Figure 1.
A: Our ULF-MRI system, which includes the x-, y-, and z-gradients (orange, red, and blue, respectively) and the magnet to generate the measurement field (green).
The polarizing and excitation coils are not shown in the figure. B: The posterior view of the system shows 47 SQUID sensors covering the posterior parts of the head.
Figure 2.
The simulated noiseless sum-of-squares (SoS) image from all 47 channels of the ULF-MRI system (left).
At different SNRs compared to the direct SoS reconstruction, SNR can be improved by incorporating the data consistency constraint (λ = 0). Using the sparsity prior (λ = 0.03, 0.1, and 0.5), the residual error can be further reduced with low SNR acquisitions. The residual errors are reported at the lower-right corner of each reconstruction.
Figure 3.
A hand sum-of-squares (SoS) image (left).
The data consistency constraint (λ = 0) reduces significantly the noticeable vertical strip artifact (middle). Further, the sparsity prior (λ = 0.1) improves the reconstruction only marginally (right). The pSNR was indicated in each image.
Figure 4.
Brain images reconstructed by the regularized SENSE reconstructions with no acceleration (left column).
The data consistency constraint (λ = 0) improves the image by showing a strong signal in the brain parenchyma (middle column). Further, the sparsity prior (λ = 0.1) suppresses the background noise significantly to better delineate the skull and the brain (right column). The pSNR was indicated in each image.
Figure 5.
A brain image with different number of averages reconstructed by the regularized SENSE reconstruction with no acceleration.
The pSNR (cyan) and MSE (green) were reported in each image.