1970

DL2 - Deep Learning + Dictionary Learning-based Regularization for Accelerated 2D Dynamic Cardiac MR Image Reconstruction
Andreas Kofler1, Tobias Schaeffter1,2,3, and Christoph Kolbitsch1,2
1Physikalisch-Technische Bundesanstalt, Berlin and Braunschweig, Berlin, Germany, 2School of Imaging Sciences and Biomedical Engineering, King's College London, London, United Kingdom, 3Department of Biomedical Engineering, Technical University of Berlin, Berlin, Germany

Synopsis

In this work, we combine Convolutional Neural Networks (CNN)- with Dictionary Learning (DL)- and Sparse Coding (SC)-based regularization for dynamic cardiac MR image reconstruction. The regularization on the image is imposed by patch-wise sparsity with respect to a learned overcomplete dictionary and closeness to a CNN-based image-prior which is obtained from a pre-trained CNN. We compare the proposed method to two iterative methods which incorporate the different components separately. We demonstrate the combination of CNNs with DL and SC leads to improved image quality and faster convergence compared to DL+SC only.

Introduction

In 2D Dynamic Cardiac Cine MR, the data-acquisition process is usually performed during a single breathhold of the patient in order to reduce respiration-related motion artefacts. Undersampling the $$$k$$$-space accelerates this process but leads to an ill-posed reconstruction problem. Thus, regularization methods must be used to properly reconstruct images with diagnostic quality. Dictionary Learning (DL) and Sparse Coding (SC) are well-known and established methods based on local model-assumptions, i.e. patch-wise sparsity. While DL and SC have been successfully applied for MR image reconstruction1,2 and do not require ground-truth data for training, the SC of the image-estimates during the reconstruction makes the reconstruction inherently slow. Convolutional Neural networks (CNNs), have been recently widely applied for MR image reconstruction as learned regularizers3,4. They provide a global model with which an image can be processed very quickly. On the other hand, they require ground-truth data for training and can typically be applied only once since they correspond to learned mappings between two fixed manifolds.
Here, we combine DL and SC with CNNs in order to accelerate the data-acquisition process as well as the reconstruction algorithm.

Methods

We formulate the reconstruction problem as a joint optimization problem over the image, an overcomplete dictionary $$$\mathbf{\Psi}$$$ and a set of sparse codes $$$\{\boldsymbol{\gamma}_j \}_j$$$ by further including a penalty term which imposes the solution to be close to a CNN-based image prior which is previously obtained from a pre-trained CNN $$$f_{\Theta}$$$ with trainable parameters $$$\Theta$$$, i.e.
$$ \min_{\mathbf{x},\mathbf{\Psi},\{\boldsymbol{\gamma}_j \}_j} \frac{1}{2}\|\mathbf{A}\mathbf{x} - \mathbf{y} \|_2^2 + \frac{\lambda_{\mathrm{CNN}}}{2}\|\mathbf{x} - \mathbf{x}_{\mathrm{CNN}} \|_2^2+ \frac{\lambda_{\mathrm{DL/SC}}}{2} \sum_j \big(\| \mathbf{R_j} \mathbf{x} -\mathbf{\Psi}\boldsymbol{\gamma}_j \|_2^2 + \|\boldsymbol{\gamma}_j \|_0\big),$$
where $$$\mathbf{A}$$$ denotes a non-uniform FFT (NUFFT) encoding operator, $$$\lambda_{\mathrm{CNN}}, \lambda_{\mathrm{DL/SC}}>0$$$ and $$$\mathbf{x}_{\mathrm{CNN}} := f_{\Theta}( \mathbf{x}_I)$$$ denotes the CNN-output given the initial reconstruction $$$ \mathbf{x}_I:= \mathbf{A}^{\dagger} \mathbf{y}$$$ where $$$\mathbf{A}^{\dagger}:=\mathbf{W}\mathbf{A}^H$$$ with $$$\mathbf{W}$$$ being a diagonal-operator containing the entries of the density-compensation function.

The network $$$f_{\Theta}$$$ is given by a data-consistent CNN, i.e. a CNN-block followed by a subsequent conjugate gradient (CG)-block which can be trained end-to-end. The CNN-block consists of a 2D spatio-temporal U-Net which reduces the undersampling artefacts in spatio-temporal domain by the XT,YT-method5. The CNN is used to generate $$$\mathbf{x}_{\mathrm{CNN}}$$$ from $$$\mathbf{x}_I$$$, which serves as a starting point for subsequent iterative reconstruction which alternates between SC, DL and an image-update stage via a CG method. As DL and SC algorithms, we use an adaptive version of the iterative thresholding and K residual means (aITKrM)6 and adaptive orthogonal matching pursuit (aOMP)7.
We evaluated the proposed reconstruction algorithm on 36 2D cine MR images of 4 patients which were acquired during a single breathhold of approximately 10s (TR/TE=3.0/1.5ms, FA 60$$$^{\circ}$$$) using a bSSFP sequence on a 1.5T MR scanner. The spatial dimensions were $$$N_x \times N_y=320×320$$$ with an in-plane resolution of 2 mm and a slice-thickness of 8 mm. The number of cardiac phases was $$$N_t=30$$$. The images were reconstructed using $$$kt$$$-SENSE8 by sampling the $$$k$$$-space data along $$$N_{\theta}=3400$$$ (i.e. $$$R=3$$$) radial spokes. Note that $$$R=3$$$ was needed to perform the scan during a single breathhold of the patients.
From these images, we retrospectively simulated a golden-angle radial data-acquisition9 with 12 receiver coils with an undersampling of approximately $$$R=18$$$ (i.e. $$$N_{\theta}=560)$$$ and $$$R=9$$$ (i.e. $$$N_{\theta}=560)$$$.

The CNN was pre-trained on a separate set of 144 2D images using the $$$L_2$$$-error as loss-function. We compared the proposed method to an iterative reconstruction method which uses the CNN-output as image-prior10 and a method using DL and SC7 in terms of peak signal-to-noise ratio (PSNR), normalized root-mean-squared error (NRMSE) and structural similarity index measure (SSIM).

Results

From Figure 1, we see that all three methods succesfully removed the undersampling artefacts from the initially reconstructed image. As can be seen in Table 1, the combination of CNN + DL led to the best results with respect to all measures. The strongest improvements were with respect to PSNR and NRMSE, while for SSIM, the impact was lower.
From Figure 2, we also see that the proposed reconstruction algorithm leads to a faster convergence of the image estimates compared to method only using DL and SC.
Figure 3 shows another example of a dynamic image where we can see the images as well as the point-wise error-images over the whole cardiac cycle.

Discussion & Conclusion

Combining CNNs and DL and SC led to an improvement in image quality compared to using each component separately. Using the CNN-estimate as a starting point for the iteration, faster convergence is achieved compared to DL+SC.Further, although the CNN-output already provides a "good" image estimate, the adaptive DL and SC further refine possibly lost image details as well as further reduce locally present image noise which was not removed by the CNN and thus the combination further improves the reconstruction results compared to the CNN-based approach.

Conclusion

The proposed reconstruction algorithm combines a global and a local model given by a pre-trained CNN and DL and SC, respectively. It yields higher image quality than using either CNNs or DL+SC separately and in addition, highly reduces reconstruction times required by DL+SC due to faster convergence given by the suitable starting point which is obtained with the CNN.

Acknowledgements

No acknowledgement found.

References

  1. Wang, Yanhua, and Leslie Ying. "Compressed sensing dynamic cardiac cine MRI using learned spatiotemporal dictionary." IEEE transactions on Biomedical Engineering 61.4 (2013): 1109-1120.
  2. Caballero, Jose, et al. "Dictionary learning and time sparsity for dynamic MR data reconstruction." IEEE transactions on medical imaging 33.4 (2014): 979-994.
  3. Hauptmann, Andreas, et al. "Real‐time cardiovascular MR with spatio‐temporal artifact suppression using deep learning–proof of concept in congenital heart disease." Magnetic resonance in medicine 81.2 (2019): 1143-1156.
  4. Hammernik, Kerstin, et al. "Learning a variational network for reconstruction of accelerated MRI data." Magnetic resonance in medicine 79.6 (2018): 3055-3071.
  5. Kofler, Andreas, et al. "Spatio-temporal deep learning-based undersampling artefact reduction for 2D radial cine MRI with limited training data." IEEE transactions on medical imaging 39.3 (2019): 703-717.
  6. Schnass, Karin. "Dictionary learning-from local towards global and adaptive." arXiv preprint arXiv:1804.07101 (2018).
  7. Pali, Marie‐Christine, et al. "Adaptive Sparsity Level and Dictionary Size Estimation for Image Reconstruction in Accelerated 2D Radial Cine MRI." Medical Physics (2020).
  8. Tsao, Jeffrey, et al.. "k‐t BLAST and k‐t SENSE: dynamic MRI with high frame rate exploiting spatiotemporal correlations." Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 50.5 (2003): 1031-1042.
  9. Winkelmann, Stefanie, et al. "An optimal radial profile order based on the Golden Ratio for time-resolved MRI." IEEE transactions on medical imaging 26.1 (2006): 68-76.
  10. Kofler, Andreas, et al. "Neural networks-based regularization for large-scale medical image reconstruction." Physics in Medicine & Biology 65.13 (2020): 135003.

Figures

Figure 1: Results and corresponding point-wise error-images which were obatined on the test set for $$$N_{\theta}=560$$$ (top row) and $$$N_{\theta}=1130$$$ (bottom row) radial spokes. From left to right: the initial NUFFT-reconstruction, the CNN-regularized solution, the solution obtained from the DL- and SC-regularized algorithm, the solution obtained with the combination of DL+SC and the CNN, the $$$kt$$$-SENSE reconstruction obtained from $$$N_{\theta}=3400$$$ radial lines which was used as ground-truth image for the retrospective undersampling.

Figure 2: Convergence results for the proposed method (red) compared to DL and SC (blue) for the two acceleration factors $$$R=18$$$ and $$$R=9$$$ given by $$$N_{\theta}=560$$$ and $$$N_{\theta}=1130$$$ radial spokes, respectively. The solid lines correspond to the mean value of the respective mesure obtained over the entire test set. The dashed lines correspond to the mean $$$\pm$$$ the standard deviation obtained over the test set.

Note that the values differ from the ones shown in the Table because here, no masks were used to restrict the calculations to a region of interest.


Table 1: Quantitative measures averaged over the test set. As can be seen, the proposed combination of the CNN- and DL+SC-based regularization achieves the best restuls with respect to PSNR, NRMSE and SSIM.

Figure 4: An example of results and corresponding point-wise error-images obtained for an acceleration factor of $$$R=18$$$. From left to right: The initial NUFFT-reconstruction obtained from $$$N_{\theta}=560$$$ radial spokes, the CNN-regularized solution10, the DL+SC-regularized solution7, the proposed method and the $$$kt$$$-SENSE reconstruction obtained from $$$N_{\theta}=3400$$$ radial spokes from which the $$$k$$$-space data was retrospectively simulated.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
1970