Fractional Delays Controller for Discrete Non-Minimum Phase System under Unknown-but-Bounded Disturbance

Homotopic Gradients of Generative Density Priors for MR Image Reconstruction

Contact Info

Product

Resources

About