4781

Synthetic Banding for bSSFP Data Augmentation
Michael Anthony Mendoza1, Nicholas McKibben1, Grayson Tarbox1, and Neal K Bangerter1,2

1Electrical Engineering, Brigham Young University, Provo, UT, United States, 2Bioengineering, Imperial College London, London, United Kingdom

Synopsis

Balanced Steady State Free Precession (bSSFP) MRI is a highly-efficient MRI pulse sequence but suffers from banding artifacts caused by its high sensitivity to magnetic field inhomogeneity. Many algorithms exist that can effectively remove these banding artifacts, typically by requiring multiple phase-cycled acquisitions, which increase scan time. While some of the algorithms can suppress banding to some degree with two sets of phase-cycled acquisitions, much more accurate band suppression is typically achieved with at least four phase-cycled acquisitions. In this work, we present a deep learning method for synthesizing additional phase-cycled images from a set of at least two phase-cycled images that can then be used with existing band reduction techniques in order to reduce scan time.

Introduction

Banding reduction techniques for balanced steady state free precession (bSSFP) acquisition often require the collection of many phase-cycled images.1 In general, a minimum of two phase-cycled images are required to reduce banding since with only one phase-cycled image parts of the image may fall in a signal null. By having a combination of two different phase-cycled images it is possible to acquire signal for each voxel in a field-of-view. While a minimum of two phase-cycles is required, many banding reduction techniques and post-processing algorithms rely on the availability of at least four phase-cycled images.2,3,4 However, in some applications such as dynamic contrast enhanced cardiac MR, it is not possible to collect four fully sampled phase-cycles for each slice or time point. Recently, methods have been proposed that collect multiple phase-cycled images simultaneously, greatly reducing the acquisition time required.5 Inspired by these methods, we propose a deep learning model that generates phase-cycled images in order to reduce the number of acquired phase-cycled images. These generated images augment the existing phase-cycled datasets. The combination of acquired and augmented images can be used in existing band reduction methods to reduce scan time. Additionally, by producing these supplemental images a variety of post processing techniques for parameter estimation such T1, T2 or field map estimates can be generated from augmented datasets.

Theory

A bSSFP data augmentation method was creating using a deep neural network. Figure 1 outlines the image reconstruction architecture for the method. A modified U-net architecture was used as the basis for the neural network. The mean squared error between the output and truth data was used as the cost function for training. As input to the network, two phase cycled images with RF phase-cycles of 0° and 180° were used. The ground truth for the network was a pair of phase-cycled images with incremental RF phases of 90°, and 270°.

As verification of the experiment, the four phase-cycled images used in training were input into the geometric solution of the elliptical signal model (ESM) to generate a band free image. Then the combination of the acquired phase-cycled images (0°, 180°) and the generated phased-cycled images (90°, 270°) were input into the same ESM model. This verification procedure is outlined in Figure 2. The resultant band reduced images then were compared.

Methods

To train the deep neural network, four phase-cycled 3D knee volumes were acquired on a Siemens 3T TIM Trio with increasing RF phases increments of 0°, 90°, 180°, and 270°. All images were acquired using a bSSFP pulse sequence with parameters TR=2.3ms, TE=4.6ms, a flip angle ɑ of 50°, FOV 220mm, matrix size 256x256x128, and 4 averages.

Results

The synthetically generated phase-cycled images are shown along side the original images in Figure 3. After training the model, the network successfully generated additional phase-cycles (90°, 270°) from the acquired phase-cycled images (0°, 180°). The generated images are comparable to the acquired images. The synthetically generated images compared to acquired images have average mean squared error of (6.21∓1.96)e-4, mSSIM of 0.86987∓0.042, and average PSNR of 38.35∓1.58. As further verification of the method, Figure 4 shows that the reconstruction from the elliptical signal model of four acquired images exhibits band reduction similar to reconstruction from the elliptical signal model of two acquired images combined with the synthetically generated images.

Conclusion and Discussion

We have demonstrated that given two phase-cycled bSSFP acquisitions, a deep convolutional neural network can be trained to augment the dataset to four phase-cycled images without the need for additional acquisitions. Using a set of four images consisting of acquired and augmented phase-cycled images, ESM-based algorithms can be used to eliminate banding2. Additionally, this data is suitable as input to algorithms estimating tissue parameters and field maps3,7. We have also demonstrated that correct phase information is synthesized, as the solution to the ESM is very close to the true solution, allowing phase information to propagate through a post-processing pipeline. Future work could involve more complicated network architectures: recent work suggests that sharper images may be generated by generative adversarial networks.6 Some algorithms also require or are improved with more than four input phase-cycled images4, leading to the possibility of synthesizing more than two phase-cycled output images.

Acknowledgements

We would like to thank Bradley Bolster from Siemens for his support of the BYU MRI Research Facility

References

  1. Bangerter, Neal K., et al. "Analysis of multiple‐acquisition SSFP." Magnetic Resonance in Medicine, 51.5 (2004): 1038-1047.
  2. Xiang, Qing‐San, and Michael N. Hoff. "Banding artifact removal for bSSFP imaging with an elliptical signal model." Magnetic resonance in medicine 71.3 (2014): 927-933.
  3. Taylor, Meredith, et al. "MRI Field Mapping using bSSFP Elliptical Signal model." Proceedings of ISMRM 25th Joint Annual Meeting (2017).
  4. Mendoza et al. “Generation of Arbitrary Spectral Profiles using Orthonormal Basis Combinations of bSSFP MRI.” Proceedings of ISMRM 26th Joint Annual Meeting (2018).
  5. Wang, Yi, et al. "Phase‐cycled simultaneous multislice balanced SSFP imaging with CAIPIRINHA for efficient banding reduction." Magnetic resonance in medicine 76.6 (2016): 1764-1774.
  6. Liu, Fang, Li Feng, and Richard Kijowski. "MANTIS: Model-Augmented Neural neTwork with Incoherent k-space Sampling for efficient MR T2 mapping." arXiv preprint arXiv:1809.03308 (2018).
  7. Shcherbakova, Yulia, et al. "PLANET: An ellipse fitting approach for simultaneous T1 and T2 mapping using phase‐cycled balanced steady‐state free precession." Magnetic Resonance in Medicine 79.2 (2018): 711-722.

Figures

Figure 1: An illustration of the model architecture

Figure 2: An illustration of model verification using the geometric solution from elliptical signal model. (Left) Band reduced image is generated using four acquired images. (Right) Band reduce image is generated using two acquired images and two synthetically generated images

Figure 3: bSSFP knee images. (a-d) are the acquired bSSFP phase-cycled images (0°, 180°, 90° and 270°) (e-f) are the synthetically generated bSSFP phase-cycled images (90° and 270°)

Figure 4: bSSFP band reduced images using the geometric solution to the elliptical signal model. (a) Inputs into the model are four acquired bSSFP phase-cycled images. (b) Inputs into the model are two acquired and two synthetically generated bSSFP phase-cycled images.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
4781