Skip to main content

Compressed sensing based dynamic MR image reconstruction by using 3D-total generalized variation and tensor decomposition: k-t TGV-TD

Abstract

Purpose

Compressed Sensing Magnetic Resonance Imaging (CS-MRI) is a promising technique to accelerate dynamic cardiac MR imaging (DCMRI). For DCMRI, the CS-MRI usually exploits image signal sparsity and low-rank property to reconstruct dynamic images from the undersampled k-space data. In this paper, a novel CS algorithm is investigated to improve dynamic cardiac MR image reconstruction quality under the condition of minimizing the k-space recording.

Methods

The sparse representation of 3D cardiac magnetic resonance data is implemented by synergistically integrating 3D total generalized variation (3D-TGV) algorithm and high order singular value decomposition (HOSVD) based Tensor Decomposition, termed k-t TGV-TD method. In the proposed method, the low rank structure of the 3D dynamic cardiac MR data is performed with the HOSVD method, and the localized image sparsity is achieved by the 3D-TGV method. Moreover, the Fast Composite Splitting Algorithm (FCSA) method, combining the variable splitting with operator splitting techniques, is employed to solve the low-rank and sparse problem. Two different cardiac MR datasets (cardiac perfusion and cine MR datasets) are used to evaluate the performance of the proposed method.

Results

Compared with the state-of-art methods, such as k-t SLR, 3D-TGV, HOSVD based tensor decomposition and low-rank plus sparse method, the proposed k-t TGV-TD method can offer improved reconstruction accuracy in terms of higher peak SNR (PSNR) and structural similarity index (SSIM). The proposed k-t TGV-TD method can achieve significantly better and stable reconstruction results than state-of-the-art methods in terms of both PSNR and SSIM, especially for cardiac perfusion MR dataset.

Conclusions

This work proved that the k-t TGV-TD method was an effective sparse representation way for DCMRI, which was capable of significantly improving the reconstruction accuracy with different acceleration factors.

Peer Review reports

Introduction

In Magnetic Resonance Imaging (MRI), imaging speed is limited by slow acquisition of full k-space using magnetic field gradients [1]. Minimizing the k-space recording time without compromising image quality has been a main thrust of MR imaging research. With the advent of compressed sensing (CS) theory [2, 3], MR image reconstruction with sparsity-promoted regularization (e.g., â„“1-based regularization), termed as CS-MRI [4,5,6,7,8,9,10], has gained popularity for its high imaging speed. The effective exploitation of the signal sparsity enables the MR image reconstruction from far fewer k-space samples possible than conventional methods require, thus CS-MRI can significantly reduce the scan time. The compressed sensing theory has been successfully applied to both static and dynamic magnetic resonance imaging (dMRI) reconstructions [11,12,13,14].

In CS-MRI, the method used to sparsify the MR image plays an important role in the image reconstruction. The most used sparsity bases are predefined mathematical transforms, such as discrete cosine transform (DCT), and discrete wavelet transform (DWT). Recently, the singular value decomposition (SVD) method has been used as a data-adaptive sparsity basis in CS-MRI reconstruction [15, 16], and it has been found that the SVD-based method could significantly accelerate the reconstruction process and achieve better image quality than those commonly used sparsifying transforms (DCT and DWT). Majumdar et al. proposed to exploit the nuclear norm regularization to implement the CS-MRI reconstruction, and the results showed that the proposed reconstruction method was faster than other methods [6]. In addition, the linear combination of Total Variation (TV) and wavelet sparse regularization, known as TV-L1 problem, is very popular in many CS-MRI models [5, 6, 17], which can be considered as processing the MR image to be sparse by both the specific transform and finite-differences at the same time. Due to the stair-case artifacts caused by the conventional TV-based regularization [18, 19], several generalizations and extensions of TV have been introduced to improve the CS-MRI reconstruction accuracy, such as Total Generalized Variation (TGV) [18,19,20], Higher Degree Total Variation (HDTV) [21]. Nonlocal Total Variation (NLTV) [22,23,24] is another effective way to address the issue of stair-case artifacts. Although effective in practice, it involves higher computational complexity than the conventional TV method.

For dynamic MR image reconstruction, Ji, et al. adopted the difference between the reconstructed image and the reference image to represent the spatial sparsity [25]. However, when compared with the reference frame, the sparsity of the difference image got worse with the increase of the subsequent frame distance. To solve this problem, Majumdar took the difference between two adjacent sub-images as a sparse representation of the reconstructed MR image [13]. In addition, Usman put forward the concept of a sparse group of dynamic MRI, utilizing both MRI signal itself sparsity and the group structure information between signals [26, 27], which can effectively improve the image reconstruction quality. Moreover, a novel blind compressed sensing frame work was proposed to recover dynamic magnetic resonance images from undersampled measurements [28, 29], which has been proved to provide superior reconstruction performance in comparison to existing low rank and compressed sensing schemes. Recently, k-t SLR (k-t Sparisity and Low-Rank) method has been proposed to accelerate dynamic MRI by exploiting sparsity and low rank properties of the image data [30, 31]. To exploit the low-rank structure, the k-t SLR method reshaped the 3D dataset into a large 2D matrix through a two-step process: vectorize the 2D images in a dynamic sequence first and then concatenate them to form a matrix. In most of the existing dynamic CS-MRI methods, 2D/1D transforms were applied to solve the 3D dynamic problem, which, by treating the 3D data as a series of 2D images, unfolded the 3D dataset into a 2D matrix to explore the spatiotemporal redundancy [30,31,32,33]. In addition, Majumdar [34, 35] acted the dynamic MR image reconstruction problem as a least squares minimization regularized by lp-norm as the sparsity penalty and Schatten-q norm as the low-rank penalty sparsity, which can yield much better reconstruction results than k-t SLR method. However, reshaping a high-order tensor into a matrix or vector may neglect the inherent data redundancy, thus greatly degrading the reconstructed image quality. To promote the signal sparsity representation by exploring the redundancy of the high-dimension data format, Yu et al. proposed tensor decomposition-based sparsifying transform, that is, high-order Singular Value Decomposition (HOSVD) [36], which can outperform the conventional sparse recovery methods for high-dimensional cardiac imaging reconstruction accuracy given the same amount of k-space data set [37].

In this paper, we will further improve the HOSVD based CS-MRI method to synergistically integrate 3D-TGV algorithm and HOSVD-based Tensor Decomposition, termed as k-t TGV-TD method. In the proposed method, the low rank structure of the 3D dynamic cardiac MR data is performed by the HOSVD method, and the localized image sparsity is achieved by the 3D-TGV method. Meanwhile, the Fast Composite Splitting Algorithm (FCSA) method [6], combining the variable splitting with operator splitting techniques, is employed to solve the low-rank and sparse problem [38]. Two different cardiac MR datasets (cardiac perfusion and cine MR datasets) are used to evaluate the performance of the proposed method.

Theory of k-t TGV-TD method

In the proposed CS-MRI technique, 3D-TGV and HOSVD based tensor decomposition are used to promote the sparsity of the dynamic MR signals, and the k-t TGV-TD optimization problem can be formed as:

$$\mathop {\arg \min }\limits_{\chi } \left\{ {\left\| {A_{u} (\chi ) - b} \right\|_{2}^{2} + \lambda_{1} \cdot TGV_{\alpha }^{2} (\chi ) + \lambda_{2} \varphi (\chi )} \right\}$$
(1)

where \(A_{u}\) is undersampled Fourier operator of the MR image, b is undersampled measurement of k-space data, and \(\chi\) is a third order tensor used to represent the spatial–temporal 3D cardiac MR data. \(\lambda_{1} \;{\text{and}}\;\lambda_{2}\) are two positive regularization parameters that determine the trade-off between the data consistency and the sparsity regularization terms. \(\varphi (\chi )\) is the tensor decomposition, and \(TGV_{\alpha }^{2} (\chi )\) is the second order TGV penalty function.

Fast composite splitting algorithm processes the original and composite regularization problem into two simpler sub-problems, which are then solved by using the fast iterative shrinkage-threshold algorithm (FISTA). In this way, we could finally reconstruct the dynamic images via an iterative combination [39]. Specifically, the complex composite reconstruction problem in Eq. (1) can be decomposed into two simpler regularization subproblems, that is, TGV subproblem and TD subproblem, as shown in Eqs. (2) and (3):

$$\mathop {\arg \min }\limits_{\chi } \left\{ {\frac{1}{2}\left\| {A_{u} (\chi ) - b} \right\|_{2}^{2} + \lambda_{1} \cdot TGV_{\alpha }^{2} (\chi )} \right\}$$
(2)
$$\mathop {\arg \min }\limits_{\chi } \left\{ {\frac{1}{2}\left\| {A_{u} (\chi ) - b} \right\|_{2}^{2} + \lambda_{2} \varphi (\chi )} \right\}$$
(3)

The basic idea of FISTA is to build regularization for the linearized differentiable part of the objective function in each iteration [38,39,40]. Therfore, the subproblem Eqs. (2) and (3) can be extended into two parts respectively:

$$\min \left\{ {F(\chi ) \equiv f(\chi ) + g(\chi ):\chi \in C^{{I_{1} \times I_{2} \times I_{3} }} } \right\}$$
(4)

where \(f(\chi ) = \frac{1}{2}\left\| {A{}_{u}(\chi ) - b} \right\|_{2}^{2}\) is a smooth convex function which is continuously differentiable with Lipschitz constant Lf (usually large); and \(g(\chi ) = \lambda_{1} \cdot TGV_{\alpha }^{2} (\chi )\) or \(\lambda_{2} \varphi (\chi )\) is a continuous convex function which is nonsmooth. According to the FISTA algorithm, given a continuous function g(u) and any scalar L > 0, the proximal map associated with function \(g(\chi )\) can be built as follows:

$$prox_{L} \{ g(u),\chi \} = \arg \mathop {\min }\limits_{u} \{ g(u) + \frac{L}{2}\left\| {u - \chi } \right\|^{2} \}$$
(5)

Equations (2) and (3) are solved in an iterative fashion. Let X1 be the solution of the TGV subproblem Eq. (2) and X2 be the solution of the TD subproblem Eq. (3) respectively; in each k iteration, the solutions \(\chi_{k}\) to the overall problem Eq. (1) can be found by a linear combination as follows:

$$\chi_{k} = \frac{1}{2}(X_{1} + X_{2} )$$
(6)

The FCSA-based algorithm for solving the k-t TGV-TD problem-based CS-dMRI reconstruction can be described in the algorithm 1.

figure a

Materials and methods

To test the reconstruction performance of the proposed k-t TGV-TD method, two different MR raw datasets, i.e. cardiac perfusion and cardiac cine, were employed in this study. The cardiac perfusion MR datasets was obtained on a 3 T Siemens scanner with saturation recovery sequence (TR/TE = 2.5/1 ms, saturation recovery time = 100 ms) at the University of Utah [30]. It contained 70 frames, and the data from a single slice was acquired on a Cartesian grid with a k-space matrix of 90 × 190 (phase encoding × frequency encoding) at a temporal resolution of one heartbeat. The cardiac cine MR data was acquired on a 1.5 T Philips system at Yonsei University Medical Center [41, 42]. The dataset is composed of 25 frames of full k-space data. The matrix size for scanning is 256 × 256, which corresponds to 256 phase encoding steps and 256 samples in frequency encoding. The cardiac cine dataset was obtained using steady-state free precession (SSFP) sequence with a flip angle of 50 degree and TR = 3.45 ms. The field of view was 345 mm × 270 mm, and the slice thickness was 10 mm. In addition, the radial sampling pattern was used to undersample the k-space of these two datasets, which was simulated by rounding the sample locations to the nearest Cartesian location [30, 41].

In each experiment, the regularization parameters λ1 and λ2 in the problem (1) were determined by parameter sweeping. The following stopping criteria were adopted for all experimental settings: the tolerance as shown in the FCSA-based reconstruction algorithm was set as tol = 10–4, and the maximum number of iterations was 30. All reconstructions were implemented in the Matlab programming environment (Version 2017b, Mathworks, Natick, MA), and the experiments were performed on a personal computer with 3.6 GHz Intel Core i9-9900 K CPU, 32 GB of memory and Windows 10 operating system. In addition, we compared the proposed k-t TGV-TD reconstruction method with four state-of-the-art dynamic CS-MRI reconstruction methods, that is, HOSVD-based tensor decomposition method [37, 43], k-t SLR [30], 3D-TGV, and low-rank plus sparse method [44].

To further quantitatively evaluate the reconstruction methods, the peak SNR (PSNR) and structural similarity index (SSIM) were adopted in this work [45]. Furthermore, the reconstructed images and the corresponding error images (the absolute difference between reconstructed image and the full sampled MR image) were also compared visually. The PSNR and SSIM were formulated as follows:

$$PSNR = 20\log_{10} \left( {{{MAX_{{I_{ref(r)} }} } \mathord{\left/ {\vphantom {{MAX_{{I_{ref(r)} }} } {\sqrt {\frac{1}{mn}\sum\nolimits_{i = 1}^{m} {\sum\nolimits_{j = 1}^{n} {\left( {I_{ref} (r) - I(r)} \right)} } } }}} \right. \kern-\nulldelimiterspace} {\sqrt {\frac{1}{mn}\sum\nolimits_{i = 1}^{m} {\sum\nolimits_{j = 1}^{n} {\left( {I_{ref} (r) - I(r)} \right)} } } }}} \right)$$
(7)
$$SSIM\left( {I_{{{\text{ref}}}} ,I} \right) = {{\left( {2\mu_{{I_{ref} }} \mu_{I} + c_{1} } \right)\left( {2\sigma_{{I_{ref} I}} + c_{2} } \right)} \mathord{\left/ {\vphantom {{\left( {2\mu_{{I_{ref} }} \mu_{I} + c_{1} } \right)\left( {2\sigma_{{I_{ref} I}} + c_{2} } \right)} {\left( {\left( {\mu_{{_{{I_{ref} }} }}^{2} + \mu_{{_{I} }}^{{^{2} }} + c_{1} } \right)\left( {\sigma_{{I_{ref} }}^{2} + \sigma_{I}^{2} + c_{2} } \right)} \right)}}} \right. \kern-\nulldelimiterspace} {\left( {\left( {\mu_{{_{{I_{ref} }} }}^{2} + \mu_{{_{I} }}^{{^{2} }} + c_{1} } \right)\left( {\sigma_{{I_{ref} }}^{2} + \sigma_{I}^{2} + c_{2} } \right)} \right)}}$$
(8)

where \(MAX_{{I_{ref(r)} }}\) is the maximum signal intensity of \(I_{ref} (r)\). \(\mu_{{I_{ref} }}\) and \(\mu_{I}\) are mean signal values of \(I_{ref} (r)\) and \(I(r)\), \(\sigma_{{I_{ref} }}^{2}\) and \(\sigma_{I}^{2}\) are variance of \(I_{ref} (r)\) and \(I(r)\), \(\sigma_{{I_{ref} I}}\) is the covariance of \(I_{ref} (r)\) and \(I(r)\), \(c_{1}\) and \(c_{2}\) are two variables to stabilize the equation when the denominator is too small.

Results

Comparisons on the cardiac perfusion dataset

The proposed k-t TGV-TD method was employed to reconstruct the cardiac perfusion data with variously reduced k-space sampling data. Figure 1 displays the performance evaluation of k-t TGV-TD method in comparison with HOSVD, k-t SLR, 3D-TGV, and L + S [44] methods on the cardiac perfusion MR dataset with acceleration factor 6. The reconstructed MR image of one representative frame (13th frame) and the error map were provided as a comparison. As can be seen in Fig. 1, the proposed k-t TGV-TD method outperformed the HOSVD, k-t SLR, 3D-TGV and L + S methods in reducing artifacts, which was presented clearly in the error maps (the second row).

Fig. 1
figure 1

Performance evaluation of k-t TGV-TD method in comparison with HOSVD, k-t SLR, 3D-TGV, and L + S methods on the cardiac perfusion dataset. The first row is the reconstructed MR image with acceleration factor 6; the second row is the error map. a MRI reconstructed with fully sampled k space data; b–f reconstructed MRI and the corresponding error images by using HOSVD, k-t SLR, 3D-TGV, L + S, and k-t TGV-TD respectively

As listed in Table 1, the PSNR and SSIM of cardiac perfusion dataset were provided by using HOSVD, k-t SLR, 3D-TGV, L + S, and k-t TGV-TD with acceleration factors: 6, 8, and 10 respectively. It can be found that, compared with the other methods, the proposed k-t TGV-TD method can achieve significantly better reconstruction results in terms of PSNR and SSIM. In addition, notched box plots about PSNR and SSIMR comparison among the above five methods were provided for cardiac perfusion datasets, as shown in Fig. 2. Since the notches in the box plot do not overlap, it is concluded that, the proposed k-t TGV-TD method outperforms state-of-the-art methods with 95% confidence. Meanwhile, PSNR and SSIM of the reconstructed MR images by using k-t TGV-TD method were more stable than those by HOSVD, k-t SLR, and L + S methods.

Table 1 Comparisons in terms of PSNR and SSIM of different acceleration factors on the cardiac perfusion dataset
Fig. 2
figure 2

The PSNR and SSIM of the reconstructed cardiac perfusion images by using HOSVD, k-t SLR, 3D-TGV, L + S, and k-t TGV-TD, with acceleration factor 6. a PSNR; b SSIM

Comparisons on the cardiac cine dataset

Figure 3 shows the visual comparisons of the reconstructed results by the proposed k-t TGV-TD method, HOSVD, k-t SLR, 3D-TGV, and L + S methods with radial sampling pattern at reduction factor 6. The first row was the reconstructed MRI (13th frame of the cardiac cine data), and the second row showed the error map. Visually, as shown in Fig. 3, one can find that the proposed method outperformed the HOSVD, k-t SLR, 3D-TGV, and L + S methods in reconstructing MRI with better defined borders and less reconstruction artifacts. Reconstructions by using the HOSVD and k-t SLR methods were contaminated by aliasing artifacts and noises, and the image reconstructed by using 3D-TGV method was over-smoothed.

Fig. 3
figure 3

Performance evaluation of k-t TGV-TD method in comparison with HOSVD, k-t SLR, 3D-TGV, and L + S methods on the cardiac cine dataset. The first row is the reconstructed MR image with acceleration factor 6; the second row is the error map. a MRI reconstructed with fully sampled k-space data; b–f reconstructed MRI and the corresponding error images by using HOSVD, k-t SLR, 3D-TGV, L + S, and k-t TGV-TD respectively

As listed in Table 2, the PSNR and SSIM of cardiac cine dataset were provided by using HOSVD, k-t SLR, 3D-TGV, L + S, and k-t TGV-TD with acceleration factor: 4, 6, and 8 respectively. It can be found that, compared with the other methods, the proposed k-t TGV-TD method can achieve better reconstruction results in terms of PSNR and SSIM. In addition, notched box plots about PSNR and SSIMR comparison among the above five methods with acceleration factor 6 were provided for cardiac cine datasets, as shown in Fig. 4. It can be found obviously that the proposed k-t TGV-TD method outperforms the other methods with higher PSNR and SSIM in most of the frames.

Table 2 Comparisons in terms of PSNR and SSIM of different acceleration factors on the cardiac perfusion dataset
Fig. 4
figure 4

The PSNR and SSIM of the reconstructed cardiac cine images by using HOSVD, k-t SLR, 3D-TGV, L + S, and k-t TGV-TD, with acceleration factor 6. a PSNR; b SSIM

Discussion

In this work, based on combination of the tensor decomposition and 3D-TGV method, the k-t TGV-TD method was proposed to reconstruct highly undersampled cardiac perfusion and cine MR images. The tensor decomposition-based sparsity regularization method exploited both the intra and inter sparsity of each frame, therefore it was an effective way to make use of the three-dimensional redundancy in dynamic cardiac datasets. Moreover, the TGV method can effectively alleviate the staircase artifacts of TV based MR image reconstruction, and the 3D-TGV method can further apply the sparsity between and within the frames to improve the reconstruction accuracy.

From the reconstruction results, as shown in Figs. 1 and 3, it can be observed that the proposed method can outperform the HOSVD, k-t SLR, 3D-TGV and L + S methods in the investigated different dynamic cardiac datasets. The k-t TGV-TD method can reconstruct the MR images with less error artifacts than those by using HOSVD, k-t SLR, 3D-TGV and L + S methods. From the quantitative evaluation indexes PSNR and SSIM, advantages of the proposed k-t TGV-TD method over other methods were apparent.

In this work, the proposed k-t TGV-TD method explored the correlations and sparsity of the dynamic cardiac datasets, but did not integrate with partially parallel imaging (PPI). In addition, a combination of compressed sensing and parallel imaging was proposed to reconstruct the MR image [46, 47], which can further reduce the k-space acquisition. In the future work, we will consider combining the k-t TGV-TD method with the PPI reconstruction method to further improve dynamic cardiac MR image reconstruction quality at higher reduction factors.

Conclusion

In this paper, a novel technique, termed k-t TGV-TD, that combines the tensor decomposition and 3D total generalized variation, was proposed for dynamic cardiac MR imaging reconstruction. The method was evaluated with cardiac perfusion and cine datasets. The experimental results indicated that, compared with the HOSVD, k-t SLR, 3D-TGV and L + S methods, the proposed k-t TGV-TD method could achieve improved reconstruction accuracy in all the cases under investigation.

Availability of data and materials

The cardiac perfusion data was acquired at the University of Utah and publicly available via web-link: http://www.engineering.uiowa.edu/~jcb/Software/ktslr_matlab/Software.html. The MR data were acquired at Yonsei University Medical center in Korea and are publicly available via the web-link: https://github.com/jasonbme/k-t-focuss

Abbreviations

MRI:

Magnetic Resonance Imaging

DCMRI:

Dynamic cardiac MR imaging

CS:

Compressed sensing

dMRI:

Dynamic magnetic resonance imaging

HOSVD:

High order singular value decomposition

TV:

Total variation

TGV:

Total generalized variation

DCT:

Discrete cosine transform

DWT:

Discrete wavelet transform

SVD:

Singular value decomposition

HDTV:

Higher degree total variation

NLTV:

Nonlocal total variation

k-t SLR:

K-t sparisity and low-rank method

FCSA:

Fast composite splitting algorithm

SSFP:

Steady-state free precession

FOV:

Field of view

FISTA:

Fast iterative shrinkage-threshold algorithm

References

  1. Westbrook C, Talbot J. MRI in practice (5th edn). New York: Wiley-Blackwell; 2018.

    Google Scholar 

  2. Donoho DL. Compressed sensing. IEEE Trans Inf Theory. 2006;52(4):1289–306. https://doi.org/10.1109/tit.2006.871582.

    Article  Google Scholar 

  3. Lustig M, Donoho DL, Santos JM, Pauly, et al. Compressed sensing MRI. IEEE Signal Process Mag 2008, 25(2), 72–82, doi:https://doi.org/10.1109/MSP.2007.914728

  4. 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. https://doi.org/10.1002/mrm.21391.

    Article  PubMed  Google Scholar 

  5. Shahdloo M, Ilicak E, Tofighi M, et al. Projection onto epigraph sets for rapid self-tuning compressed sensing MRI. IEEE Trans Med Imaging. 2019;38(7):1677–89. https://doi.org/10.1109/tmi.2018.2885599.

    Article  PubMed  Google Scholar 

  6. Yang G, Yu SM, Dong H, et al. DAGAN: deep de-aliasing generative adversarial networks for fast compressed sensing MRI reconstruction. IEEE Trans Med Imaging. 2017;37(6):1310–21. https://doi.org/10.1109/TMI.2017.2785879.

    Article  Google Scholar 

  7. Baron CA, Dwork N, Pauly JM, et al. Rapid compressed sensing reconstruction of 3D non-Cartesian MRI. Magn ResonMed. 2018;79(5):2685–92. https://doi.org/10.1002/mrm.26928.

    Article  Google Scholar 

  8. Ye JC. Compressed sensing MRI: a review from signal processing perspective. BMC Biomed Eng. 2019;1:1–17. https://doi.org/10.1186/s42490-019-0006-z.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Bustin A, Fuin N, Botnar RM, et al. From compressed-sensing to artificial intelligence-based cardiac MRI reconstruction. Front Cardiovasc Med. 2020;7:17. https://doi.org/10.3389/fcvm.2020.00017.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Delattre BMA, Boudabbous S, Hansen C, et al. Compressed sensing MRI of different organs: ready for clinical daily practice? Eur Radiol. 2020;30(1):308–19. https://doi.org/10.1007/s00330-019-06319-0.

    Article  PubMed  Google Scholar 

  11. Liu Y, Liu Q, Zhang M, et al. IFR-Net: Iterative feature refinement network for compressed sensing mri. IEEE Trans Comput Imaging. 2019;6:434–46. https://doi.org/10.1109/TCI.2019.2956877.

    Article  Google Scholar 

  12. Sun L, Fan Z, Fu X, et al. A deep information sharing network for multi-contrast compressed sensing MRI reconstruction. IEEE Trans Image Process. 2019;28(12):6141–53. https://doi.org/10.1109/TIP.2019.2925288.

    Article  PubMed  Google Scholar 

  13. Liu Y, Liu T, Liu J, Zhu C. Smooth robust tensor principal component analysis for compressed sensing of dynamic MRI. Pattern Recognit. 2020;102: 107252. https://doi.org/10.1016/j.patcog.2020.107252.

    Article  Google Scholar 

  14. Zhao B, Haldar JP, Christodoulou AG, et al. Image reconstruction from highly undersampled (k, t)-space data with joint partial separability and sparsity constraints. IEEE Trans Med Imaging. 2012;31(9):1809–20. https://doi.org/10.1109/tmi.2012.2203921.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Giese D, Schaeffter T, Kozerke S. Highly undersampled phase-contrast flow measurements using compartment-based k-t principal component analysis. Magn Reson Med. 2013;69(2):434–43. https://doi.org/10.1002/mrm.24273.

    Article  PubMed  Google Scholar 

  16. Velikina JV, Samsonov AA. Reconstruction of dynamic image series from undersampled MRI data using data-driven model consistency condition (MOCCO). Magn Reson Med. 2015;74(5):1279–90. https://doi.org/10.1002/mrm.25513.

    Article  PubMed  Google Scholar 

  17. Tsaig Y, Donoho DL. Extensions of compressed sensing. Signal Process. 2006;86(3):549–71. https://doi.org/10.1016/j.sigpro.2005.05.029.

    Article  Google Scholar 

  18. Bredies K, Kunisch K, Pock T. Total generalized variation. SIAM J Imaging Sci. 2010;3:492–526. https://doi.org/10.1137/090769521.

    Article  Google Scholar 

  19. Knoll F, Bredies K, Pock T, et al. Second order total generalized variation (TGV) for MRI. Magn Reson Med. 2011;65(2):480–91. https://doi.org/10.1002/mrm.22595.

    Article  PubMed  Google Scholar 

  20. Guo W, Qin J, Yin W. A new detail-preserving regularization Scheme. SIAM J Imaging Sci. 2014;7:1309–34. https://doi.org/10.1137/120904263.

    Article  Google Scholar 

  21. Hu Y, Jacob M. Higher degree total variation (HDTV) regularization for image recovery. IEEE Trans Image Process. 2012;21(5):2559–71. https://doi.org/10.1109/TIP.2012.2183143.

    Article  PubMed  Google Scholar 

  22. Huang J, Yang F. Compressed magnetic resonance imaging based on wavelet sparsity and nonlocal total variation. IEEE Int Symp Biomed Imaging. 2012. https://doi.org/10.1109/ISBI.2012.6235718.

    Article  Google Scholar 

  23. Murphy M, Alley M, Demmel J, et al. Fast ℓ1-SPIRiT compressed sensing parallel imaging MRI: scalable parallel implementation and clinically feasible runtime. IEEE Trans Med Imaging. 2012;31(6):1250–62. https://doi.org/10.1109/tmi.2012.2188039.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Qu X, Hou Y, Lam F, et al. Magnetic resonance image reconstruction from undersampled measurements using a patch-based nonlocal operator. Med Image Anal. 2014;18(6):843–56. https://doi.org/10.1016/j.media.2013.09.007.

    Article  PubMed  Google Scholar 

  25. Ji J, Lang T. Dynamic MRI with compressed sensing imaging using temporal correlations. IEEE Int Symp Biomed Imaging. 2008. https://doi.org/10.1109/ISBI.2008.4541321.

    Article  Google Scholar 

  26. Usman M, Prieto C, Schaeffter T, et al. k-t Group sparse: a method for accelerating dynamic MRI. Magn Reson Med. 2011;66(4):1163–76. https://doi.org/10.1002/mrm.22883.

    Article  CAS  PubMed  Google Scholar 

  27. Prieto C, Usman M, Wild JM, et al. Group sparse reconstruction using intensity-based clustering. Magn Reson Med. 2013;69(4):1169–79. https://doi.org/10.1002/mrm.24333.

    Article  CAS  PubMed  Google Scholar 

  28. Lingala SG, Jacob M. Blind compressive sensing dynamic MRI. IEEE Tran Med Imaging. 2013;32(6):1132–45. https://doi.org/10.1109/TMI.2013.2255133.

    Article  Google Scholar 

  29. Majumdar A. Improving synthesis and analysis prior blind compressed sensing with low-rank constraints for dynamic MRI reconstruction. Magn Reson Imaging. 2015;33(1):174–9. https://doi.org/10.1016/j.mri.2014.08.031.

    Article  PubMed  Google Scholar 

  30. Lingala SG, Hu Y, DiBella E, et al. Accelerated dynamic MRI exploiting sparsity and low-rank structure: k-t SLR. IEEE Trans Med Imaging. 2011;30(5):1042–54. https://doi.org/10.1109/tmi.2010.2100850.

    Article  PubMed  PubMed Central  Google Scholar 

  31. Lingala SG, DiBella E, Adluru G, et al. Accelerating free breathing myocardial perfusion MRI using multi coil radial k-t SLR. Phys Med Biol. 2013;58(20):7309–27. https://doi.org/10.1088/0031-9155/58/20/7309.

    Article  PubMed  PubMed Central  Google Scholar 

  32. Petrov AY, Herbst M, Stenger VA. Improving temporal resolution in fMRI using a 3D spiral acquisition and low rank plus sparse (L+S) reconstruction. Neuroimage. 2017;157:660–74. https://doi.org/10.1016/j.neuroimage.2017.06.004.

    Article  PubMed  Google Scholar 

  33. Sun A, Zhao B, Li Y, et al. Real-time phase-contrast flow cardiovascular magnetic resonance with low-rank modeling and parallel imaging. J Cardiovasc Magn Reson. 2017;19:19. https://doi.org/10.1186/s12968-017-0330-1.

    Article  PubMed  PubMed Central  Google Scholar 

  34. Majumdar A. Improved dynamic MRI reconstruction by exploiting sparsity and rank-deficiency. Magn Reson Imaging. 2013;31(5):789–95. https://doi.org/10.1016/j.mri.2012.10.026.

    Article  PubMed  Google Scholar 

  35. Majumdar A, Ward RK, Aboulnasr T. Non-convex algorithm for sparse and low-rank recovery: application to dynamic MRI reconstruction. Magn Reson Imaging. 2013;31(3):448–55. https://doi.org/10.1016/j.mri.2012.08.011.

    Article  PubMed  Google Scholar 

  36. Lathauwer LD, Moor BD, Vandewalle J. A multilinear singular value decomposition. SIAM J Matrix Anal Appl. 2000;21:1253–78. https://doi.org/10.1137/s0895479896305696.

    Article  Google Scholar 

  37. Zuo XN, Yu Y, Jin J, et al. Multidimensional compressed sensing MRI using tensor decomposition-based sparsifying transform. PLoS One 2014, 9, doi:https://doi.org/10.1371/journal.pone.0098441.

  38. Beck A, Teboulle M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J Imaging Sci. 2009;2:183–202. https://doi.org/10.1137/080716542.

    Article  Google Scholar 

  39. Jiang M, Jin J, Liu F, et al. Sparsity-constrained SENSE reconstruction: An efficient implementation using a fast composite splitting algorithm. Magn Reson Imaging. 2013;31(7):1218–27. https://doi.org/10.1016/j.mri.2012.12.003.

    Article  PubMed  Google Scholar 

  40. Candès EJ, Recht B. Exact matrix completion via convex optimization. Found Comput Math. 2009;9:717–72. https://doi.org/10.1007/s10208-009-9045-5.

    Article  Google Scholar 

  41. Jung H, Ye JC, Kim EY. Improved k-t BLAST and k-t SENSE using FOCUSS. Phys Med Biol. 2007;52(11):3201–26. https://doi.org/10.1088/0031-9155/52/11/018.

    Article  PubMed  Google Scholar 

  42. Jung H, Sung K, Nayak KS, et al. k-t FOCUSS: A general compressed sensing framework for high resolution dynamic MRI. Magn Reson Med. 2009;61(1):103–16. https://doi.org/10.1002/mrm.21757.

    Article  PubMed  Google Scholar 

  43. Nam S, Akçakaya M, Basha T, et al. Compressed sensing reconstruction for whole-heart imaging with 3D radial trajectories: a graphics processing unit implementation. Magn Reson Med. 2013;69(1):91–102. https://doi.org/10.1002/mrm.24234.

    Article  PubMed  Google Scholar 

  44. Tanner J, Vary S. Compressed sensing of low-rank plus sparse matrices. arXiv preprint arXiv:2007.09457, 2020.

  45. 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. https://doi.org/10.1109/TIP.2003.819861.

    Article  PubMed  Google Scholar 

  46. Feng L, Srichai MB, Lim RP, et al. Highly accelerated real-time cardiac cine MRI using k-t SPARSE-SENSE. Magn Reson Med. 2013;70(1):64–74. https://doi.org/10.1002/mrm.24440.

    Article  PubMed  Google Scholar 

  47. Paul J, Divkovic E, Wundrak S. High-resolution respiratory self-gated golden angle cardiac MRI: comparison of self-gating methods in combination with k-t SPARSE SENSE. Magn Reson Med. 2015;73(1):292–8. https://doi.org/10.1002/mrm.25102.

    Article  PubMed  Google Scholar 

Download references

Acknowledgements

The authors would like to thank S.G. Lingala for sharing the k-t SLR matlab code, and thanks Yu Y for providing the HOSVD sparsity basis code.

Funding

This work is supported in part by the National Natural Science Foundation of China (61672466, 62011530130), in part by Joint Fund of Zhejiang Provincial Natural Science Foundation (LSZ19F010001), in part by the Key Research and Development Program of Zhejiang Province (2020C03060, 2020C03016, 2022C03111).

Author information

Authors and Affiliations

Authors

Contributions

JZ & LH designed the k-t TGV-TD reconstruction method. JS, ZW, WX, LX and MJ contributed to the algorithm development and data analysis. All authors have been involved in drafting and revising the manuscript and approved the final version to be published. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Mingfeng Jiang.

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 licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. 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 in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Zhang, J., Han, L., Sun, J. et al. Compressed sensing based dynamic MR image reconstruction by using 3D-total generalized variation and tensor decomposition: k-t TGV-TD. BMC Med Imaging 22, 101 (2022). https://doi.org/10.1186/s12880-022-00826-1

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12880-022-00826-1

Keywords