Aliasing artifacts

[mckib2]

Aliasing artifacts

[estab1]

Thank you very much for your reply. I replicated your results with PULSAR_recon.py using my own phantom data and realized that PULSAR for matlab is producing the same results. I got confused about the results using basic_grappa.py because I forgot to add Gaussian noise to the k-space like I was doing in my matlab script. And thank you for the tip how to improve the image quality.

[mckib2]

Hi @estab1 , thanks for reporting! I'll have more time to look at this a little later today, but I think I've run into relative scaling issues in the past, so normalizing both the reference and reconstruction before comparing may help. I've not used PULSAR before, but I can take a look and see if there are any obvious differences in reconstruction technique -- is the demo recon.m script a good place to start?

[estab1]

Thanks for your quick reply. recon.m contains all the parameters needed for the reconstruction. And grappa.m contains the algorithm itself. So you probably need both to understand the reconstruction technique.

[mckib2]

So I took a look at that recon.m and grappa.m and I think these are similar parameters to what's happening there:

• R=2
• 32 ACS lines
• 90% of FE lines used

The following modified basic_grappa.py script attempts to replicate these:

PULSAR_recon.py

'''Replicate PULSAR recon.'''

import numpy as np
import matplotlib.pyplot as plt
from phantominator import shepp_logan

from pygrappa import grappa


if __name__ == '__main__':
    # Generate fake sensitivity maps: mps
    N = 128
    ncoils = 4
    xx = np.linspace(0, 1, N)
    x, y = np.meshgrid(xx, xx)
    mps = np.zeros((N, N, ncoils))
    mps[..., 0] = x**2
    mps[..., 1] = 1 - x**2
    mps[..., 2] = y**2
    mps[..., 3] = 1 - y**2

    # generate 4 coil phantom
    ph = shepp_logan(N)
    imspace = ph[..., None]*mps
    imspace = imspace.astype('complex')
    ax = (0, 1)
    kspace = np.fft.fftshift(np.fft.fft2(
        np.fft.ifftshift(imspace, axes=ax), axes=ax), axes=ax)

    # crop 32x90% window from the center of k-space for calibration
    pd = 16
    ctr = int(N/2)
    calib = kspace[ctr-pd:ctr+pd, int(.05*N):-int(.05*N), :].copy()

    # calibrate a kernel
    kernel_size = (7, 7)

    # undersample by a factor of 2 in ky
    kspace[:, 1::2, ...] = 0

    # reconstruct:
    res = grappa(
        kspace, calib, kernel_size, coil_axis=-1, lamda=0.01)

    # replace calib
    #res[ctr-pd:ctr+pd, int(.05*N):-int(.05*N), :] = calib

    # SOS recon
    res = np.sqrt(np.sum(np.abs(np.fft.fftshift(np.fft.ifft2(
        np.fft.ifftshift(res, axes=ax), axes=ax), axes=ax))**2, axis=-1))

    # SOS truth
    truth = np.sqrt(np.sum(np.abs(imspace)**2, axis=-1))
    residual = np.abs(res - truth)

    plt.title('Residual | recon - truth |')
    plt.imshow(residual, vmin=0, vmax=residual.flatten().max(), cmap='gray')
    plt.colorbar()
    plt.show()

The results I'm getting seem alright -- I was not able to run PULSAR because I don't have easy access to MATLAB. Do you know if it runs in Octave? A couple notes:

• The unmodified basic_grappa.py is reconstructing a different scenario: R=4 (Rx=Ry=2) with a smaller calibration region (20x20)
• you are comparing the coil reconstructions to truth -- this in general will not look good because the SNR of any single coil is worse than the combined reconstruction. I recommend doing SOS on the reconstructed coil images for a good comparison
• try tuning the kernel size: larger kernels run slower but produce less residual, I found (7, 7) to be a good compromise
• replacing the calibration region into your reconstructed k-space will help (if the calibration region came from your collected data, that is)
• The quality of the coil sensitivities might also be in question -- I generated 4 very simple non-complex coil sensitivities. You might try experimenting with more/better coils

EDIT: screen shots for posterity

Cranking kernel size and replacing calibration region: