1089

DAS-Net: A Generative Adversarial Net to Suppress Artifact-Generating Echoes in DENSE MRI
Mohammad Abdishektaei1, Xue of Feng1, Craig H Meyer1,2, and Frederick H Epstein1,2

1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, 2Radiology and Medical Imaging, University of Virginia, Charlottesville, VA, United States

Synopsis

In DENSE, displacement-encoded stimulated echoes are acquired with an artifact-generating signal due to T1 relaxation. Phase-cycling acquisitions are generally used to suppress the artifact-generating echoes which can result in imperfect artifact suppression when there is motion between the two acquisitions. To avoid this problem, a generative adversarial convolutional neural network (DAS-Net) is proposed to suppress the artifacts from a single acquisition. DAS-Net was trained on a DENSE dataset acquired from healthy volunteers. Results show that DAS-Net can effectively suppress the artifact-generating echoes and has the potential to obviate the need for phase-cycling acquisitions

Introduction

Cine Displacement Encoding with Stimulated Echoes (DENSE) provides an accurate and reproducible measure of tissue displacement [1]. In DENSE, displacement-encoded STimulated Echoes (STE) are acquired with an artifact-generating signal due to T1 relaxation. As a remedy, additional phase-cycled data are acquired and subtraction is used to cancel the T1-relaxation signal. However, phase cycling can result in imperfect artifact suppression when there is motion between the two acquisitions. [2]. Alternatively, a deep Convolutional Neural Net (CNN) is proposed to suppress the T1-relaxation echo from a single acquisition (without phase cycling), potentially obviating the need to acquire the phase-cycled data.

Methods

DENSE Artifact Suppression Net (DAS-Net) (Fig. 1) is an instance of a generative adversarial network [3] consisting of a generator and a discriminator competing in a zero-sum game. A U-Net [4] with 5 encoding and 5 decoding blocks was used as the generator trying to identify and suppress the T1-relaxation echo. A deep CNN with a dense layer as the last layer was used as the discriminator trying to evaluate the quality of the artifact suppression done by the generator. DENSE images from 27 healthy volunteers were acquired using 3T scanners (Siemens) leading to 49 DENSE datasets (40 short-axis and 9 long-axis datasets) with in-plane x and y displacement encoding as well as the background encoding. This set was divided into training (n=32), validation (n=8) and test (n=9) subsets. Spatial transformation (rotations and translations) in the image domain was used to augment the training set by a factor of 4 (Fig. 2) for both displacement-encoded and background-encoded data. Then the images were converted into kx-t/ky-t planes for displacement encoded data (24064 planes) and kx-t planes for background encoded data (12032 planes). These non-phase-cycled complex k-t planes were split into real and imaginary parts as two-channel k-t planes. The validation set was used to decide on network’s hyperparameters. DAS-Net was trained with displacement- and background- encoded data separately with the same hyperparameters. Finally, the performance of the trained network was evaluated on the test set. The output k-t planes of the DAS-Net on the test set were converted back to the image domain and were compared to the phase-cycled data using Root-Mean-Square Error (RMSE) evaluated on real and imaginary parts.

Results

The performance of DAS-Net is demonstrated on example images randomly drawn from the test set for x, y and background encodings (Fig. 3). The T1-relaxation echo was suppressed in k-space and the corresponding striping artifacts in the image domain were removed. The RMSE on validation and test sets are 8.8±3.3% 7.9±2.8% for displacement-encoded and 10.9±3.7% and 10.5±4.3% for background-encoded data respectively.

Conclusion

Our results show that CNNs have the potential to be an alternative solution to the phase-cycling acquisitions as DAS-Net effectively suppressed the T1-relaxation echo in DENSE MRI. With DAS-Net, the complications associated with phase-cycling in DENSE may be avoided, leading to significant improvement of DENSE in clinical settings. Furthermore, the method may be used to suppress residual artifacts due to imperfect phase-cycling subtractions. Evaluation of the strain maps extracted from output images will be performed in the future

Acknowledgements

No acknowledgement found.

References

  1. Kim D, Gilson WD, Kramer CM, Epstein FH, Myocardial Tissue Tracking with Two-dimensional Cine Displacement-encoded MR Imaging: Development and Initial Evaluation, Radiology, 2004.
  2. Cai X, Epstein FH, Free‐breathing cine DENSE MRI using phase cycling with matchmaking and stimulated‐echo image‐based navigators, MRM, 2018.
  3. Goodfellow IJ, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A, Bengio Y, Generative Adversarial Networks, arXiv, 2014.
  4. Ronneberger O, Fischer Ph, Brox T, U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv, 2015.

Figures

Architecture of DAS-Net

Augmentation of the training set with spatial transformations. Each input frame and its corresponding target image (phase-cycled subtracted) was rotated and translated on displacement (x, y) and background (b) encoded data

Comparison between output of DAS-Net to the ground truth for a randomly selected example within the test set. DAS-Net suppressed the T1-relaxation artifacts

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
1089