4646

Signal Stability and Sensitivity of Referenceless Reconstructions by a Neural Network in Simultaneous Multi-Slice Imaging
Klaus Eickel1,2, Martin Blaimer3, and Matthias Günther1,2

1MR-Imaging & Spectroscopy, Faculty 01 (Physics/Electrical Engineering), University of Bremen, Bremen, Germany, 2MR Physics, Fraunhofer MEVIS, Bremen, Germany, 3Fraunhofer IIS, Würzburg, Germany

Synopsis

A deep neural network for reconstruction of SMS data without the need of additional reference data to calibrate for the spatial encoding information of the multi-coil receiver is presented. Noise-propagation through the reconstruction process is investigated in form of g-factors. Simulations with pseudo-multiple replicas showed robustness and stability of this new method. In addition, the sensitivity for physiological signal variations of this approach is considered in BOLD-signal dynamics. Results are compared to conventional methods like split slice-GRAPPA.

Introduction

Simultaneous multi-slice (SMS) imaging1-3 has emerged as a promising acceleration technique for multi-coil systems4 in magnetic resonance imaging (MRI) where undersampled data are recovered by utilizing the spatial encoding information inherent in multi-coil receiver arrays. However, the prominent methods like (split) slice-GRAPPA5,6 require additional, scan-specific reference data, i.e. auto-calibration signal (ACS), to disentangle overlapping image content7. ACS-acquisition can be time-consuming and a source of reconstruction-errors, if ACS data are corrupted. Here, a deep neural network (DNN), named SMSnet, reconstructs the SMS data without the need of any reference-scans8. The predicted images by SMSnet are compared to conventional approaches, i.e. split slice-GRAPPA (SSG) in terms of image quality, noise propagation and sensitivity to physiological changes. The spatially varying noise amplification in SMS can be characterized by the g-factor delivered by the pseudo-multiple-replica method (PMRM)9,10 to investigate signal stability. SMSnet’s sensitivity to small signal variations is evaluated in a time-series of blood-oxygenation-level dependent (BOLD) signal11.

Methods

The architecture depicted in Figure 1 is similar to8. Raw-data of 38 GRE and MPRAGE scans measured with a 20-channel head-coil at 3T (MAGNETOM Skyra, Siemens Healthineers, Erlangen, Germany) were augmented by downsampling and different CAIPIRINHA-patterns12 into a pool of Ntrain=41939 (Nval=24660) source- and target-pairs for training (validation). SMSnet was set up in Keras13 and training of 7 epochs was performed on a GTX1080 graphics card (Nvidia, Santa Clara, United States). A combined loss-function E = MSE x TV of mean-squared (MSE) and total-variation (TV) error accounted for global errors as well as borders mismatches. Acquired test data simulated a whole head scan in 6 slices (25mm gap), the parameters of the MRI protocol were: α=30°, TE=5ms, TR=126ms with an acquisition matrix of 96x104 (zero-filled to 128x128). 5 measurements yielded a ‘noise-free’ ground-truth (REF). A visual comparison between the 4 reconstruction methods was done first. The reconstruction performance was then quantified with respect to REF using the structural similarity index (SSIM). SSIM assesses perceptual image quality by taking advantage of characteristics of the human visual system14. Thereafter, PMRM was run for 100 repetitions, SMS data were synthesized thereafter for a multi-band-factor (MB) of 2 and a CAIPIRINHA-pattern of FOV-shift=1/4. These SMS data were processed in 4 reconstructions: a) SMSnet, b) SSGstd with correct ACS, c) SSGavg,h&p with corrupted ACS (averaged from multiple phantom- and head scans) and d) SSGavg,h also with corrupted ACS (averaged from multiple head scans).
In a second experiment, an accelerated (MB=2, FOV-shift=1/2) multi-echo (TE=8.6s/19.1s/30s) mutli-shot EPI scan (EPI-factor=13) was run for 130s (120 measurements) while the volunteer (31 years, male) performed finger-tapping (paradigm: 20s [rest] / 20s [active]). BOLD-dynamics were calculated after signal-separation into S0 and T2* by a mono-exponential fit15 and compared for SSGstd and SMSnet reconstructions of identical data.

Results

Visual inspection generally approves the quality of SMSnet reconstructions (Fig. 2). The 4 compared reconstruction methods reached mean SSIMs of 0.91 (SMSnet), 0.99 (SSGstd), 0.81 (SSGavg,h&p) and 0.87 (SSGavg,h) with respect to REF. Figure 3 shows the local, reconstruction-related noise-enhancement. While SMSnet and SSGstd perform almost equally in average (mean g-factors 1.71 and 1.72), higher local g-factors appear in SMSnet. Noise is clearly enhanced in SSGavg,h&p and SSGavg,h (mean g-factors 1.97 and 2.15) in accordance with the apparent leakage-artifacts. Signal-separation of the finger-tapping SMS-data displayed in Figure 4 gives reasonable T2*-curves corresponding to the paradigm. Both reconstructions lead to correlated results (r=0.97).

Discussion

In general, the reconstructions by SMSnet show promising and robust results with an acceptable g-factor penalty, while ensuring sensitivity not only to anatomical structures, but also physiological variations. Structures in proximity to regions of high susceptibility suffer from higher SNR-reductions (Fig. 2) which is physically plausible as coil-sensitivities are expect to vary stronger near these perturbations in the B-field homogeneity.
Main limitations of this work are that the presented test data have relatively large slice-distances and minimal acceleration (MB=2). Smaller slice-distances with higher MBs will require suitable training data and extension of the DNN’s architecture which can be realized easily with the modular design of SMSnet. Furthermore, transfer to coil-systems other than the head-coil may be of interest, in particular for applications where ACS can not be acquired reliably.

Conculsion

It is shown, that SMSnet reconstructs data with minor g-factor penalties compared to conventional SSG. However, SMSnet outperforms SSG reconstructions where correct ACS is missing as it does not rely on any scan-specific ACS to disentangle overlapping slices. SMSnet’s sensitivity to relatively small signal changes is proved in BOLD-dynamics induced by controlled finger-tapping. Furthermore, the reconstruction itself is fast, as no heavy computations have to be done after training.

Acknowledgements

The authors gratefully thank Markus Wenzel and Hans Meine for valuable, interdisciplinary discussions on deep learning in image-processing.

References

1. Weaver JB. Simultaneous multislice acquisition of MR images. Magn Reson Med 1988;8:275–284. 2. Müller S. Multifrequency selective rf pulses for multislice MR imaging. Magn Reson Med 1988;6:364–371. 3. Souza SP. Simultaneous Multislice Acquisition of MR Images by Hadamard-Encodes Excitation SOUZA_1988.pdf. 4. Larkman DJ, Hajnal J V., Herlihy AH, Coutts GA, Young IR, Ehnholm G. Use of multicoil arrays for separation of signal from multiple slices simultaneously excited. J Magn Reson Imaging 2001;13:313–317. 5. Setsompop K, Gagoski BA, Polimeni JR, Witzel T, Wedeen VJ, Wald LL. Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty. Magn Reson Med 2012;67:1210–1224. 6. Cauley SF, Polimeni JR, Bhat H, Wald LL, Setsompop K. Interslice leakage artifact reduction technique for simultaneous multislice acquisitions. Magn Reson Med 2014;72:93–102. 7. Barth M, Breuer F, Koopmans PJ, Norris DG, Poser BA. Simultaneous multislice (SMS) imaging techniques. Magn Reson Med 2016;75:63–81. 8. Eickel K, Günther M. A Neural Network for Referenceless Reconstruction in Simultaneous Multi-Slice Imaging.Proceedings ISMRM 2018 Paris; 2782 9. Robson PM, Grant AK, Madhuranthakam AJ, Lattanzi R, Sodickson DK, McKenzie CA. Comprehensive quantification of signal-to-noise ratio and g -factor for image-based and k -space-based parallel imaging reconstructions. Magn Reson Med. 2008;60: 895–907. doi:10.1002 10. Kellman P, McVeigh ER. Image reconstruction in SNR units: A general method for SNR measurement. Magn Reson Med. 2005;54: 1439–1447. doi:10.1002/mrm.20713/mrm.21728 11. Eickel K, Günther M, Lüdemann L. Dynamic T2*-Mapping using Segmented EPI with Multi-TE. Klöck S, editor. Dreiländertagung der Medizinischen Phys Zürich 2014 SGSMP, ÖGMP, DGMP, Hrsg Stephan Klöck. Zürich: Stephan Klöck; 2014; 410. 12. Breuer FA, Blaimer M, Heidemann RM, Mueller MF, Griswold MA, Jakob PM. Controlled aliasing in parallel imaging results in higher acceleration (CAIPIRINHA) for multi-slice imaging. Magn Reson Med 2005;53:684–691. 13. Chollet F et al. Keras. 2015. 14. Wang Z, Bovik AC, Sheikh HR, Member S, Simoncelli EP, Member S. Image Quality Assessment: From Error Visibility to Structural Similarity. 2004;13:600–612. 15.Poser BA, Norris DG. Investigating the benefits of multi-echo EPI for fMRI at 7 T. Neuroimage. 2009;45: 1162–1172. doi:10.1016/j.neuroimage.2009.01.007

Figures

The 20-channel preprocessed complex input data are split into real and imaginary components which are concatenated along channel dimension resulting in 40 input-channels. In the left tree, named CS-path, features similar to coil sensitivities are derived from reduced k-space data (32x32x40). These are merged by multiplication with the input image data (128x128x40) after passing the right tree, named Im-path, similar to the unfolding process in SSG. The last section, called merge-path, reduces the number of channels. The combined channels yielding a single magnitude image per slice (128x128) at the output. SMSnet for MB=2 has about 9.9x106 trainable parameters.

Reconstructions of synthetic SMS data (MB=2, FOV-shift=1/4). REF represent the noise-free ground-truth. SMSnet and SSGstd recover the test data without obvious artifacts unlike SSG reconstructions with corrupted ACS (SSGavg,h&p,SSGavg,h). Images were slice-wise normalized before plotting.

Monte-Carlo generated g-factor maps show the reconstruction-related noise-enhancement for all 4 reconstruction methods. Masks were applied to remove irritating background and limit the calculation of the mean g-factor to voxels inside the object.

T2*-dynamics, i.e. BOLD-signal, of identical data, but reconstructed differently. Results of SMSnet and SSGstd correlate with r=0.97. Signal was averaged from 6 neighboring voxels.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
4646