def compute_corr_3d(src=[], target=[], x=0, xshift=0, xsize=0, y=0, yshift=0, ysize=0, z=0, zshift=0, zsize=0, xtarget=0, ytarget=0, ztarget=0, zrange=[], verbose=1, save_suffix='', gaussian_weighting=True, path_output='../'):
"""
Find z that maximizes correlation between src and target 3d data.
:param src: 3d source data
:param target: 3d target data
:param x:
:param xshift:
:param xsize:
:param y:
:param yshift:
:param ysize:
:param z:
:param zshift:
:param zsize:
:param xtarget:
:param ytarget:
:param ztarget:
:param zrange:
:param verbose:
:param save_suffix:
:return:
"""
# parameters
thr_corr = 0.2 # disc correlation threshold. Below this value, use template distance.
# get dimensions from src
nx, ny, nz = src.shape
# Get pattern from template
pattern = target[
xtarget - xsize: xtarget + xsize + 1,
ytarget + yshift - ysize: ytarget + yshift + ysize + 1,
ztarget + zshift - zsize: ztarget + zshift + zsize + 1]
pattern1d = pattern.ravel()
# initializations
I_corr = np.zeros(len(zrange))
allzeros = 0
# current_z = 0
ind_I = 0
# loop across range of z defined by src
for iz in zrange:
# if pattern extends towards the top part of the image, then crop and pad with zeros
if z + iz + zsize + 1 > nz:
# print 'iz='+str(iz)+': padding on top'
padding_size = z + iz + zsize + 1 - nz
data_chunk3d = src[
x - xsize: x + xsize + 1,
y + yshift - ysize: y + yshift + ysize + 1,
z + iz - zsize: z + iz + zsize + 1 - padding_size]
data_chunk3d = np.pad(data_chunk3d, ((0, 0), (0, 0), (0, padding_size)), 'constant',
constant_values=0)
# if pattern extends towards bottom part of the image, then crop and pad with zeros
elif z + iz - zsize < 0:
# print 'iz='+str(iz)+': padding at bottom'
padding_size = abs(iz - zsize)
data_chunk3d = src[
x - xsize: x + xsize + 1,
y + yshift - ysize: y + yshift + ysize + 1,
z + iz - zsize + padding_size: z + iz + zsize + 1]
data_chunk3d = np.pad(data_chunk3d, ((0, 0), (0, 0), (padding_size, 0)), 'constant',
constant_values=0)
else:
# print 'iz='+str(iz)+': no padding'
data_chunk3d = src[
x - xsize: x + xsize + 1,
y + yshift - ysize: y + yshift + ysize + 1,
z + iz - zsize: z+ iz + zsize + 1]
# if verbose == 2 and iz in range(0, nz, 10):
# # display template and subject patterns
# plt.figure(11)
# plt.subplot(131)
# plt.imshow(np.flipud(np.mean(pattern[:, :, :], axis=0).transpose()), origin='upper',
# cmap=plt.cm.gray,
# interpolation='none')
# plt.title('Template pattern')
# plt.subplot(132)
# plt.imshow(np.flipud(np.mean(data_chunk3d[:, :, :], axis=0).transpose()), origin='upper',
# cmap=plt.cm.gray,
# interpolation='none')
# plt.title('Subject pattern at iz=0')
# # save figure
# plt.figure(11), plt.savefig('../fig_pattern'+save_suffix+'_z'+str(iz)+'.png'), plt.close()
# convert subject pattern to 1d
data_chunk1d = data_chunk3d.ravel()
# check if data_chunk1d contains at least one non-zero value
# if np.any(data_chunk1d): --> old code which created issue #794 (jcohenadad 2016-04-05)
if (data_chunk1d.size == pattern1d.size) and np.any(data_chunk1d):
#I_corr[ind_I] = np.corrcoef(data_chunk1d, pattern1d)[0, 1]
# data_chunk2d = np.mean(data_chunk3d, 1)
# pattern2d = np.mean(pattern, 1)
I_corr[ind_I] = mutual_information(data_chunk1d, pattern1d, nbins=16)
else:
allzeros = 1
# printv('.. WARNING: iz='+str(iz)+': Data only contains zero. Set correlation to 0.', verbose)
ind_I = ind_I + 1
# ind_y = ind_y + 1
if allzeros:
printv('.. WARNING: Data contained zero. We probably hit the edge of the image.', verbose)
# adjust correlation with Gaussian function centered at the right edge of the curve (most rostral point of FOV)
if gaussian_weighting:
gaussian_std_factor = 0.5
from scipy.signal import gaussian
gaussian_window = gaussian(len(I_corr) * 2, std=len(I_corr) * gaussian_std_factor)
I_corr = np.multiply(I_corr, gaussian_window[0:len(I_corr)])
# Find global maximum
# ind_peak = ind_peak[np.argmax(I_corr[ind_peak])]
if np.any(I_corr):
# if I_corr contains at least a non-zero value
ind_peak = [i for i in range(len(I_corr)) if I_corr[i] == max(I_corr)][0] # index of max along z
# ind_peak[1] = np.where(I_corr == I_corr.max())[1] # index of max along y
printv('.. Peak found: z=' + str(zrange[ind_peak]) + ' (correlation = ' + str(I_corr[ind_peak]) + ')', verbose)
# check if correlation is high enough
if I_corr[ind_peak] < thr_corr:
printv('.. WARNING: Correlation is too low. Using adjusted template distance.', verbose)
ind_peak = zrange.index(0) # approx_distance_to_next_disc
# ind_peak[1] = int(round(len(length_y_corr)/2))
else:
# if I_corr contains only zeros
printv('.. WARNING: Correlation vector only contains zeros. Using adjusted template distance.', verbose)
ind_peak = zrange.index(0) # approx_distance_to_next_disc
# display patterns and correlation
if verbose == 2:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# display template pattern
plt.figure(11, figsize=(15, 7))
plt.subplot(131)
plt.imshow(np.flipud(np.mean(pattern[:, :, :], axis=0).transpose()), origin='upper', cmap=plt.cm.gray, interpolation='none')
plt.title('Template pattern')
# display subject pattern at best z
plt.subplot(132)
iz = zrange[ind_peak]
data_chunk3d = src[
x - xsize: x + xsize + 1,
y + yshift - ysize: y + yshift + ysize + 1,
z + iz - zsize: z + iz + zsize + 1]
plt.imshow(np.flipud(np.mean(data_chunk3d[:, :, :], axis=0).transpose()), origin='upper', cmap=plt.cm.gray, clim=[0, 800], interpolation='none')
plt.title('Subject at iz=' + str(iz))
# display correlation curve
plt.subplot(133)
plt.plot(zrange, I_corr)
plt.title('corr(subject, template)')
plt.plot(zrange[ind_peak], I_corr[ind_peak], 'ro'), plt.draw()
plt.axvline(x=zrange.index(0), linewidth=1, color='black', linestyle='dashed')
plt.axhline(y=thr_corr, linewidth=1, color='r', linestyle='dashed')
plt.grid()
# save figure
plt.figure(11), plt.savefig(path_output + 'fig_pattern'+save_suffix+'.png'), plt.close()
# return z-origin (z) + z-displacement minus zshift (to account for non-centered disc)
return z + zrange[ind_peak] - zshift
def compute_corr_3d(src, target, x, xshift, xsize, y, yshift, ysize, z, zshift,
zsize, xtarget, ytarget, ztarget, zrange, verbose,
save_suffix, gaussian_std, path_output):
"""
Find z that maximizes correlation between src and target 3d data.
:param src: 3d source data
:param target: 3d target data
:param x:
:param xshift:
:param xsize:
:param y:
:param yshift:
:param ysize:
:param z:
:param zshift:
:param zsize:
:param xtarget:
:param ytarget:
:param ztarget:
:param zrange:
:param verbose:
:param save_suffix:
:param gaussian_std:
:return:
"""
# parameters
thr_corr = 0.2 # disc correlation threshold. Below this value, use template distance.
# get dimensions from src
nx, ny, nz = src.shape
# Get pattern from template
pattern = target[xtarget - xsize:xtarget + xsize + 1,
ytarget + yshift - ysize:ytarget + yshift + ysize + 1,
ztarget + zshift - zsize:ztarget + zshift + zsize + 1]
pattern1d = pattern.ravel()
# initializations
I_corr = np.zeros(len(zrange))
allzeros = 0
# current_z = 0
ind_I = 0
# loop across range of z defined by src
for iz in zrange:
# if pattern extends towards the top part of the image, then crop and pad with zeros
if z + iz + zsize + 1 > nz:
# sct.printv('iz='+str(iz)+': padding on top')
padding_size = z + iz + zsize + 1 - nz
data_chunk3d = src[x - xsize:x + xsize + 1,
y + yshift - ysize:y + yshift + ysize + 1, z +
iz - zsize:z + iz + zsize + 1 - padding_size]
data_chunk3d = np.pad(data_chunk3d,
((0, 0), (0, 0), (0, padding_size)),
'constant',
constant_values=0)
# if pattern extends towards bottom part of the image, then crop and pad with zeros
elif z + iz - zsize < 0:
# sct.printv('iz='+str(iz)+': padding at bottom')
padding_size = abs(iz - zsize)
data_chunk3d = src[x - xsize:x + xsize + 1,
y + yshift - ysize:y + yshift + ysize + 1, z +
iz - zsize + padding_size:z + iz + zsize + 1]
data_chunk3d = np.pad(data_chunk3d,
((0, 0), (0, 0), (padding_size, 0)),
'constant',
constant_values=0)
else:
data_chunk3d = src[x - xsize:x + xsize + 1,
y + yshift - ysize:y + yshift + ysize + 1,
z + iz - zsize:z + iz + zsize + 1]
# convert subject pattern to 1d
data_chunk1d = data_chunk3d.ravel()
# check if data_chunk1d contains at least one non-zero value
if (data_chunk1d.size == pattern1d.size) and np.any(data_chunk1d):
I_corr[ind_I] = mutual_information(data_chunk1d,
pattern1d,
nbins=16,
normalized=False)
else:
allzeros = 1
ind_I = ind_I + 1
# ind_y = ind_y + 1
if allzeros:
sct.printv(
'.. WARNING: Data contained zero. We probably hit the edge of the image.',
verbose)
# adjust correlation with Gaussian function centered at the right edge of the curve (most rostral point of FOV)
from scipy.signal import gaussian
gaussian_window = gaussian(len(I_corr) * 2, std=len(I_corr) * gaussian_std)
I_corr_gauss = np.multiply(I_corr, gaussian_window[0:len(I_corr)])
# Find global maximum
if np.any(I_corr_gauss):
# if I_corr contains at least a non-zero value
ind_peak = [
i for i in range(len(I_corr_gauss))
if I_corr_gauss[i] == max(I_corr_gauss)
][0] # index of max along z
sct.printv(
'.. Peak found: z=' + str(zrange[ind_peak]) + ' (correlation = ' +
str(I_corr_gauss[ind_peak]) + ')', verbose)
# check if correlation is high enough
if I_corr_gauss[ind_peak] < thr_corr:
sct.printv(
'.. WARNING: Correlation is too low. Using adjusted template distance.',
verbose)
ind_peak = zrange.index(0) # approx_distance_to_next_disc
else:
# if I_corr contains only zeros
sct.printv(
'.. WARNING: Correlation vector only contains zeros. Using adjusted template distance.',
verbose)
ind_peak = zrange.index(0) # approx_distance_to_next_disc
# display patterns and correlation
if verbose == 2:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# display template pattern
plt.figure(11, figsize=(15, 7))
plt.subplot(131)
plt.imshow(np.flipud(np.mean(pattern[:, :, :], axis=0).transpose()),
origin='upper',
cmap=plt.cm.gray,
interpolation='none')
plt.title('Template pattern')
# display subject pattern at best z
plt.subplot(132)
iz = zrange[ind_peak]
data_chunk3d = src[x - xsize:x + xsize + 1,
y + yshift - ysize:y + yshift + ysize + 1,
z + iz - zsize:z + iz + zsize + 1]
plt.imshow(np.flipud(
np.mean(data_chunk3d[:, :, :], axis=0).transpose()),
origin='upper',
cmap=plt.cm.gray,
clim=[0, 800],
interpolation='none')
plt.title('Subject at iz=' + str(iz))
# display correlation curve
plt.subplot(133)
plt.plot(zrange, I_corr)
plt.plot(zrange, I_corr_gauss, 'black', linestyle='dashed')
plt.legend(['I_corr', 'I_corr_gauss'])
plt.title('Mutual Info, gaussian_std=' + str(gaussian_std))
plt.plot(zrange[ind_peak], I_corr_gauss[ind_peak], 'ro'), plt.draw()
plt.axvline(x=zrange.index(0),
linewidth=1,
color='black',
linestyle='dashed')
plt.axhline(y=thr_corr, linewidth=1, color='r', linestyle='dashed')
plt.grid()
# save figure
plt.figure(11), plt.savefig(
os.path.join(path_output,
"fig_pattern" + save_suffix + '.png')), plt.close()
# return z-origin (z) + z-displacement minus zshift (to account for non-centered disc)
return z + zrange[ind_peak] - zshift