Shouchang Guo1 and Douglas C. Noll2
1Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, United States, 2Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States
1Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, United States, 2Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States
The proposed recurrent U-Net with two levels of temporal regularization presents higher quality fMRI reconstruction than other methods.
Functional results including activation maps, temporal SNR maps, and time courses. The proposed approach has well preserved functional signals with similar activation map and time course to the fully sampled case.
Because the data shared images were reconstructed by combining k-space of every 10 slow time points, the time-series of data shared images were generated by repeating each data shared image for 10 times along slow time.
The proposed network that reconstructs nc = 10 fast time images as a sequence. The network combines U-Nets with a recurrent layer. The recurrent layer “fw” is located at the bottleneck of the U-Net, and takes learned representations from the U-Net encoder and hidden states hi (i = 1, 2, ... , nc) from the previous fast time to generate the hidden state for the next fast time frame. xi denotes zero-filled fast time image, di denotes data shared image, and yi denotes two-channel (real and imaginary) output image from the network.
