Synopsis

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Ferumoxytol can support high quality SGE cardiac cine with blood-myocardial CNR similar to SSFP cine, but free of off-resonance artifact. Much work is ongoing to accelerate the acquisition of multi-slice, breath held cardiac cine. Deep learning-based reconstruction methods can accelerate the image acquisition and reconstruction but needs a large amount of data to train. When compared with compressed sensing and low rank reconstructions, our network showed sharper images and higher consistency with the reference and significantly better quantitative evaluation metrics. We showed that a network trained with non-contrast images could generalize to accelerated ferumoxytol-enhanced cardiac cine MRI with 10x acceleration.

INTRODUCTION

Ferumoxytol has increasing applications for cardiovascular MRI in patients with renal impairment and/or congenital heart disease (CHD). Conventionally, multi-slice 2D cine images are collected using ECG-gating during multiple breath-holds, which results in a long acquisition time and possible slice misalignments.¹ Parallel imaging (PI) and compressed sensing (CS) reconstruction have been proposed to accelerate image acquisition. However, PI has a limited acceleration rate and CS requires extensive computational resources and a long reconstruction time.^2-4 The recent development of deep learning enables more accurate modeling of the image priors and faster reconstruction compared to CS reconstructions.^5,6 However, the generalizability of the trained network on contrast-enhanced images has not been investigated. In this work, we aimed to investigate the performance of a neural network trained with non-contrast images on ferumoxytol-enhanced cardiac cine MRI.

METHODS

Variable splitting
CS reconstruction can be formulated as minimizing the optimization problem:
$$\underset{m}\min\{||Am-y||^2_2 + \lambda R(m)\}$$
where m is the complex-valued MR image, y is the undersampled k-space and A represents coil sensitivity, Fourier transform and undersampling operations. The first term is the data fidelity term, and the second term is the regularization term. Inspired by the fast convergence of the variable splitting method, the optimization problem is transferred to a denoising problem, a data consistency problem and a weighted sum problem.⁷

Network architecture and training
An unrolled spatiotemporal network was developed to integrate the iterative process. As shown in Figure 1, the variable-splitting network has multiple stages, each consisting of a denoiser, a data consistency block and a weighted sum block. The denoiser is learned by a 3D spatiotemporal CNN. The inputs to the network are undersampled 2D cine images, undersampled k-space, sensitivity maps and undersampling masks. The network utilizes mean-squared-error (MSE) loss to minimize the difference between the output and the reference.

The network was trained using 10x undersampled non-contrast cardiac cine MRI and tested using ferumoxytol-enhanced cardiac cine MRI. The network was trained for 26 epochs using an Adam optimizer with momentum β=0.9, mini-batch size of 1, and an initial learning rate of 0.0001. Weights for the network were initiated with random normal distributions with a variance of σ = 0.01 and mean µ=0. The training was completed in Python v3.7 on a graphics processing unit (NVIDIA Tesla V100, 32GB).

Data acquisition and preparation
The study was HIPAA compliant and approved by the institutional review board. To train and evaluate the network, non-contrast 2D SSFP cine images with 30 cardiac phases (2x GRAPPA) were acquired with ECG retrogating in 42 patients on a 1.5T scanner (MAGNETOM Avanto^fit, Siemens Healthcare, Erlangen, Germany), totally 531 images. The images were divided into 424 training data and 107 validation data. Ferumoxytol-enhanced testing data were acquired using 2D ECG-retrogated GRE sequences from 6 CHD adult patients on the 1.5T scanner (30 phases and 2x GRAPPA) and 6 patients and 5 healthy subjects on a 3T scanner (MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany, 25 phases and 3x GRAPPA), totally 192 images.

A research application tool written in C++ using the image reconstruction framework of the MRI scanner was used to reconstruct the aliasing-free multi-channel complex images and to extract the sensitivity maps for each image. The images were retrospectively undersampled using variable-density sampling across phase-encoding and cardiac phase dimensions. An example undersampling mask is shown in Figure 2.

Evaluation
The proposed network was evaluated on retrospectively undersampled images and compared with CS reconstruction with spatial l1-wavelet and temporal total variation (TV) regularizations⁸ and low rank reconstruction⁹, both of which were completed using the Bart toolbox¹⁰. To evaluate the image quality of different reconstructions, normalized MSE (NMSE), structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) were computed for all test subjects. Paired t-tests were conducted to determine the statistical significance between the network and other methods. Cardiac function analysis including left ventricular end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV) and ejection fraction (EF) were calculated for 5 CHD patients using cvi42 (Circle Cardiovascular Imaging, Calgary, AB, Canada). The EDV, ESV, SV and EF measured from different methods were compared to those measured from the reference using Bland-Altman analysis with mean difference (MD) and limits of agreement (LoA=MD±1.96×SD).

RESULTS

Figure 3 shows the network, CS and low rank reconstructions of a patient with 10x acceleration. All reconstruction methods removed ghosting artifacts and the network showed sharper images and higher consistency compared to CS and low rank reconstructions. NMSE, SSIM and PSNR evaluation results are shown in Figure 4. Paired t-tests show that the proposed network had significantly lower NMSE (p<0.0001), significantly higher SSIM (p<0.0001) and PSNR (p<0.0001) than the other reconstructions. Figure 5 shows the Bland-Altman plots of EDV, ESV, SV and EF from all reconstruction methods. The network produced smaller ranges of LoA compared to the other reconstructions.

DISCUSSION AND CONCLUSION

In this study, we developed a spatiotemporal deep neural network to reconstruct ferumoxytol-enhanced accelerated cardiac cine images using the variable splitting method. Results in this study show that even not trained with contrast-enhanced images, the network trained with non-contrast images can be applied to reconstruct artifact-free ferumoxytol-enhanced cardiac cine images using 10x accelerated acquisition. This network may allow more efficient ferumoxytol-enhanced cardiac cine imaging.

Acknowledgements

No acknowledgement found.