4661
Non-Cartesian k-space Deep Learning for Accelerated MRI1Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea, Republic of
Synopsis
The annihilating filter-based low-rank Hankel matrix approach (ALOHA) [1] is one of the most recent compressed sensing (CS) approaches that directly interpolates the missing k-space data using low-rank Hankel matrix completion. Inspired by the recent low-rank Hankel matrix decomposition using data-driven framelet basis [2], we propose a completely data-driven deep learning algorithm for k-space interpolation. In particular, our method can be applied directly by simply adding an additional re-gridding layer to non-Cartesian k-space trajectories such as radial trajectories.
Introduction
Recently, inspired by the tremendous success of deep learning in low-level computer vision problems, many researchers have investigated deep learning approaches for various biomedical image reconstruction problems and successfully demonstrated significant performance gain.
In particular, Wang et al [1] used the deep learning reconstruction either as an initialization or a regularization term. Deep network architecture using unfolded iterative compressed sensing (CS) algorithm was also proposed. The authors [2] tried to learn a set of regularizers under a variational framework. Multilayer perceptron [3] was introduced in for accelerated parallel MRI. An extreme form of the neural network called AUtomated TransfOrm by Manifold APproximation (AUTOMAP) [4] even attempts to estimate the Fourier transform itself using fully connected layers. All these pioneering works have consistently demonstrated superior reconstruction performances over the CS at significantly lower run-time computational complexity.
However, although the end-to-end recovery approach like AUTOMAP may directly recover the image without ever interpolating the missing k-space samples, it works only for the sufficiently small size images due to its huge memory requirement for fully connected layers. Accordingly, most of the popular deep learning MR reconstruction algorithms are either in the form of image domain post-processing, or iterative updates between the k-space and the image-domain using a cascaded network.
Therefore, The proposed deep learning approach directly interpolates the missing k-space data so that accurate reconstruction can be obtained by simply taking the Fourier transform of the interpolated k-space data. Specifically, the recent theory of deep convolutional framelets [5] showed that an encoder-decoder network emerges from the data-driven low-rank Hankel matrix decomposition [6], whose rank structure is controlled by the number of filter channels. This discovery gives us important clues to develop a successful deep learning approach for k-space interpolation. We further show that our deep learning approach for Non-Cartesian k-space interpolation such as radial, spiral. Moreover, all the network are implemented in the form of convolutional neural network (CNN) without requiring fully connected layer, so the GPU memory requirement is minimal.
Methods
As shown in Fig. 1, most of the existing deep learning approaches were applied directly to an image domain, whereas our deep neural network is directly applied to k-space domain. Furthermore, to handle non-Cartesian k-space measurements, the proposed k-space deep learning added the re-gridding layer that is placed before the weighting or skipped connection layer, as shown in Fig. 2. Also, as an input of the proposed network, the Fourier domain data was converted into two-channel real values, and each channel is a real part and an imaginary part. The network backbone was used in U-Net, which consists of convolution, batch normalization, rectified linear unit (ReLU), and skipped-connection with concatenation as shown in Fig. 1.Network training
The k-space dataset was synthesized from Human Connectome Project (HCP) MR dataset (https://db.humanconnectome.org). The radial and spiral k-space data were generated using MRI simulator (http://bigwww.epfl.ch/algorithms/mri-reconstruction). The proposed network was optimized by stochastic gradient descent (SGD) method, and regularization parameter was λ = 10-4. The number of epochs was 300, and the batch size was 4. The initial learning rate was 10-5 and gradually dropped to 10-6Results
Fig. 3(a) shows the reconstruction images from x6 accelerated radial sampling patterns using the trajectory of the green box in Fig. 3(a). The 2th column shows the downsampled image, and 3rd and last columns show the image-domain CNN and the proposed k-space deep learning results, respectively. The proposed k-space deep learning provides realistic image quality and preserves the detailed structures as well as the textures. However, the image-domain CNN does not preserve the textures and structures. In addition, the proposed k-space deep learning provides smallest NMSE values. Fig. 3(b) shows the reconstruction images from the x4 spiral trajectory of the green box in Fig. 3(b). As in the radial case, the proposed k-space deep learning outperforms the image-domain.Discussion
To improve the network’s performance using ALOHA [6], a signal must be sparse. However, in general, the signal is not satisfied with sparseness. To meet the requirement, ALOHA introduces the k-space weighting kernel . As our k-space deep learning is inspired by ALOHA, the concept of weights is easily implemented using a weighting and un-weighting layer as shown in Fig. 2(a). In addition, sparsifying the signal can be implemented by the skipped connection layer as shown in Fig. 2(b).Conclusion
We propose a novel k-space deep learning method for reconstructing the non-Cartesian trajectories. Although the image-domain CNN method works only in the image-domain, the proposed k-space deep learning directly interpolates k-space in the under-sampled measurements. The proposed method is superior to the image-domain CNN.Acknowledgements
This work is supported by Korea Science and Engineering Foundation, Grant number NRF2016R1A2B3008104.References
[1] Wang, Shanshan, et al. "Accelerating magnetic resonance imaging via deep learning." Biomedical Imaging (ISBI), 2016 IEEE 13th International Symposium on. IEEE, 2016.
[2] Hammernik, Kerstin, et al. "Learning a variational network for reconstruction of accelerated MRI data." Magnetic resonance in medicine 79.6 (2018): 3055-3071.
[3] Kwon, Kinam, Dongchan Kim, and HyunWook Park. "A parallel MR imaging method using multilayer perceptron." Medical physics 44.12 (2017): 6209-6224.
[4] Zhu, Bo, et al. "Image reconstruction by domain-transform manifold learning." Nature 555.7697 (2018): 487.APA
[5] Ye, Jong Chul, Yoseob Han, and Eunju Cha. "Deep convolutional framelets: A general deep learning framework for inverse problems." SIAM Journal on Imaging Sciences 11.2 (2018): 991-1048.
[6] Jin, Kyong Hwan, Dongwook Lee, and Jong Chul Ye. "A general framework for compressed sensing and parallel MRI using annihilating filter based low-rank Hankel matrix." IEEE Transactions on Computational Imaging 2.4 (2016): 480-495.
Figures
