Jonathan M Trevathan1, Jonathan M Scott1, Joshua D Trzasko1, Armando Manduca1, John Huston1, Richard L Ehman1, and Matthew C Murphy1
1Mayo Clinic, Rochester, MN, United States
1Mayo Clinic, Rochester, MN, United States
A convolutional neural
network was trained to estimate a phasor corresponding to noise-free
displacements given noisy, complex-valued MRE data. The net outperformed its
filtering counterpart in noisy and noise-free simulation data, and decreased
test-retest repeatability error in vivo.
Simulated data with noise. Top) Left: Correlation with no filter.
Middle: Correlation with LPS filter. Right: Correlation with neural net
denoising. Bottom) from left to right: Original displacements, no filter, LPS filter,
neural net denoising.
Sagittal view of calculated x component
of the curl from in vivo data. Left: No Filter. Middle: LPS filter. Right:
Neural net denoised.
