SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Christopher Sandino, MSc
Ph.D. Student
Stanford University
Peng Lai, PhD
Scientist
GE Healthcare
Martin Janich, PhD
Manager
GE Healthcare
Anja Brau, PhD
General Manager
GE Healthcare
Shreyas Vasanawala, MD
Professor
Stanford University
Joseph Cheng, PhD
Instructor
Stanford University
Background: Neural networks can leverage previous exam data to learn better priors for iterative reconstruction of highly accelerated dynamic MRI data [1,2]. Information about the data acquisition model (i.e. Fourier transform, and coil sensitivities) has been incorporated to further improve reconstruction accuracy and robustness to hallucinations [3,4]. Here, we extend previous work on dynamic MRI [1] by incorporating ESPIRiT-based coil sensitivity information [5] into the unrolled network [6]. Our method is compared against conventional compressed sensing on 10x accelerated 2D cardiac CINE data and evaluated with respect to common image quality metrics.
Methods: The proposed unrolled network architecture [6] iterates between data consistency and ResNet [7] blocks (Fig. 1). Data consistency layers incorporate multiple sets of coil sensitivity maps computed using ESPIRiT to project back and forth between k-space and image domains. Two sets of ESPIRiT sensitivity maps are used for improved robustness to acquisition model errors [5]. Two coil-combined images, each corresponding to one set of maps, are stacked together and jointly de-aliased using networks with 3D convolutions to exploit spatiotemporal correlations. The neural network is trained end-to-end on the average l1 loss [8] between output and ground truth image sets.
With IRB approval, fully sampled bSSFP 2D cardiac CINE datasets were acquired from 12 volunteers at different cardiac views and slice locations on 1.5T and 3.0T GE scanners. For training, ten volunteer datasets are split slice-by-slice to create 155 unique examples, which are further augmented by random flipping, cropping along readout, reducing phase FOV (0-15%) to simulate anatomy overlap, partial echo (20-30%), and variable-density undersampling. For evaluation, the remaining two volunteer datasets are retrospectively undersampled to simulate 10-fold acceleration with 25% partial echo readout. The same evaluation datasets are reconstructed slice-by-slice using the proposed method and compressed sensing with spatial Wavelet and temporal total variation constraints (l1-ESPIRiT using two sets of maps). Reconstruction quality is evaluated with respect to peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).
Results: The proposed method outperforms l1-ESPIRiT with respect to both PSNR and SSIM metrics (Figs 2 & 3). The improvement of the proposed method over l1-ESPIRiT is most apparent inside the heart with less spatiotemporal blurring of myocardium and papillary muscles in deep reconstructions (Fig. 2). ESPIRiT with deep priors more effectively suppresses residual aliasing arising from acquisition model errors, such as those caused by gradient non-linearity (Fig. 3).
Conclusion: Preliminary results suggest that ESPIRiT with deep priors can reconstruct 2D cardiac CINE data more accurately and robustly than compressed sensing.