Introduction

3D whole‐heart coronary MR angiography (CMRA) has shown significant potential for the diagnosis of coronary artery disease (CAD). The acquisition of 3D whole-heart CMRA remains lengthy, as high-resolution data needs to be acquired for quantification of luminal stenosis. Scan time reduction has been previously achieved by combining undersampling with respiratory non-rigid motion correction enabling free-breathing 3D isotropic CMRA within ~10min scan time^1,2. Recently, deep learning-based undersampling reconstruction methods have been proposed to shorten reconstruction time and enabling high-resolution scans in ~7min predictable scan time³. However, spatial resolution is still limited compared to coronary CT angiography (~0.6mm³) and scan time remains relatively long.
An alternative approach for accelerating the image acquisition while simultaneously increasing spatial resolution exploits deep learning-based super-resolution⁴. Images are acquired in low-resolution (with or without imaging acceleration) and retrospectively reconstructed to the high-resolution target in a single image super-resolution setting. Different deep learning strategies have been employed to utilize spatio-temporal information sharing on a patch- or image-level^5-8.
In this work, we propose a deep learning-based super-resolution framework, combined with non-rigid respiratory motion correction (SR-CMRA), to further shorten the acquisition time. Furthermore, in contrast to previous methods, we propose a deep-learning generative adversarial super-resolution network that generalizes to different input resolutions and can operate on an image or patch-level enabling flexible integration into either image or patch-based reconstruction methods. In this study, we analyse the proposed method for a 16-fold increase in spatial resolution by reconstructing a high-resolution (HR) CMRA (0.9mm³ or 1.2mm³) dataset from low-resolution (LR) acquisitions (1.2x3.6x3.6mm³ or 1.2x4.8x4.8mm³; superior-inferior x left-right x anterior-posterior). The supervised SR-CMRA was trained in a cohort of 47 CMRA patients with suspected CAD whereas testing was performed in retrospectively downsampled images (LR: 1.2x3.6x3.6mm³ and 1.2x4.8x4.8mm³) of 5 patients to quantify performance and in 5 prospectively acquired patients with a LR CMRA acquisition (LR: 1.2x4.8x4.8mm³) in less than one minute scan time.

Methods

The proposed SR framework is depicted in Fig.1. ECG-triggered 3D whole-heart Cartesian bSSFP CMRA was acquired under free-breathing in coronal orientation in patients with suspected CAD using a prototype sequence at 1.5T (Aera, Siemens Healthcare, Erlangen Germany). Imaging parameters are stated in². Briefly, data was acquired with a variable-density CASPR sampling in ~7min (2.5x acceleration) for an isotropic resolution of 1.2mm³ and in ~8min (5x acceleration) for a resolution of 0.9mm³. 2D image navigators were used for retrospective beat-to-beat translational respiratory motion correction and respiratory binning of the 3D CMRA data. Images were reconstructed with a non-rigid motion-compensated PROST² serving then as HR target. LR images were acquired with matching parameters in ~50s (1x fully sampled) for a resolution of 1.2x4.8x4.8mm³.
We propose a generative adversarial network which consists of two cascaded Enhanced Deep Residual Network for SR⁹ generator, a trainable discriminator¹⁰ and a pre-trained perceptual loss network (VGG-16¹¹). The EDSR is built up of two stages each performing a 2-fold upsampling in each direction in 8 consecutive residual blocks with 3x3 convolution filters of stride 1 and 64 kernels. The patch discriminator network is built as a convolutional neural network with dyadic kernel increase and alternating stride of 1 and 2. The input to the network is the LR-CMRA (1.2x3.6x3.6mm³ or 1.2x4.8x4.8mm³) whereas the output is the corresponding isotropic SR-CMRA image (0.9mm³ or 1.2mm³). The input can be either an axial 2D patch with a size in the range of 10% - 90% of the full image slice or an axial 2D image slice (i.e. a single 2D patch of 100% size). The network is trained in a supervised manner on 460,000 axial pairs of HR-CMRA (0.9mm³ and 1.2mm³) and retrospectively downsampled LR images from 47 patients. Retrospectively simulating a 25% phase and slice resolution provides the LR training input (readout resolution was not affected). The perceptual loss network and the discriminator were pre-trained on a cohort of 50 patients with suspected CAD scanned by CT coronary angiography. In total, the network consists of ~6.3 million trainable parameters optimized under mean absolute error (MAE), structural similarity index (SSIM), adversarial and perceptual loss by an Adam optimizer (batch size=16, 60 epochs). Testing was done on 5 patients (retrospective cohort; not seen in training) with retrospectively downsampled LR images from HR target (0.9mm³ or 1.2mm³) and in 5 patients (prospective cohort; not seen in training) with prospectively acquired LR-CMRA of 1.2x4.8x4.8mm³ and with a separate HR acquisition (1.2mm³).

Results and Discussion

The LR input, HR acquisition and SR-CMRA output of one prospectively acquired test patient and one retrospectively downsampled test patient are shown in Fig. 2 and 3, respectively. The proposed network generalized for varying imaging resolutions and for patch- or image-input as shown in Fig. 4. Qualitatively and quantitatively the proposed SR-CMRA showed significant improvement in edge sharpness and vessel delineation as well as depiction of coronary arteries with respect to LR-CMRA and comparable image quality to the HR-CMRA (Fig. 5). The proposed SR-CMRA required ~186hrs for training but only ~3s for reconstruction.

Conclusion

In conclusion, the proposed SR-CMRA has the potential to provide a 16-fold increase in resolution with comparable image quality to the HR-CMRA image while reducing the predictable scan time to <1min.

Acknowledgements

This work was supported by EPSRC (EP/P001009, EP/P032311/1, EP/P007619/1), BHF programme grant RG/20/1/34802 and Wellcome EPSRC Centre for Medical Engineering (NS/ A000049/1).

References

1. Cruz et al. MRM 2017;77(5).
2. Bustin et al. JCMR 2020;22(1).
3. Fuin et al. Magnetic Resonance Imaging 2020;70.
4. Lundervold et al. Zeitschrift für Medizinische Physik 2019;29(2).
5. Xie et al. ISMRM 2018. p1050.
6. Lin et al. arXiv preprint arXiv:200510626 2020.
7. Zhao et al. Comput Med Imaging Graph 2020;80.
8. Steeden et al. Journal of Cardiovascular Magnetic Resonance 2020;22(1).
9. Lim et al. CVPR 2017. p136-44.
10. Ledig et al. CVPR 2017. p4681-90.
11. Russakovsky et al. IJCV 2015;115(3).