Roberto Souza^{1}, Youssef Beauferris^{1}, Wallace Loos^{1}, Mariana Bento^{1}, Robert Marc Lebel^{2}, and Richard Frayne^{1}

^{1}University of Calgary, Calgary, AB, Canada, ^{2}GE, Calgary, AB, Canada

Magnetic resonance (MR) compressed sensing reconstruction explores image sparsity to make MR acquisition faster while still reconstructing high quality images. Modern picture archiving and communication systems allow efficient access to previous scans acquired of the same subject. In this work, we propose to use previous scans to enhance the reconstruction of follow-up scans using a deep learning model. Our model is composed of a reconstruction network that outputs an initial MR reconstruction, which is used as input to an enhancement network along with a co-registered previous scan. Our enhancement network improved quantitative metrics on average by 15%.

^{1}D. Lee, J. Yoo, and J. C. Ye, “Deep residual learning
for compressed sensing MRI,” in IEEE International Symposium on Biomedical
Imaging. Conference Proceedings, pp. 15–18, 2017.

^{2}J. Schlemper, J. Caballero, J. Hajnal, A. Price, and D. Rueckert, “A
deep cascade of convolutional neural networks for dynamic MR image
reconstruction,” IEEE Transactions on Medical Imaging, vol. 37, no. 2, pp.
491–503, 2018.

^{3}L. Xiang, Y. Chen, W. Chang, Y. Zhan, W. Lin, Q. Wang,
and D. Shen, “Deep learning based multi-modal fusion for fast MR
reconstruction,” IEEE Transactions on Biomedical Engineering, 2018.

^{4}E. G. Larsson, D. Erdogmus, R. Yan, J. C. Principe,
and J. R. Fitzsimmons, “SNR-optimality of sum-of-squares reconstruction for
phased-array magnetic resonance imaging,” Journal of Magnetic Resonance, vol.
163, no. 1, pp. 121–123, 2003.

^{5}Jenkinson, M.,
Bannister, P., Brady, J. M. and Smith, S. M. "Improved Optimisation for the
Robust and Accurate Linear Registration and Motion Correction of Brain Images".
NeuroImage, 17(2), 825-841, 2002.

^{6}Souza, R., Frayne,
R., “W-net: A Hybrid Compressed Sending MR Reconstruction Model”, ISMRM 2019.

^{7}Ronneberger, O.,
Fischer, P. and Brox, T., "U-net: Convolutional networks for
biomedical image segmentation". In International Conference on Medical
image computing and computer-assisted intervention (pp. 234-241), 2015.

^{8}Wilcoxon, F. "Individual Comparisons by Ranking
Methods". Biometrics Bulletin, 1(6), 80-83, 1945.

Figure 1 Flow chart depicting the proposed methodology for the
data acquisition and reconstruction process. Traditional reconstruction refers
to MR image reconstruction that does not leverage past scans. If a previous
scan for a subject is available in the Picture Archive and Communications
System (PACS), then the traditional reconstruction could then be enhanced by
leveraging this past scan. This methodology can potentially be used to
accelerate even further MR acquisition.

Figure 2 Flowchart of the proposed methodology. The model
receives as input multi-channel (MC) under-sampled k-spaces, which is processed
by the reconstruction network resulting into MC images that are combined using
root sum of squares to yield an initial reconstruction R_{0}. The
subject previous scan (PS) is registered to R_{0} resulting in PS_{reg}.
R_{0} and PS_{reg} are combined into a 2-channel image that is
processed by the enhancement network resulting in the final reconstruction R_{enhanced}.

Figure 3 SSIM, NRMSE and pSNR plotted with
error bars (standard deviation) for each acceleration. For higher accelerations
(R>10), the improvement of the enhancement
network is more noticeable. The size of the error bars can be explained by the
initial slices which contain little anatomy. First 50 slices were discarded from the metrics
computation for all subjects to reduce the effects of this noise on the
metrics.

Figure 4 Sample reconstruction and residual images of a
slice comparing the quality of the traditional reconstruction against the
enhanced reconstruction for acceleration factors of R=5×, 10×, 15×, and 20×.
Notice not just the improvement in the metrics, but also the brain edges become
sharper after being processed by the enhancement network.