def test_kmeans():
x = np.concatenate((np.ones(10) * 10, np.ones(10) * 50))
centers, labels, maps = kmeans(x,
k=2,
scale_max=100,
scale_min=0,
map_keys=['a', 'b'])
assert (len(centers) == 2)
assert (sorted(centers) == [10, 50])
assert (len(labels) == len(x))
assert (len(maps) == 2)
assert (np.allclose(maps['a'], np.concatenate(
(np.ones(10), np.zeros(10)))))
assert (np.allclose(maps['b'], np.concatenate(
(np.zeros(10), np.ones(10)))))
def init_values(data, k, scale_range, scale_sigma):
"""
Initialize theta and labels. The means of each label are initialized using
kmeans. The standard deviation of each label is chosen randomly from a uniform
distribution between 0 and scale_sigma.
Input
-----
data : array (n, m)
Image to segment.
k : int
Number of labels.
scale_range : tuple
(min, max) Range of values from which start kmeans.
scale_sigma : float
Sigma will be initialized with a random value between 1 and scale_sigma.
Output
------
thetas : list of lists
Contains [mu, sigma] for each label.
labels : int array (n, m)
Labels for each pixel. The labels are assigned using kmeans (distance of
each pixel's intensity to label means).
"""
scale_min, scale_max = scale_range
mus, labels, _ = kmeans(data.ravel(), k, scale_max=scale_max, scale_min=scale_min)
labels = labels.reshape(data.shape)
thetas = []
for i in range(k):
sigma = np.random.rand() * (scale_sigma - 1) + 1
thetas.append([mus[i], sigma])
return thetas, labels.astype(int)
from fmri_utils.segmentation.kmeans import kmeans
subjects = ['sub-10159'] #, 'sub-10171', 'sub-10189']
# Current directory
my_dir = dirname(__file__)
# Data directory (project-red/data)
data_dir = pjoin(my_dir, '..', '..', '..', 'data')
for s in subjects:
img = nib.load(
pjoin(data_dir, 'ds000030', s, 'anat', '%s_T1w_brain.nii.gz' % s))
data = img.get_data()
# Segment
_, klabel, kmap = kmeans(data.ravel())
# Reshape things
csf = kmap['csf'].reshape(data.shape)
gm = kmap['gray'].reshape(data.shape)
wm = kmap['white'].reshape(data.shape)
# Save
nifti_csf = nib.Nifti1Image(csf, affine=img.affine)
nifti_gm = nib.Nifti1Image(gm, affine=img.affine)
nifti_wm = nib.Nifti1Image(wm, affine=img.affine)
nib.nifti1.save(
nifti_csf,
pjoin(data_dir, 'ds000030', s, 'segmentation_results',
'kmeans_csf.nii'))
nib.nifti1.save(
my_dir = dirname(__file__)
# Data directory (project-red/data)
data_dir = pjoin(my_dir, '..', '..', '..', 'data')
# Report figures directory (project-red/report/figures)
report_dir = pjoin(data_dir, '..', 'report', 'figures')
# Segmentation for subjects
for s in subjects:
img = nib.load(
pjoin(data_dir, 'ds000030', s, 'anat', '%s_T1w_brain.nii.gz' % s))
data = img.get_data()
n_slices = data.shape[-1]
slice_s = data[..., np.floor(n_slices / 2).astype(int)] # middle slice
# Segment
_, klabel, kmap = kmeans(slice_s.ravel())
# Plots
plt.figure(figsize=(15, 15))
plt.subplot(1, 4, 1)
plt.imshow(klabel.reshape(slice_s.shape))
plt.title('Segmented slice')
plt.subplot(1, 4, 2)
plt.imshow(kmap['csf'].reshape(slice_s.shape), cmap='gray')
plt.title('CSF')
plt.subplot(1, 4, 3)
plt.imshow(kmap['gray'].reshape(slice_s.shape), cmap='gray')
plt.title('Gray matter')
plt.subplot(1, 4, 4)
im = plt.imshow(kmap['white'].reshape(slice_s.shape), cmap='gray')
plt.title('White matter')
slice_s = data[x1:x2, y1:y2, slice_n]
# Segment
_, labels, maps = mrf_em(slice_s,
0.05,
k=2,
max_iter=10,
scale_range=(100, 400),
scale_sigma=50,
max_label_iter=10,
njobs=2,
map_labels=['gray', 'white'])
# Run kmeans for comparison
_, klabels, kmaps = kmeans(slice_s.ravel(),
k=2,
max_iter=10 ^ 4,
scale_max=400,
scale_min=200,
map_keys=['gray', 'white'])
klabels = klabels.reshape(slice_s.shape)
# Plot segmentation
fig1 = plt.figure(figsize=(10, 10))
plt.subplot(1, 3, 1)
plt.imshow(labels, interpolation="None")
plt.title('MRF EM')
plt.subplot(1, 3, 2)
plt.imshow(slice_s, interpolation="None")
plt.title('Original')
plt.subplot(1, 3, 3)
plt.imshow(np.abs(klabels - 1), interpolation="None")
plt.title('kmeans')