4791

Deep Learning Super-FOV for Accelerated bSSFP Banding Reduction
Nicholas McKibben1, Michael Mendoza1, Edward DiBella2, and Neal K. Bangerter3

1Electrical and Computer Engineering, Brigham Young University, Provo, UT, United States, 2Radiology, University of Utah, Salt Lake City, UT, United States, 3Imperial College London, London, United Kingdom

Synopsis

We present a technique for bSSFP band removal using two undersampled phase-cycled bSSFP image acquisitions.

Introduction

Balanced steady state free precession (bSSFP) is an imaging technique capable of fast image acquisition with a high signal-to-noise ratio (SNR). However, bSSFP is highly sensitive to magnetic field inhomogeneities. This sensitivity is known to create characteristic banding artifacts which degrade image quality. Several methods for reducing banding artifacts have been developed using the combination of multiple phase-cycled images in order to effectively reduce the banding.1 The acquisition of multiple phase-cycled images increases scan time and is a limiting factor. We present a technique for state of the art band reduction using undersampled bSSFP acquisitions combined with a deep neural network. This method uses fewer images compared to some techniques for band reduction, like the elliptical signal model2, and is capable of reducing scan time. We validate our technique with in vivo experimental results.

Theory

A banding artifact reduction method was created using a deep neural network. Previously it was shown that such a network could be designed to effectively reduce banding artifacts using a U-Net architecture.3,4 The ground truth data was generated using four phase-cycled images combined with the geometric solution to the elliptical signal model in order to generate a band reduced image. Using this work as a basis, the network was improved to utilize undersampled bSSFP images as the inputs to the network to solve a generalized SENSE reconstruction. This improvement is analogous to previous work for SSFP banding artifact reduction Super-FOV.5 In the Super-FOV method, the banding artifacts of SSFP are used as a sensitivity map in order to reconstruct a pair of undersampled phase-cycled bSSFP images. The sensitivity maps are used in a similar manner to other subsampled reconstruction techniques like GRAPPA and SENSE. Figure 1 outlines the imaging and reconstruction process for this technique.

Methods

We acquired sets of four phase-cycled images using a 3T Siemens TIM Trio scanner. This experiment produced the data necessary for both training and testing the model. The model was trained using two undersampled phase-cycled images and the truth data was generated from the results of the geometric solution to the elliptical signal model. As inputs to the elliptical model, four images corresponding to the RF phase increments of 0°, 90°, 180°, and 270° from coronal slices of a knee were acquired. This data was acquired using a FOV of 250 mm, TR of 4.6ms, TE of 2.3ms, a tip angle of 55°, and an acquisition matrix size of 256x256x128. Two of the bSSFP images with RF phase increments 0° and 180° were undersampled retroactively with a undersampling factor of R=2.

Results and Conclusion

The results of the experiment are shown in Figure 2. After training the model, the reconstruction exhibits band reduction similar to the geometric solution of the elliptical signal model. The network-generated banding-free images compared to the true solution to the elliptical signal model has average mean squared error of (5.38±2.06)e-7, mSSIM index of 0.9998±6.66e-5, and average PSNR of 69.02±1.61.

Discussion

We have demonstrated that a deep neural network can reduce banding artifacts in multiple acquisition bSSFP with comparable performance to the elliptical signal model and reduce overall scan time by requiring fewer phase-cycled images. There are a variety of applications which could benefit from the excellent bSSFP contrast and SNR efficiency, but which cannot tolerate severe banding at low repetition times such as musculoskeletal imaging, dynamic cardiac MR6, and transition band functional MRI7. This technique might be improved by measuring an auto-calibration signal in the center of k-space in addition to the undersampled acquisitions taken. This would increase acquisition time, but allow for much better estimation of the phase sensitivity maps required for reconstruction. The model could also be improved with the addition of an additional branching decoder that reconstructs the fully sampled input images, allowing for the inclusion of a data fidelity term in the training cost function.

Acknowledgements

Many thanks to Bradley Bolster from Siemens for his support of the BYU MRI Research Facility

References

  1. Xiang, Q., Hoff, M., Banding artifact removal for bSSFP imaging with an elliptical signal model, Magnetic Resonance in Medicine, 2014
  2. Bangerter, Neal K., et al. "Analysis of multiple‚Äźacquisition SSFP." Magnetic Resonance in Medicine, 51.5 (2004): 1038-1047.
  3. Mendoza M, McKibben N, Hales L, Taylor M, Bangerter NK. Banding Artifact Reduction in Musculoskeletal bSSFP using Deep Learning. Proceedings of ISMRM Machine Learning Workshop Part II, Oct 2018.
  4. Ronneberger, O., Fischer, P., Brox, T., U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv:1505.04597
  5. Lustig M., Santos J., Pauly J.M., A Super-FOV method for rapid SSFP banding artifact reduction. Proceedings of ISMRM, 2005.
  6. Schär, Michael, et al. "Cardiac SSFP imaging at 3 Tesla." Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 51.4 (2004): 799-806.
  7. Miller, Karla L. "FMRI using balanced steady-state free precession (SSFP)." Neuroimage 62.2 (2012): 713-719.

Figures

Undersampled k-space data from two phase-cycled acquisitions is fed to the model as input. The output is compared using mean squared error to the geometric solution to the elliptical signal model generated from four fully sampled phase-cycled acquisitions.

(a-b) is the combined k-space and image domain representation, respectively, of the two undersampled acquisitions. The combined k-space has even lines from first phase-cycled image and odd lines from the second, as it was input to the neural network for training. (c-d) show the k-space and image space representations of the geometric solution to the elliptical signal model which was used as ground truth for training. (d-e) show the model’s estimate of the banding-free image.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
4791