Qiang Zhang^{1}, Rui Guo^{1}, Huikun Qi^{1}, Di Cui^{2}, Edward S Hui^{2}, Shuo Chen^{1}, Hua Guo^{1}, and Huijun Chen^{1}

The purpose of this study is to develop a MR
fingerprinting (MRF) reconstruction algorithm using convolutional neural
network (MRF-CNN). Better MRF reconstruction fidelity was achieved using our
MRF-CNN compared with that of the conventional approach (R^{2} of T_{1}: 0.98 vs 0.97, R^{2} of T_{2}: 0.97 vs 0.59).
This study further demonstrated the performance of our MRF-CNN, which was
retrained using MR signal evolutions in the continuous parameter space with
various levels of Gaussian noise, amidst noise contamination, suggesting that
it may likely be a better alternative than the conventional MRF dictionary
matching approach.

1. Ma, D., V. Gulani, N. Seiberlich, K. Liu, J.L. Sunshine, et al., Magnetic resonance fingerprinting. Nature, 2013. 495(7440): p. 187-192.

2. Matsugu, M., K. Mori, Y. Mitari, and Y. Kaneda, Subject independent facial expression recognition with robust face detection using a convolutional neural network. Neural Networks, 2003. 16(5): p. 555-559.

3. Kleesiek, J., G. Urban, A. Hubert, D. Schwarz, K. Maier-Hein, et al., Deep MRI brain extraction: A 3D convolutional neural network for skull stripping. NeuroImage, 2016. 129(Supplement C): p. 460-469.

4. Akkus, Z., A. Galimzianova, A. Hoogi, D.L. Rubin, and B.J. Erickson, Deep Learning for Brain MRI Segmentation: State of the Art and Future Directions. Journal of Digital Imaging, 2017. 30(4): p. 449-459.

5. Bahrami, K., F. Shi, I. Rekik, and D. Shen, Convolutional Neural Network for Reconstruction of 7T-like Images from 3T MRI Using Appearance and Anatomical Features, in Deep Learning and Data Labeling for Medical Applications: First International Workshop, LABELS 2016, and Second International Workshop, DLMIA 2016, Held in Conjunction with MICCAI 2016, Athens, Greece, October 21, 2016, Proceedings, G. Carneiro, et al., Editors. 2016, Springer International Publishing: Cham. p. 39-47.

Fig. 1 The TR and FA trains used in this study.

Fig.2
The architecture of MRF reconstruction using Convolutional Neural Network. The
convolutional layer have 3×3 kernels and
the max pooling layers have 2×2 windows. Apart from the top layer which uses linear activation
function, Rectified Linear Unit (Relu) was used as the activation function.

Fig
3. The MRF reconstruction fidelity of our dictionary-trained MRF-CNN and the
conventional dictionary matching approach without noise. The red line was the
linear regression, with its formula in the top left corner.

Fig
4. The R^{2} (upper
row) and RMSE (bottom row)
of the conventional dictionary matching approach, dictionary-trained MRF-CNN
and retrained MRF-CNN for different levels of noise. The black line
indicated the dictionary-trained MRF-CNN, the red line indicated the retrained MRF-CNN,
and the blue line indicated the conventional dictionary matching approach.

Fig 5. The R^{2} (left) and RMSE (right) of retrained MRF-CNN
for different number of retrain MRF evolutions.