SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Cine Displacement Encoding with Stimulated Echoes (DENSE) provides accurate and reproducible imaging of myocardial strain . During DENSE data acquisition, an artifact-generating echo due to T1 relaxation is simultaneously acquired along with the displacement-encoded echo. Phase-cycled acquisitions are generally used to suppress the artifact-generating echo, which doubles the scan time and in some cases leads to imperfect artifact suppression . We developed and trained a deep Convolutional Neural Net (CNN) to suppress the T1-relaxation echo from a single acquisition (without phase cycling), potentially obviating the need to acquire phase-cycled data.
A Generative Adversarial Network  (GAN) (Figure 1) was employed (DAS-Net: DENSE Artifact Suppression Net) where a U-net  was embedded as the generator, trying to generate artifact-free images, and a ConvNet was embedded as the discriminator, evaluating the performance of the generator. Twenty-eight short-axis (SA) and long-axis cine DENSE datasets with phase cycling, each encoded for displacement in 2 directions, were acquired using 3T Siemens scanners. This dataset was divided into training (n=19), validation (n=4) and testing (n=5) data. The training set was augmented by a factor of 4 using image-domain spatial rotations and translations (Figure 2). The data were converted to kx-t/ky-t planes (n=10640) and pooled for training. Non-phase-cycled data were separated into real and imaginary parts and input into the CNN, and the network was trained in the k-t domain to output results matching the subtracted phase-cycled data, free of the artifact-generating T1-relaxation echo. The performance of the trained network was evaluated on the test set. All the test set images (real and imaginary parts separately) were compared to their corresponding phase-cycled subtracted data (as the ground truth) using Root-Mean-Square Error (RMSE) evaluated in the image domain.
Figure 3 shows the performance of the DAS-Net on a randomly selected example drawn from the test set for a systolic and a diastolic phase. As shown in the figure, in the k-space domain the T1-relaxation echo is suppressed and in the image domain the striping artifacts are largely suppressed. The RMSE over all the frames was calculated as (8.2±2.0)% for the validation set and was (8.4±1.9)% for the test set.
DAS-Net was found to effectively suppress the T1-relaxation echo in DENSE MRI and has potential as an alternative to phase-cycling for artifact suppression. By employing the DAS-Net, the DENSE acquisition time can be halved, which would provide a substantial improvement for clinical imaging. Additionally, the method may be used to suppress residual artifacts due to imperfect phase-cycling subtractions.