 Research Article
 Open Access
 Published:
A deep error correction network for compressed sensing MRI
BMC Biomedical Engineering volume 2, Article number: 4 (2020)
Abstract
Background
CSMRI (compressed sensing for magnetic resonance imaging) exploits image sparsity properties to reconstruct MRI from very few Fourier kspace measurements. Due to imperfect modelings in the inverse imaging, stateoftheart CSMRI methods tend to leave structural reconstruction errors. Compensating such errors in the reconstruction could help further improve the reconstruction quality.
Results
In this work, we propose a DECN (deep error correction network) for CSMRI. The DECN model consists of three parts, which we refer to as modules: a guide, or template, module, an error correction module, and a data fidelity module. Existing CSMRI algorithms can serve as the template module for guiding the reconstruction. Using this template as a guide, the error correction module learns a CNN (convolutional neural network) to map the kspace data in a way that adjusts for the reconstruction error of the template image. We propose a deep error correction network. Our experimental results show the proposed DECN CSMRI reconstruction framework can considerably improve upon existing inversion algorithms by supplementing with an errorcorrecting CNN.
Conclusions
In the proposed a deep error correction framework, any offtheshelf CSMRI algorithm can be used as template generation. Then a deep neural network is used to compensate reconstruction errors. The promising experimental results validate the effectiveness and utility of the proposed framework.
Background
MRI (Magnetic resonance imaging) is an important medical imaging technique with high resolution in soft tissues, low radiations, but the slow imaging speed is a major drawback of MRI. CS (Compressed sensing) theory [1, 2] has been a significant development of the signal acquisition and reconstruction process that has allowed for significant acceleration of MRI with less kspace measurements. The CSMRI problem can be formulated as the optimization
where x∈C^{N×1} is the complexvalued MRI to be reconstructed, F_{u}∈C^{M×N} is the undersampled Fourier matrix and y∈C^{M×1} (M≪N) are the kspace data measured by the MRI machine. The first data fidelity term ensures agreement between the Fourier coefficients of the reconstructed image and the measured data, while the second term regularizes the reconstruction to encourage certain image properties such as sparsity in a transform domain.
Recently, the compressed sensing MRI is approved by the FDA (Food and Drug Administration) to two main MRI vendors: GE and Siemens [3]. As the growing needs for application of compressed sensing MRI, improving reconstruction accuracy of the CSMRI is of great significance. In this paper, we propose a deep learning framework called DECN (deep error correction network) in which an arbitrary CSMRI inversion algorithm is combined with a deep learning error correction network. The network is trained for a specific inversion algorithm to exploit structural consistencies in the errors they produce. The final reconstruction is found by combining the information from the original algorithm with the error correction of the network.
A lot of previous works focus on proposing appropriate regularizations that lead to better MRI reconstructions. In the pioneering work of CSMRI called SparseMRI [4], this regularization adds an ℓ_{1} penalty on the wavelet coefficients and the total variation of the reconstructed image. Based on SparseMRI, more efficient optimization methods have been proposed to optimize this objective, such as TVCMRI (Total Variation ℓ_{1} Compressed MR Imaging) [5], RecPF (Reconstruction From Partial Fourier Data) [6] and FCSA (Fast Composite Splitting Algorithm) [7]. Variations on the wavelet penalty exploit geometric information of MRI, such as PBDW/PBDWS (Patch Based Directional Wavelet) [8, 9] and GBRWT (Graph Based Redundant Wavelet Transform) [10], for improved results. Dictionary learning methods [11–14] have also been applied to CSMRI reconstruction, as have nonlocal priors such as NLR (NonLocal Regularization) [15], PANO (Patch Based NonLocal Operator) [16] and BM3DMRI (BlockMatching 3D MRI) [17]. These previous works can be considered sparsitypromoting regularized CSMRI methods that are optimized using iterative algorithms. They also represent images using simple single layer features that are either predefined (e.g., wavelets) or learned from the data (e.g., dictionary learning).
Recently, deep learning approaches have been introduced for the CSMRI problem, achieving stateoftheart performance compared with conventional methods. For example, an endtoend mapping from input zerofilled MRI to a fullysampled MRI was trained using the classic CNN model in [18], or its residual network variant in [19]. In the residual network proposed in [19], a global shortcut is applied to enforce a UNet architecture input with a zerofilled MRI to learn the difference between the fullsampled MRI and its zerofilled one.
Although the work [19] shares the idea of residual learning with our approaches, there are some major differences between the two methods. In our model, the network design is motivated by exploiting the structural residual errors left by general reconstruction algorithms, the error correction module input with both the zerofilled MRI and guide image to learn the residual between the fullsampled MRI and guide image. If the error correction module is an identical mapping, the proposed DECN will be turned into the similar architecture to the compared model. However, our deep error correction network can be seen as a generalization of the compared network since the error correction module could be any offtheshelf CSMRI algorithms. Better reconstruction a guide module achieves, the smaller residual errors and the improvement under our framework are. Besides, for the input of the error correction module, the concatenation design of the zerofilled and guide MRI is motivated and justified by the observation that guide image produced by an offtheshelf MRI reconstruction algorithm is imperfect and lose details compared with zerofilled MRI, which is not discussed in the compared model.
Greater integration of the data fidelity term into the network has resulted in a DCCNN (Deep Cascade CNN) [20, 21]. The conventional iterative optimization is also unfolded as deep neural networks [22] called ADMMNet where the transform domain is learnable in a full supervised manner. The adversarial training strategy is also introduced in CSMRI [23–25] to help the reconstructed MRI more realistic. In DAGAN proposed in [23], frequency domain information is incorporated in the adversarial learning framework. A refinement Unet is designed as generator with a content loss to preserve details. A cyclic loss is introduced with a chain refinement strategy is proposed in [24] called RefineGAN for compressed sensing MRI. Similar GAN architecture is also evaluated in rapid MRI in [25].
Compared with previous models proposed for CSMRI inversion, deep learning is able to capture more intricate patterns within the data in both image domain and frequency domain [26, 27], which leads to their improved performance.
Previous work has also tried to exploit regularities in the reconstruction error in different ways. In the popular dynamic MRI reconstruction method kt FOCUSS (kt FOCal Underdetermined System Solver) [28, 29], the original signal is decomposed into a predicted signal and a residual signal. The predicted signal is estimated by temporal averaging, while the highly sparse residual signal has a l_{1}norm regularization. An iterative feature refinement strategy called IFRCS for CSMRI was proposed in [30]. The IFRCS method is an iterative optimization based approach. In certain iteration in this model, a sparsity promotion module using total variation (TV) is applied on the input noisy MR image to obtain a rough estimation first. Then a manually designed feature extractor is used on the rough estimation to generate a feature. The feature is calibrated by the difference between the noisy MR image input into the sparsity promotion module and the rough estimation to produce a refined feature. Then this refined feature is added back to the rough estimation to obtain the output in the iteration. The optimization iterates till it converges. Compared with our deep error correction network (DECN), the feature extraction of IFRCS is handcrafted, whereas a deep network can better extract features automatically in DECN. Also DECN model is more general because all compressed sensing MRI methods can be used to generate guide/rough image. The IFRNet is a variant of the IFRCS method using deep convolutional neural networks [31]. The IFRNet unrolls the IFRCS using deep learning architecture, which improves the transform domain and feature learning. The IFRNet shares similarities in using deep models for error correction with proposed DECN model although it is based on IFRCS formulation. In [32], the kspace measurements are divided into high and low frequency regions and reconstructed separately. In [33] the MR image is decomposed into a smooth layer and a detail layer which are estimated using total variation and wavelet regularization separately. In [34], the low frequency information is estimated using parallel imaging techniques. These methods each employ a fixed transform basis.
Methods
Problem formulation
Exploiting structural regularities in the reconstruction error of CSMRI is a good approach to compensate for imperfect modeling. Starting with the standard formulation of CSMRI in Eq. 1, we formulate our objective function as
where x_{p} is an intermediate reconstruction of the MRI. Due to the imperfect modeling, we model this intermediate reconstruction x_{p} as the summation of a “guidance” image \(\overline x_{p}\) and the error image of the reconstruction Δx_{p},
Substituting this into Eq. 2, we obtain
The guidance image \(\overline x_{p}\) is the reconstructed MRI using any chosen CSMRI method; thus x_{p} can be formed using existing software prior to using our proposed method for the final reconstruction. The reconstruction error Δx_{p} is between the ground truth fullsampled MRI x_{fs} and the reconstruction \(\overline x_{p}\). Since we don’t know this at testing time, we use training data to model this error image with a neural network \({f_{\theta }(\mathcal {X}) }\), where θ represents the network parameters and \(\mathcal {X}\) is the input to the network. Thus, Eq. 4 can be rewritten as
For a new MRI, after obtaining the guidance image \(\overline x_{p}\) (using a preexisting algorithm) and the welllearned mapping \(\Delta x_{p} = {f_{\theta }(\mathcal {X}) }\) (using a feedforward neural network trained on data), the proposed framework produces the final output MRI by solving the least square problem of Eq. 5.
Deep error correction network (DECN)
Following the formulation of our CSMRI framework above and in Fig. 1, we turn to a more detailed discussion of the optimization procedure. We next discuss each module of the proposed Deep Error Correction Network (DECN) framework.
Guide module
With the guide module, we seek a reconstruction of the MRI \(\overline x_{p}\) that approximates the fullysampled MRI using a standard “offtheshelf” CSMRI approach. We denote this as
We first illustrate with reconstructions for three CSMRI methods: TLMRI (transform learning MRI) [14], PANO (patchbased nonlocal operator) [16] and GBRWT (graphbased redundant wavelet transform) [10]. The PANO and GBRWT models achieve impressive reconstruction qualities because they use an nonlocal prior and adaptive graphbased wavelet transform to exploit image structures. In TLMRI, the sparsifying transform learning and the reconstruction are performed simultaneously in more efficient way than DLMRI (dictionary learning MRI) [11]. The three methods represent the stateoftheart performance in the nondeep CSMRI models. In Fig. 2, we show the reconstructions error for zerofilled (itself a potential reconstruction “algorithm”), TLMRI, PANO and GBRWT on a complexedvalued brain MRI using 30% Cartesian undersampling. The error display ranges from 0 to 0.2 with normalized data. The parameter setting will be elaborated in the Results section. We observe the reconstruction errors show high degree of sparsity and obvious image structures. From sparse representation theory, a more sparse signal can be recovered with less measurements [1, 35], which provide a solid ground that the sparse structural reconstruction error can be well approximated.
We also consider the representative deep learning DCCNN model [20] as the guide module. We also give the reconstruction error in Fig. 2. We observe the zerofilled, TLMRI, PANO, GBRWT and DCCNN models all suffer the structural reconstruction errors, while the DCCNN model achieves the highest reconstruction quality with minimal errors because of its powerful model capacity. Another advantage of this CNN model is that, once the network is trained, testing is very fast compared with conventional sparseregularization CSMRI models. This is because no iterative algorithm needs to be run for optimization during testing since the operations are a simple feed forward function of the input. We compare the reconstruction time of TLMRI, PANO, GBRWT and DCCNN for testing for Fig. 2 in Table 1. Note the DCCNN is implemented on GPU and other nondeep methods are implemented on CPU. However, the major reason for the difference in running speed among deep and nondeep models lies in the noniterative forward reconstruction property of the deep models when testing.
Error correction module
Using the guidance image \(\overline x_{p}\), we can train a deep error correction module on the residual. To perform this task, we need access during training to pairs of the true, fully sampled MRI x_{fs}, as well as its reconstruction \(\overline x_{p}\) found by manually undersampling the kspace of this image according to a predefined mask and inverting. We then optimize the following objective function over network parameter θ,
where \(\mathcal {Z}(y)\) indicates the reconstructed MRI using zerofilled and the input to the error correction module \(\mathcal {X}\) is the concatenation of the zerofilled MRI \(\mathcal {Z}(y)\) and the guidance MRI \(\overline x_{p}\) as shown in Fig. 1. Therefore, the errorcorrecting network is learning how to map the concatenation of the zerofilled, compressively sensed MRI and the guidance image to the residual of the true MRI using a corresponding offtheshelf CSMRI inversion algorithm. Now we give the rationales and explanations for the concatenation operation.
In the CSMRI inversions, the zerofilled MR images usually serve as the starting point in the iterative optimization. Although the iterative dealiasing can effectively remove the artifacts and achieve much more pleasing visual quality compared with zerofilled reconstruction, the distortion and information loss is inevitable in the reconstruction. To further illustrate this phenomenon, we compare the pixelwise reconstruction errors among the zerofilling reconstruction and other nondeep reconstruction models of the MR image in Fig. 2.
We take the difference between the absolute reconstruction error of the reconstructed MRI produced by compared CSMRI methods and zerofilled and only keep the nonnegative values, which can be formulated as
Where the operator (·)_{+} set the negative values to zero. We only keep the nonnegative values in the map, which results the filtered difference map. We show the corresponding filtered difference map m_{d} in Fig. 3 in the form of color map ranging from 0 to 0.1 with a 1D Cartesian 30% undersamling mask. On certain pixel of the reconstruction, if the guide reconstruction is less accurate compared with zerofilling, the difference on this pixel would be positive. Because we hope to find out if the zerofilled MRI is more accurate on some pixels, the negative values are not our interests and filtered. In the filtered difference map, the bright region means the better accuracy of zerofilled reconstruction. We observe the zerofilling reconstruction provide better reconstruction accuracy compared with different methods on some regions, indicating the information loss in the reconstruction occurs.
To alleviate the information loss in the guide module, we introduce the concatenation operation to utilize the information from both the zerofilled MR image and guidance image as the input to the error correction network. In later Discussion section, we further validate it by the ablation study.
We again note that the network \({f_{\theta } }\left (\mathcal {Z}(y),\overline x_{p} \right)\) is paired with a particular inversion algorithm invMRI(y), since each algorithm may have unique and consistent characteristics in the errors they produce. The network \({f_{\theta } }\left (\mathcal {Z}(y),\overline x_{p} \right)\) can be any deep learning network trained using standard methods.
Data fidelity module
After the error correction network is trained, for a new undersampled kspace data y for which the true x_{fs} is unknown, we use its corresponding guidance image \(\overline x_{p} = \text {invMRI}(y)\) and the approximated reconstructed error \({f_{\theta }}\left ({\mathcal {Z}(y),\overline x_{p}} \right)\) to optimize the data fidelity module by solving the following optimization problem
The data fidelity module is utilized in our proposed DECN framework to correct the reconstruction by enforcing greater agreement at the sampled kspace locations [11, 12]. Using the properties of the fast Fourier transform (FFT), we can simplify the optimization by working in the Fourier domain using the common technique described in, e.g., [12]. The optimal values for \(\hat x\) in kspace can be found pointwise. This yields the closedform solution
The regularization parameter α is usually set very small in the noisefree environment. We found that α=5e−5 worked well in our lownoise experiments.
Results
Data
In the experiment section, we present experimental results using complexvalued MRI datasets. The T1 weighted MRI dataset (size 256×256) is acquired on 40 volunteers with total 3800 MR images at Siemens 3.0T scanner with 12 coils using the FLASH (Fast Low Angle SHot) sequence (TR/TE = 55/3.6ms, 220 mm^{2} field of view, 1.5mm slice thickness). The SENSE (SENSitivity Encoding) reconstruction is introduced to compose the gold standard full kspace, which is used to emulate the singlechannel MRI. For SENSE reconstruction, each coil receives partial MRI signal and produce the corresponding parallel MRI images. Then the coil sensitivity maps are computed for each coil and used for generating the fullsampled MRI data by matrix inversion. The similar simulation setting can be found in [8]. We randomly select 75% MR images as training set, 5% as validation set and 20% as testing set. Informed consent was obtained from the imaging subject in compliance with the Institutional Review Board policy. The magnitude of the fullsampled MR image is normalized to unity by dot dividing the image by its largest pixel magnitude. The real and imaginary parts of a complex MRI data are input into the deep neural networks in twochannel manner [20].
We also validate our deep error correction network on the publicized MRI brain datasets MRBrainS13 [36]. The dataset is acquired at UMC Utrecht from patients. Each imaging subject is scanned to acquire multimodality MRI brain data including T1, T1IR and T2FLAIR modalities. Here we use T2FLAIR MRI throughout our paper. Bias correction has been applied on all scans and the data of each patient aligned. The voxel size is 0.958mm×0.958mm×3.00mm. There are total 5 scans in the training datasets. We use the fifth scan for testing and the rest 4 scans for training.
Undersampled kspace measurements are manually obtained via Cartesian and Random sampling mask with random phase encodes. Different undersampling ratios are adopted in the experiments.
Network architecture
For the deep guide module (i.e., learning \(\overline x_{p}\)), we use the CNN architecture called deep cascade CNN [20], where the nonadjustable data fidelity layer is also incorporated into the model. This guide module consists of four blocks. Each block is formed by four consecutive convolutional layers with a shortcut and a data fidelity layer. For each convolutional layer, except the last one within a block, there are total of 64 feature maps. We use ReLU (Rectified Linear Unit) [37] as the activation function.
For the error correction module (i.e., learning \(f_{\theta }(\mathcal {Z}(y),\overline x_{p})\)), we adopt the network architecture shown in Fig. 1. There are 18 convolutional layers with a skip layer connection as proposed in [38, 39] to alleviate the gradient vanish problem. We again adopt ReLU as the activation function, except for the last layer where the identity function is used to allow negative values. All convolution filters are set to 3×3 with stride set to 1.
Experimental setup
We train and test the two deep algorithms using Tensorflow [40] for the Python environment on a NVIDIA GeForce GTX 1080 with 8GB GPU memory. Padding is applied to keep the size of features the same. We use the Xavier method [41] to initialize the network parameters, and we apply ADAM [42] with momentum. The implementation uses the initialized learning rate 0.0001, firstorder momentum 0.9 and second momentum 0.99. The weight decay regularization parameter is set to 0.0005. The size of training batch is 4. We report our performance after 20000 training iteration of DCCNN guide module and 40000 iterations of error correction module.
In the guidance module, we implement the stateoftheart CSMRI models with the following parameter settings. In TLMRI [14], we set the data fidelity parameter 1e6/(256×256), the patch size 36, the number of training signals 256×256, the sparsity fraction 4.6%, the weight on the negative logdeterminat+Frobenius norm terms 0.2, the patch overlap stride 1, the DCT (Discrete Cosine Transform) matrix is used as initial transform operator, the iterations 50 times for optimization. The above parameter setting follows the advices from the author [14]. In PANO [16], we use the implementation with parallel computation provided by [16]. The data fidelity parameter is set 1e6 with zerofilled MR image as initial reference image. The nonlocal operation is implemented twice to yield the MRI reconstruction. In GBRWT [10], we set the data fidelity parameter 5×1e3. The Daubechies redundant wavelet sparsity is used as regularization to obtain the reference image. The graph is trained 2 times.
Experimental results
We evaluate the proposed DECN framework using PSNR and SSIM (structural similarity index) [43] as quantitative image quality assessment measures. We give the quantitative reconstruction results of all the test data on different undersampling patterns and different undersampling ratios in Table 2. We show the Cartesian 30% undersampling mask in Fig. 4 and the Random 20% undersampling mask in Fig. 5. We observe that DECN improved all offtheshelf CSMRI inversion methods on all the undersampling patterns. Since the 2D Random mask enjoys the more incoherence than the 1D Cartesian mask with the same undersampling ratio, the CSMRI achieves better reconstruction quality on the Random masks. We observe all different regular CSMRI inversions can be improved in PSNR and SSIM metrics. Also, we observe the plain DCCNN model already achieves better reconstruction accuracy than other compared models, leaving less structural errors for the error correction module, leading to less performance improvement. However, in the field of medical imaging where the quantitative accuracy matters, the small improvement in reconstruction quality is also valuable.
In Fig. 4, we show reconstruction results and the corresponding error images of an example from the test data on the 1D 30% undersampling mask. With local magnification on the red box, we observe that by learning the error correction module, the fine details, especially the lowcontrast structures are better preserved, leading to a better reconstruction.
In Fig. 5, we also compare the MR images produced by the TLMRI, PANO, GBRWT and DCCNN with their DECN counterparts on the 2D 20% undersampling mask. The results are consistent with our observation in Cartesian undersampling case.
We note the sparse and nonlocal prior based models are improved more significantly than the deep learning model DCCNN, which can be attributed to the highly accurate reconstruction of DCCNN, which leaves less structural residual information as demonstrated in Fig. 2.
We compare the DECNbased models with another two stateoftheart deep learning compressed sensing MRI methods: Residual UNet [19] and ADMMNet [22] in Figs. 4 and 5. The Residual UNet achieves the 34.39 dB in PSNR and 0.909 in SSIM compared to ADMMNet with 28.12dB in PSNR and 0.727 in SSIM on 1D Cartesian 30% mask. On 2D Random 20% mask, the Residual UNet achieves the 35.64 dB in PSNR and 0.878 in SSIM compared to ADMMNet with 37.05dB in PSNR and 0.951 in SSIM. We observe the DCCNNDECN outperforms both methods.
We also test our DECN approach on the publicized MRBrainS13 dataset with PANO being guide module and show the visual results in Fig. 6. We observe the error correction strategy efficiently improve the reconstruction quality on this datasets. The objective results on PSNR and SSIM are shown in Table 3.
Discussion
To validate the architecture of the proposed DECN model, we conduct the ablation study by comparing the DECN framework with other Baseline network architectures in Fig. 7, which we refer the model in Fig. 7 as DECNNICNEC (DECN with No Input Concatenation and Error Correction). With the guide module, a later cascaded CNN module learns the mapping from the prereconstructed MR image to the fullsampled MR image. Likewise, we name the models in Fig. 7 (DECNICNEC) and Fig. 7 (DECNNICEC). By comparing the DECNNICNEC framework with the DECNICNEC framework, we evaluate the benefit brought by the concatenating the zerofilled MR images and corresponding guide MR images as the input to compensate the information loss in the guide module. In Fig. 3, we give the illustration the information from zerofilled MR images and guide images can be shared. By comparing the DECNNICNEC framework with the DECNNICEC framework, we evaluate how the error correction strategy improves the reconstruction accuracy compared with simple cascade manner.
Here we show the experimental results of the ablation study using PANO with the Cartesian undersampling mask shown in Fig. 4 as the guide module. We give the averaged PSNR (peak signaltonoise ratio) and SSIM (structural similarity index) results over the testing datasets in Fig. 8 and the standard deviation. We observe the PANODECNICNEC and PANODECNNICEC both outperforms the PANODECNNICNEC with the similar margins about 0.2dB in PSNR. While the proposed PANODECN model with the input concatenation and error correction outperforms the PANODECNNICNEC about 0.5 dB in PSNR. We can obtain the similar observations with other CSMRI methods as guide module. The ablation study shows the input concatenation and error correction strategies can effectively improve the model performance in the DECN framework.
Conclusions
We have proposed a deep error correction framework for the CSMRI inversion problem. Using any offtheshelf CSMRI algorithm to construct a template, or “guide” for the final reconstruction, we use a deep neural network that learns how to correct for errors that typically appear in the chosen algorithm. Experimental results show that the proposed model achieves consistently improves a variety of CSMRI inversion techniques.
Availability of data and materials
The MRBrainS13 dataset is publicly available.
Abbreviations
 BM3D:

BlockMatching 3D MRI
 CNN:

Convolutional Neural Network
 CS:

Compressed Sensing
 DCCNN:

Deep Cascade Convolutional Neural Network
 DCT:

Discrete Cosine Transform
 DECNNICNEC:

DECN with No Input Concatenation and Error Correction
 DECN:

Deep Error Correction Network
 DLMRI:

Dictionary Learning Magnetic Resonance Imaging
 FCSA:

Fast Composite Splitting Algorithm
 FDA:

Food and Drug Administration
 FLASH:

Fast Low Angle SHot
 GBRWT:

Graph Based Redundant Wavelet Transform
 kt FOCUSS:

kt FOCal Underdetermined System Solver
 MRI:

Magnetic Resonance Imaging
 NLR:

NonLocal Regularization
 PANO:

Patch Based NonLocal Operator
 PBDW/PBDWS:

Patch Based Directional Wavelet
 PSNR:

Peak SignaltoNoise Ratio
 SSIM:

Structural SIMilarity index RecPF: Reconstruction From Partial Fourier Data
 ReLU:

Rectified Linear Unit
 TLMRI:

Transform Learning Magnetic Resonance Imaging
 TVCMRI:

Total Variation ℓ_{1} Compressed MR Imaging
References
Candès EJ, Romberg J, Tao T. Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information. IEEE Trans Inf Theory. 2006; 52(2):489–509.
Donoho DL. Compressed sensing. IEEE Trans Inf Theory. 2006; 52(4):1289–306.
Fessler JA. Medical image reconstruction: a brief overview of past milestones and future directions. arXiv preprint arXiv:1707.05927. 2017.
Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magn Reson Med. 2007; 58(6):1182–95.
Ma S, Yin W, Zhang Y, Chakraborty A. An efficient algorithm for compressed MR imaging using total variation and wavelets. In: IEEE Conference on Computer Vision and Pattern Recognition. IEEE: 2008. p. 1–8. https://doi.org/10.1109/cvpr.2008.4587391.
Yang J, Zhang Y, Yin W. A fast alternating direction method for TVL1L2 signal reconstruction from partial Fourier data. IEEE J Sel Topics Sig Process. 2010; 4(2):288–97.
Huang J, Zhang S, Metaxas D. Efficient MR image reconstruction for compressed MR imaging. Med Image Anal. 2011; 15(5):670–9.
Qu X, Guo D, Ning B, Hou Y, Lin Y, Cai S, Chen Z. Undersampled MRI reconstruction with patchbased directional wavelets. Magn Reson Imaging. 2012; 30(7):964–77.
Ning B, Qu X, Guo D, Hu C, Chen Z. Magnetic resonance image reconstruction using trained geometric directions in 2D redundant wavelets domain and nonconvex optimization. Magn Reson Imaging. 2013; 31(9):1611–22.
Lai Z, Qu X, Liu Y, Guo D, Ye J, Zhan Z, Chen Z. Image reconstruction of compressed sensing MRI using graphbased redundant wavelet transform. Med Image Anal. 2016; 27:93–104.
Ravishankar S, Bresler Y. MR image reconstruction from highly undersampled kspace data by dictionary learning. IEEE Trans Med Imaging. 2011; 30(5):1028.
Huang Y, Paisley J, Lin Q, Ding X, Fu X, Zhang XP. Bayesian nonparametric dictionary learning for compressed sensing MRI. IEEE Trans Image Process. 2014; 23(12):5007–19.
Liu Q, Wang S, Ying L, Peng X, Zhu Y, Liang D. Adaptive dictionary learning in sparse gradient domain for image recovery. IEEE Trans Image Process. 2013; 22(12):4652–63.
Ravishankar S, Bresler Y. Efficient blind compressed sensing using sparsifying transforms with convergence guarantees and application to magnetic resonance imaging. SIAM J Imaging Sci. 2015; 8(4):2519–57.
Dong W, Shi G, Li X, Ma Y, Huang F. Compressive sensing via nonlocal lowrank regularization. IEEE Trans Image Process. 2014; 23(8):3618–32.
Qu X, Hou Y, Lam F, Guo D, Zhong J, Chen Z. Magnetic resonance image reconstruction from undersampled measurements using a patchbased nonlocal operator. Med Image Anal. 2014; 18(6):843–56.
Eksioglu EM. Decoupled algorithm for MRI reconstruction using nonlocal block matching model: BM3DMRI. J Math Imaging Vision. 2016; 56(3):430–40.
Wang S, Su Z, Ying L, Peng X, Zhu S, Liang F, Feng D, Liang D. Accelerating magnetic resonance imaging via deep learning. In: IEEE International Symposium on Biomedical Imaging. IEEE: 2016. p. 514–7. https://doi.org/10.1109/isbi.2016.7493320.
Lee D., Yoo J., Ye J. C.Deep residual learning for compressed sensing MRI. In: IEEE International Symposium on Biomedical Imaging. IEEE: 2017. p. 15–8. https://doi.org/10.1109/isbi.2017.7950457.
Schlemper J, Caballero J, Hajnal JV, Price A, Rueckert D. A deep cascade of convolutional neural networks for MR image reconstruction. In: International Conference on Information Processing in Medical Imaging. Springer: 2017. p. 647–58. https://doi.org/10.1007/9783319590509_51.
Schlemper J, Caballero J, Hajnal JV, Price AN, Rueckert D. A deep cascade of convolutional neural networks for dynamic MR image reconstruction. IEEE Trans Med Imaging. 2018; 37(2):491–503.
Sun J, Li H, Xu Z, et al.Deep ADMMNet for compressive sensing MRI. In: Advances in Neural Information Processing Systems: 2016. p. 10–8.
Yang G, Yu S, Dong H, Slabaugh G, Dragotti PL, Ye X, Liu F, Arridge S, Keegan J, Guo Y, et al.Dagan: Deep dealiasing generative adversarial networks for fast compressed sensing MRI reconstruction. IEEE Trans Med Imaging. 2018; 37(6):1310–21.
Quan TM, NguyenDuc T, Jeong WK. Compressed sensing MRI reconstruction using a generative adversarial network with a cyclic loss. IEEE Trans Med Imaging. 2018; 37(6):1488–97.
Mardani M, Gong E, Cheng JY, Vasanawala SS, Zaharchuk G, Xing L, Pauly JM. Deep generative adversarial neural networks for compressive sensing mri. IEEE Trans Med Imaging. 2018; 38(1):167–79.
Zhu B, Liu JZ, Cauley SF, Rosen BR, Rosen MS. Image reconstruction by domaintransform manifold learning. Nature. 2018; 555(7697):487.
Eo T, Jun Y, Kim T, Jang J, Lee HJ, Hwang D. KIKInet: crossdomain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magn Reson Med. 2018; 80(5):2188–201.
Ye JC, Tak S, Han Y, Park HW. Projection reconstruction MR imaging using FOCUSS. Magn Reson Med. 2007; 57(4):764–75.
Jung H, Sung K, Nayak KS, Kim EY, Ye JC. kt FOCUSS: a general compressed sensing framework for high resolution dynamic MRI. Magn Reson Med. 2009; 61(1):103–16.
Wang S, Liu J, Liu Q, Ying L, Liu X, Zheng H, Liang D. Iterative feature refinement for accurate undersampled MR image reconstruction. Phys Med Biol. 2016; 61(9):3291.
Liu Y, Liu Q, Zhang M, Yang Q, Wang S, Liang D, et al. IFRNet: Iterative feature refinement network for compressed sensing mri. arXiv preprint arXiv:1909.10856. 2019.
Yang Y, Liu F, Xu W, Crozier S. Compressed sensing MRI via twostage reconstruction. IEEE Trans Biomed Eng. 2015; 62(1):110–18.
Park S, Park J. Compressed sensing MRI exploiting complementary dual decomposition. Med Image Anal. 2014; 18(3):472–86.
Sung K, Hargreaves BA. Highfrequency subband compressed sensing MRI using quadruplet sampling. Magn Reson Med. 2013; 70(5):1306–18.
Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Trans Image Process. 2017; 26(7):3142–55.
Mendrik AM, Vincken KL, Kuijf HJ, Breeuwer M, Bouvy WH, De Bresser J, Alansary A, De Bruijne M, Carass A, ElBaz A, et al.MRBrainS challenge: online evaluation framework for brain image segmentation in 3T MRI scans. Comput Intell Neurosci. 2015; 2015:1.
Nair V, Hinton GE. Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th International Conference on Machine Learning. JMLR: 2010. p. 807–14.
He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE: 2016. p. 770–8. https://doi.org/10.1109/cvpr.2016.90.
He K, Zhang X, Ren S, Sun J. Identity mappings in deep residual networks. In: Computer Vision ? ECCV 2016. Springer: 2016. p. 630–45. https://doi.org/10.1007/9783319464930_38.
Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, Devin M, Ghemawat S, Irving G, Isard M, et al.Tensorflow: a system for largescale machine learning. In: OSDI: 2016. p. 265–283.
Glorot X, Bengio Y. Understanding the difficulty of training deep feedforward neural networks. In: International Conference on Artificial Intelligence and Statistics. JMLR: 2010. p. 249–56.
Kingma D, Ba J. ADAM: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. 2014.
Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process. 2004; 13(4):600–12.
Acknowledgements
We would like to thank the editor and reviewers for valuable comments on this work.
Funding
This work was supported in part by the National Natural Science Foundation of China under Grants 61571382, 81671766, 61571005, 81671674, 61671309 and U1605252, in part by the Fundamental Research Funds for the Central Universities under Grant 20720160075, 20720180059, in part by the 2019 Principal Fund Innovation Team Cultivation Project under Grants 20720190116, in part by the CCFTencent open fund and, the Natural Science Foundation of Fujian Province of China (No.2017J01126). L. Sun conducted portions of this work at Columbia University under China Scholarship Council grant (No.201806310090). The funding bodies did not contribute to the design of the study, to the collection, analysis, nor interpretation of data nor in writing the manuscript.
Author information
Authors and Affiliations
Contributions
LS and XD developed new algorithms. LS was one of the major contributors in writing the manuscript. YW and ZF performed simulation and experimental study. YH collected and interpreted medical data. LS reviewed literatures. JP polished the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not Applicable.
Consent for publication
Not Applicable.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
About this article
Cite this article
Sun, L., Wu, Y., Fan, Z. et al. A deep error correction network for compressed sensing MRI. BMC biomed eng 2, 4 (2020). https://doi.org/10.1186/s4249002000375
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s4249002000375
Keywords
 Fast imaging
 Magnetic resonance imaging
 Deep convolutional neural network