Introduction

Despite recent advances in compressed sensing (CS) and deep learning (DL) methods to accelerate MR imaging, there are still limitations for wide clinical adoption in cardiac MR imaging; therefore, parallel imaging remains the preferred image acceleration technique. CS reconstruction time remains long, even with the state-of-art hardware system, is only available for specific sequences (e.g., cardiac cine), and often uses spatial-temporal redundancy resulting in considerable temporal blurring. In addition, other than denoising algorithms, the majority of prior DL techniques have been tested offline using retrospective data without rigorous prospective evaluation. Furthermore, the generalizability of DL models for different sequences, slice orientation, and ease of integration into the standard clinical system remains challenging. In this study, we developed and evaluated an image reconstruction framework to accelerate cardiac MR imaging through a cascade of parallel imaging and generative adversarial neural network (PI-GAN) that combines widely available parallel imaging with GAN for the reconstruction of highly accelerated cardiac MR images. Images collected using late gadolinium enhancement (LGE) were used to train the PI-GAN model, which was later used to reconstruct cine images. Subsequently, we deployed the reconstruction model inline that allows the reconstruction of highly accelerated cardiac MR imaging on the scanner.

Method

Figure 1.a shows schematic of the PI-GAN model. The fully-sampled k-space data are undersampled by collecting R_GAN of k_y lines, followed by R_PI from undersampled k-space data divided into two regions: a) high-frequency and b) low-frequency. The resulting k-space data is first reconstructed using a matrix size, corresponding to a low-resolution parallel imaging image using conventional parallel imaging. In this study, we used a conventional GRAPPA acceleration factor of 2. The out of the initial PI reconstruction is de-aliased images with low spatial resolution. Subsequently, images reconstructed from GRAPPA are fed into a GAN reconstruction model to reconstruct the images into a larger matrix size. The architecture of the GAN consists of the following neural networks (Figure 1.b). 1) Generative Network has 23 dense residual blocks and 3 convolutional layers. This network removes artifacts and reconstructs high-resolution images from low-resolution images. 2) Discriminative network has 10 convolutional layers to train the generative network by discriminating whether the input of this network is the generated or original fully-sampled reconstructed. Furthermore, 3) Pre-trained VGG network extracts features and enables to compare the image at the feature level. Networks are trained using the following loss functions: 1) Pixel loss uses mean squared error of images. 2) Adversarial loss uses relative average GAN loss between predicted label distributions. 3) VGG feature based on L1-loss measures the difference between the VGG features. Furthermore, 4) Spatial frequency loss is L1-loss between the spatial frequencies of images. We set the weights of each cost function as adversarial loss = 0.005, pixel loss and spatial frequency loss = 0.01, and VGG loss = 1.0. The hyperparameters were batch size 36 and training patch size 48 by 48. We trained the neural networks for 210 epochs using Adam optimizer with learning rate= 0.0001, beta1=0.9, and beta2=0.999. We implemented the networks with Python and PyTorch library and trained on a DGX-1 workstation (NVIDIA, Santa Clara, California). The network was trained using 3D LGE images from 363 patients. Data were acquired at 3T (MAGNETOM Vida, Siemens Healthcare, Erlangen, Germany), and the k-space data were extracted from the vendor reconstruction pipeline. The datasets were divided into training (291 patients) and validation (72 patients). 3D LGE images were collected using R_PI = 2. To simulate the R_GAN rate 2-4 (i.e., overall acceleration between 4 and 8), we retrospectively discarded the phase-encoding k-space lines by selecting only 25% to 50% of the center of k-space lines. Afterwards, we generated training datasets consisting of pairs of GRAPPA-reconstructed images with low and high-resolution LGE images (Figure 2). We also integrated the GAN network inline using a Siemens prototype Framework for Image Reconstruction (FIRE)¹. Low-resolution images were reconstructed in a Siemens Image Reconstruction Environment (ICE) and fed to PI-GAN using Fire in ISMRMRD format². The reconstructed image is sent back to the ICE pipeline to correct distortion and generate a DICOM image. For evaluation, we prospectively acquired ECG-segmented cine bSSFP images from 29 subjects with GRAPPA rate 2 and PI-GAN accelerated segmented cine images and R_GAN=4 (total acceleration of 8, 4 sec per slice). For DL-accelerated acquisition, we use the trained GAN for reconstruction. We quantified LVEF and LV mass from the cine images and compared the conventional and the proposed methods.

Results

PI-GAN model, trained using LGE images, successfully reconstructed highly accelerated cine images in any slice orientation (Figure 3). Bland-Altman graph (Figure 4) shows excellent agreement between various structural and functional parameters measured using standard ECG-segmented cine images vs. PI-GAN accelerated cine images. Figure 5 shows inline reconstructed images on the scanner. The total reconstruction time for PI-GAN on the scanner was 133 seconds for 11 slices.

Conclusion

PI-GAN can reconstruct highly accelerated cine images with different imaging contrast and orientation as the ones in the training set. An inline implementation allows seamless integration of the proposed acceleration approach into clinical workflow. Further studies are warranted to evaluate the performance of PI-GAN in other cardiac MR sequences.

Acknowledgements

No acknowledgement found.