4790

Simultaneous Multi-Slice Deep RecOnstruction NEtwork (SMS-DRONE)

Ouri Cohen¹

¹Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States

Synopsis

Recently, MR fingerprinting (MRF) has been proposed as a means of disentangling simultaneously excited slices by exciting each slice with a distinct acquisition schedule. A notable drawback of this approach, which is particularly acute for multi-parametric dictionaries, is the linear increase in reconstruction time with the number of slices and the potential reduction in accuracy. Here we describe an extension to our previously described MRF-DRONE method that can overcome these issues. Our method can enable larger acceleration factors and faster reconstruction of multi-parametric data.

Introduction

In conventional simultaneous multi-slice (SMS) the excited slices are separated using the differential sensitivities of multi-channel RF coils[1]. Typical multi-channel coils have limited coverage in the slice direction and suffer from g-factor noise amplification which limits the achievable acceleration factors. Additionally, the reconstructed images are qualitative and are susceptible to instrumental contributions to the signal. In recent work[2],MR fingerprinting (MRF) was proposed as a means of disentangling the simultaneous slices by exciting each slice with a distinct acquisition schedule. The resulting signal is then pattern-matched voxel-wise to separate dictionaries with the best matching entry in each dictionary used to assign the quantitative tissue values in each slice. A notable drawback of this approach, which is particularly acute for multi-parametric dictionaries, is the linear increase in reconstruction time with the number of slices. Moreover, because each dictionary assumes only a single isochromat per voxel, the correlations between each dictionary and the measured signal get smaller with increasing number of slices. Here we describe an extension to our previously described MRF-DRONE method [3] that can overcome these issues. Our method can enable larger acceleration factors and faster reconstruction of multi-parametric data. We demonstrate the proof-of-principle with simulations on a numerical brain phantom.

Methods

A pulse sequence for an illustrative acceleration factor of R=2 is shown in Figure 1. A random schedule of flip angles (FA) and repetition times (TR) of length N=75 (corresponding to a 10 second acquisition) was used with an EPI readout [4] and with each RF pulse exciting two slices at different spatial locations. The FAs were chosen from the range 0-90° and the TR from the range 75-200 ms with the same TR used for both slices. A four layer neural network was defined in TensorFlow [5] consisting of an input layer, output layer and two hidden layers with 300×300 nodes (Figure 2). To train the network a 5000 entries, four dimensional dictionary was defined. Each dictionary atom consisted of four parameters representing the T1 and T2 values in the two simultaneously excited slices. The T1 and T2 values were chosen from the ranges T1=1-3000 ms and T2=1-500 ms using MATLAB’s (The Mathworks, Natick, MA) lhsdesign() function. The network was trained to convergence on an NVIDIA (Nvidia Inc., Santa Clara, CA) Tesla P40 GPU with 24 GB of RAM. To test the trained network a SMS acquisition of a numerical brain phantom[6] was simulated in MATLAB. The simulated data was used as input to the trained network which outputted the underlying T1 and T2 maps of each slice. Reconstruction of the four tissue maps with the trained network required approximately 400 ms.

Results

The quantitative T1 and T2 maps reconstructed by SMS-DRONE are shown in Figure 3 in comparison to the true values. A percentage error map was also calculated as Error = 100×|Recon-True|/True. The reconstructed tissue maps show excellent agreement to the true values with a mean error of 4-11% for this random acquisition schedule.

Discussion

In SMS-DRONE a functional mapping is found between the measured data and the underlying quantitative tissue maps. An important benefit of this approach is that a small training set suffices for accurate reconstructions [3], [7]. Unlike conventional dictionary matching, SMS-DRONE provides continuous-valued tissue maps and is not susceptible to the discretization artifacts inherent to multi-dimensional dictionaries[8]. The random acquisition schedule used in this work is likely far from optimal so optimization of the acquisition schedule [4] is expected to significantly improve the results and may further shorten scan times as well.

Conclusion

SMS-DRONE enables simultaneous quantification of tissue parameters arising from multiple slices without requiring the use of multi-channel coils. Future work will focus on further reducing the error and increasing the achievable acceleration factors.

Acknowledgements

Memorial Sloan Kettering Cancer Center

References

[1] M. Barth, F. Breuer, P. J. Koopmans, D. G. Norris, and B. A. Poser, “Simultaneous multislice (SMS) imaging techniques,” Magn. Reson. Med., vol. 75, no. 1, pp. 63–81, 2016.

[2] Y. Jiang et al., “Use of pattern recognition for unaliasing simultaneously acquired slices in simultaneous multislice MR fingerprinting,” Magn. Reson. Med., 2016.

[3] O. Cohen, B. Zhu, and M. S. Rosen, “MR fingerprinting deep reconstruction network (DRONE),” Magn. Reson. Med., 2018.

[4] O. Cohen and M. S. Rosen, “Algorithm comparison for schedule optimization in MR fingerprinting,” Magn. Reson. Imaging, 2017.

[5] M. Abadi et al., “Tensorflow: Large-scale machine learning on heterogeneous distributed systems,” ArXiv Prepr. ArXiv160304467, 2016.

[6] D. L. Collins et al., “Design and construction of a realistic digital brain phantom,” IEEE Trans. Med. Imaging, vol. 17, no. 3, pp. 463–468, 1998.

[7] O. Cohen, B. Zhu, and M. S. Rosen, “Characterization of Sparsely Trained Deep Learning Reconstruction of Noisy MR Fingerprinting Data,” in Proceedings of the International Society of Magnetic Resonance in Medicine, Paris, France, 2018.

[8] O. Cohen, C. T. Farrar, B. Zhu, and M. S. Rosen, “Fast Deep Learning Reconstruction of Highly Multi-dimensional MR Fingerprinting Data,” presented at the ISMRM Workshop on Machine Learning, Pacific Grove, CA, 2018.

Figures

Figure 1: The MRF-EPI SMS pulse sequence. Each RF pulse simultaneously excited 2 slices, labeled ‘a’ and ‘b’, with different, randomly selected, flip angles. The TR was the same for both slices but varied for each excitation according to the random acquisition schedule.

Figure 2: Illustration of the SMS-DRONE framework. The signal obtained from the simultaneous excitation of multiple slices is fed voxel-wise to a trained neural network which outputs the corresponding tissue parameters (T1,T2) of each slice.

Figure 3: Reconstructed quantitative T1 and T2 tissue maps obtained with SMS-DRONE in comparison to the true values. A percent error map is shown for each slice along with the mean error for each slice. Note the good agreement between the SMS-DRONE reconstruction and the true maps.

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

4790