Oral Abstract Session
SCMR 22nd Annual Scientific Sessions
Non-contrast thoracic MRA (NC-MRA) based on T2-prepared b-SSFP readout [1, 2] is a promising method for evaluation of thoracic aortic aneurysms, because it does not require radiation or gadolinium-based contrast agent. We accelerated and simplified thoracic NC-MRA by using stack-of-stars k-space sampling with GRASP compressed sensing(CS) reconstruction  for to achieve 1.5x1.5x1.5mm3 spatial resolution at 6 min scan time . While the results from this study showed excellent agreement with contrast-enhanced MRA, the lengthy reconstruction (~4 hrs) remains an obstacle to clinical translation. The purpose of this study is to develop and test a convolutional neural network (CNN) to drastically reduce reconstruction time.
For more details on the image acquisition and CS reconstruction incorporating self-navigation of respiratory motion, please see reference . We performed NC-MRA in 22 patients (17 males, mean age=56.1±12.6 years) undergoing clinical MRI with FOV=288x288x120mm3, acquisition matrix=192x192x80, and scan time=350 heart beats. Each patient produced time resolved 3D volume (or 80 2D-time series) with 6 respiratory phases. We built a convolutional neural network (CNN) composed of 4 layers, with depth of 16, and 3x3x3(x-y-time) kernels and trained the network running on GPU Tesla P100 workstation equipped with tensorflow, by mapping zero-filled undersampled data (input) to CS reconstructed data (output) from first 10 randomly selected patients (Fig1A). Next, for validation, zero-filled undersampled data from the remaining 12 patients were reconstructed using the trained network and CS  (Fig1B). To evaluate data fidelity, we calculated DICE, SSIM, NRMSE, and apparent SNR. We also measured aortic diameters at seven standardized locations .
The mean image reconstruction time was 212 min and 1 min 14s for CS and CNN, respectively. Figure 2 shows representative maximum-intensity-projections derived from CS and CNN. As summarized in Table 1, data fidelity was high for all metrics. As shown in Figure 2, the mean difference and coefficient of repeatability (CR) were similarly small for all pairs: CS vs. CNN (mean = 3.4 cm; mean difference = -0.01 cm [0.3% relative to mean], CR= 0.26 cm [7.6% relative to mean]), CS vs CS (3.4 cm, 0.0 cm [0% relative to mean], 0.29 cm [8.5% relative to mean], respectively), and CNN vs. CNN (3.4 cm, -0.03 cm [0.9% relative to mean], 0.25 cm [7.4% relative to mean], respectively), indicating similar intra-observer agreement among different reconstruction methods.
Our proposed CNN achieves 171-fold faster image reconstruction compared with GRASP reconstruction for thoracic NC-MRA. The accelerated image reconstruction time of 1 min 14 s is amendable to clinical translation.