def segment_glomeruli2d(input_file, tissue_mask_file, output_file, voxel_xy):
kmask = io.imread(tissue_mask_file)
if kmask.max() == 0:
tifffile.imsave(output_file, kmask, compress=5)
return
# normalize image
img = io.imread(input_file)
img = ndimage.median_filter(img, 3)
img = img * 255. / img.max()
# remove all intensity variations larger than maximum radius of a glomerulus
d = mahotas.disk(int(float(glomeruli_maxrad) / voxel_xy))
img = img - mahotas.open(img.astype(np.uint8), d)
img = img * 255. / img.max()
ch = img[np.where(kmask > 0)]
# segment glomeruli by otsu thresholding only if this threshold is higher than the 75-th percentile in the kidney mask
t = mahotas.otsu(img.astype(np.uint8))
cells = None
if t > np.percentile(ch, 75) * 1.5:
cells = img > t
cells[np.where(kmask == 0)] = 0
cells = mahotas.open(
cells, mahotas.disk(int(float(glomeruli_minrad) / 2. / voxel_xy)))
else:
cells = np.zeros_like(img)
tifffile.imsave(output_file, img_as_ubyte(cells), compress=5)
def watershedSegment(image, diskSize=20):
gradmag = gradientMagnitudue(image)
## compute foreground markers
# open image to create flat regions at cell centers
se_disk = pymorph.sedisk(diskSize)
image_opened = mahotas.open(image, se_disk);
# define foreground markers as regional maxes of cells
# this step is slow!
foreground_markers = mahotas.regmax(image_opened)
## compute background markers
# Threshold the image, cast it to the right datatype, and then calculate the distance image
image_black_white = image_opened > mahotas.otsu(image_opened)
image_black_white = image_black_white.astype('uint16')
# note the inversion here- a key difference from the matlab algorithm
# matlab distance is to nearest non-zero pixel
# python distance is to nearest 0 pixel
image_distance = pymorph.to_uint16(nd.distance_transform_edt(np.logical_not(image_black_white)))
eight_conn = pymorph.sebox()
distance_markers = mahotas.label(mahotas.regmin(image_distance, eight_conn))[0]
image_dist_wshed, image_dist_wshed_lines =mahotas.cwatershed(image_distance, distance_markers, eight_conn, return_lines=True)
background_markers = image_distance_watershed_lines - image_black_white
all_markers = np.logical_or(foreground_markers, background_markers)
# impose a min on the gradient image. assumes int64
gradmag2 = imimposemin(gradmag.astype(int), all_markers, eight_conn)
# call watershed
segmented_cells, segmented_cell_lines = mahotas.cwatershed(gradmag2, mahotas.label(all_markers)[0], eight_conn, return_lines=True)
# seperate watershed regions
segmented_cells[gradientMagnitudue(segmented_cells) > 0] = 0
return segmented_cells > 0, segmented_cells
#
# It is made available under the MIT License
import numpy as np
import mahotas as mh
# Load our example image:
image = mh.imread('../SimpleImageDataset/building05.jpg')
# Convert to greyscale
image = mh.colors.rgb2gray(image, dtype=np.uint8)
# Compute a threshold value:
thresh = mh.thresholding.otsu(image)
print('Otsu threshold is {0}'.format(thresh))
# Compute the thresholded image
otsubin = (image > thresh)
print('Saving thresholded image (with Otsu threshold) to otsu-threshold.jpeg')
mh.imsave('otsu-threshold.jpeg', otsubin.astype(np.uint8) * 255)
# Execute morphological opening to smooth out the edges
otsubin = mh.open(otsubin, np.ones((15, 15)))
mh.imsave('otsu-closed.jpeg', otsubin.astype(np.uint8) * 255)
# An alternative thresholding method:
thresh = mh.thresholding.rc(image)
print('Ridley-Calvard threshold is {0}'.format(thresh))
print('Saving thresholded image (with Ridley-Calvard threshold) to rc-threshold.jpeg')
mh.imsave('rc-threshold.jpeg', (image > thresh).astype(np.uint8) * 255)
def extract_polygons(self, y_offset, x_offset):
'''Creates a polygon representation for each segmented object.
The coordinates of the polygon contours are relative to the global map,
i.e. an offset is added to the :class:`Site <tmlib.models.site.Site>`.
Parameters
----------
y_offset: int
global vertical offset that needs to be subtracted from
*y*-coordinates (*y*-axis is inverted)
x_offset: int
global horizontal offset that needs to be added to *x*-coordinates
Returns
-------
Generator[Tuple[Union[int, shapely.geometry.polygon.Polygon]]]
label and geometry for each segmented object
'''
bboxes = mh.labeled.bbox(self.array)
# We set border pixels to zero to get closed contours for
# border objects. This may cause problems for very small objects
# at the border, because they may get lost.
# We recreate them later on (see below).
plane = self.array.copy()
plane[0, :] = 0
plane[-1, :] = 0
plane[:, 0] = 0
plane[:, -1] = 0
for label in np.unique(plane[plane > 0]):
bbox = bboxes[label]
obj_im = self._get_bbox_image(plane, bbox)
logger.debug('find contour for object #%d', label)
# We could do this for all objects at once, but doing it on the
# bounding box for each object individually ensures that we get the
# correct number of objects and that polygons are in the
# correct order, i.e. sorted according to their corresponding label.
mask = obj_im == label
if np.sum(mask > 0) > 1:
# We need to remove single pixel extensions on the border of
# objects because they can lead to polygon self-intersections.
# However, this should only be done if the object is larger
# than 1 pixel.
mask = mh.open(mask)
# NOTE: OpenCV returns x, y coordinates. This means one would need
# to flip the axis for numpy-based indexing (y,x coordinates).
_, contours, hierarchy = cv2.findContours(
(mask).astype(np.uint8) * 255,
cv2.RETR_CCOMP, # two-level hierarchy (holes)
cv2.CHAIN_APPROX_NONE)
if len(contours) == 0:
logger.warn('no contours identified for object #%d', label)
# This is most likely an object that does not extend
# beyond the line of border pixels.
# To ensure a correct number of objects we represent
# it by the smallest possible valid polygon.
coords = np.array(np.where(plane == label)).T
y, x = np.mean(coords, axis=0).astype(int)
shell = np.array([[x - 1, x + 1, x + 1, x - 1, x - 1],
[y - 1, y - 1, y + 1, y + 1, y - 1]]).T
holes = None
elif len(contours) > 1:
# It may happens that more than one contour is
# identified per object, for example if the object
# has holes, i.e. enclosed background pixels.
logger.debug('%d contours identified for object #%d',
len(contours), label)
holes = list()
for i in range(len(contours)):
child_idx = hierarchy[0][i][2]
parent_idx = hierarchy[0][i][3]
# There should only be two levels with one
# contour each.
if parent_idx >= 0:
shell = np.squeeze(contours[parent_idx])
elif child_idx >= 0:
holes.append(np.squeeze(contours[child_idx]))
else:
# Same hierarchy level. This shouldn't happen.
# Take only the largest one.
lengths = [len(c) for c in contours]
idx = lengths.index(np.max(lengths))
shell = np.squeeze(contours[idx])
break
else:
shell = np.squeeze(contours[0])
holes = None
if shell.ndim < 2 or shell.shape[0] < 3:
logger.warn('polygon doesn\'t have enough coordinates')
# In case the contour cannot be represented as a
# valid polygon we create a little square to not loose
# the object.
y, x = np.array(mask.shape) / 2
# Create a closed ring with coordinates sorted
# counter-clockwise
shell = np.array([[x - 1, x + 1, x + 1, x - 1, x - 1],
[y - 1, y - 1, y + 1, y + 1, y - 1]]).T
# Add offset required due to alignment and cropping and
# invert the y-axis as required by Openlayers.
add_y = y_offset + bbox[0] - 1
add_x = x_offset + bbox[2] - 1
shell[:, 0] = shell[:, 0] + add_x
shell[:, 1] = -1 * (shell[:, 1] + add_y)
if holes is not None:
for i in range(len(holes)):
holes[i][:, 0] = holes[i][:, 0] + add_x
holes[i][:, 1] = -1 * (holes[i][:, 1] + add_y)
poly = shapely.geometry.Polygon(shell, holes)
if not poly.is_valid:
logger.warn(
'invalid polygon for object #%d - trying to fix it', label)
# In some cases there may be invalid intersections
# that can be fixed with the buffer trick.
poly = poly.buffer(0)
if not poly.is_valid:
raise ValueError('Polygon of object #%d is invalid.' %
label)
if isinstance(poly, shapely.geometry.MultiPolygon):
logger.warn(
'object #%d has multiple polygons - '
'take largest', label)
# Repair may create multiple polygons.
# We take the largest and discard the smaller ones.
areas = [g.area for g in poly.geoms]
index = areas.index(np.max(areas))
poly = poly.geoms[index]
yield (int(label), poly)
def segment_layer(filename, params):
'''
Segment one layer in a stack
'''
start = time.time()
#extract pixel size in xy and z
xsize, zsize = extract_zoom(params.folder)
#load image
img = tifffile.imread(params.inputfolder + params.folder + filename)
#normalize image
img = ndimage.median_filter(img, 3)
img = img * 255. / img.max()
##segment kidney tissue
sizefactor = 10.
small = ndimage.interpolation.zoom(
img, 1. / sizefactor) #scale the image to a smaller size
imgf = ndimage.gaussian_filter(small, 3. / xsize) #Gaussian filter
median = np.percentile(imgf, 40) #40-th percentile for thresholding
kmask = imgf > median * 1.5 #thresholding
kmask = mahotas.dilate(kmask, mahotas.disk(5))
kmask = mahotas.close_holes(kmask) #closing holes
kmask = mahotas.erode(kmask, mahotas.disk(5)) * 255
#remove objects that are darker than 2*percentile
l, n = ndimage.label(kmask)
llist = np.unique(l)
if len(llist) > 2:
means = ndimage.mean(imgf, l, llist)
bv = llist[np.where(means < median * 2)]
ix = np.in1d(l.ravel(), bv).reshape(l.shape)
kmask[ix] = 0
kmask = ndimage.interpolation.zoom(kmask,
sizefactor) #scale back to normal size
kmask = normalize(kmask)
kmask = (kmask > mahotas.otsu(kmask.astype(
np.uint8))) * 255. #remove artifacts of interpolation
#save indices of the kidney mask
ind = np.where(kmask > 0)
ind = np.array(ind)
np.save(
params.inputfolder + '../segmented/masks/kidney/' + params.folder +
filename[:-4] + '.npy', ind)
#segment glomeruli, if there is a kidney tissue
if kmask.max() > 0:
#remove all intensity variations larger than maximum radius of a glomerulus
d = mahotas.disk(int(float(params.maxrad) / xsize))
img = img - mahotas.open(img.astype(np.uint8), d)
img = img * 255. / img.max()
ch = img[np.where(kmask > 0)]
#segment glomeruli by otsu thresholding only if this threshold is higher than the 75-th percentile in the kidney mask
t = mahotas.otsu(img.astype(np.uint8))
if t > np.percentile(ch, 75) * 1.5:
cells = img > t
cells[np.where(kmask == 0)] = 0
cells = mahotas.open(
cells, mahotas.disk(int(float(params.minrad) / 2. / xsize)))
else:
cells = np.zeros_like(img)
else:
cells = np.zeros_like(img)
#save indices of the glomeruli mask
ind = np.where(cells > 0)
ind = np.array(ind)
np.save(
params.inputfolder + '../segmented/masks/glomeruli/' + params.folder +
filename[:-4] + '.npy', ind)
def watershedSegment(image, diskSize=20):
"""This routine implements the watershed example from
http://www.mathworks.com/help/images/examples/marker-controlled-watershed-segmentation.html,
but using pymorph and mahotas.
:param image: an image (2d numpy array) to be segemented
:param diskSize: an integer used as a size for a structuring element used
for morphological preprocessing.
:returns: tuple of binarized and labeled segmention masks
"""
def gradientMagnitudue(image):
sobel_x = nd.sobel(image.astype('double'), 0)
sobel_y = nd.sobel(image.astype('double'), 1)
return np.sqrt((sobel_x * sobel_x) + (sobel_y * sobel_y))
def imimposemin(image, mask, connectivity):
fm = image.copy()
fm[mask] = -9223372036854775800
fm[np.logical_not(mask)] = 9223372036854775800
fp1 = image + 1
g = np.minimum(fp1, fm)
j = infrec(fm, g)
return j
def infrec(f, g, Bc=None):
if Bc is None: Bc = pymorph.secross()
n = f.size
return fast_conditional_dilate(f, g, Bc, n)
def fast_conditional_dilate(f, g, Bc=None, n=1):
if Bc is None:
Bc = pymorph.secross()
f = pymorph.intersec(f, g)
for i in xrange(n):
prev = f
f = pymorph.intersec(mahotas.dilate(f, Bc), g)
if pymorph.isequal(f, prev):
break
return f
gradmag = gradientMagnitudue(image)
## compute foreground markers
# open image to create flat regions at cell centers
se_disk = pymorph.sedisk(diskSize)
image_opened = mahotas.open(image, se_disk)
# define foreground markers as regional maxes of cells
# this step is slow!
foreground_markers = mahotas.regmax(image_opened)
## compute background markers
# Threshold the image, cast it to the right datatype, and then calculate the distance image
image_black_white = image_opened > mahotas.otsu(image_opened)
image_black_white = image_black_white.astype('uint16')
# note the inversion here- a key difference from the matlab algorithm
# matlab distance is to nearest non-zero pixel
# python distance is to nearest 0 pixel
image_distance = pymorph.to_uint16(
nd.distance_transform_edt(np.logical_not(image_black_white)))
eight_conn = pymorph.sebox()
distance_markers = mahotas.label(mahotas.regmin(image_distance,
eight_conn))[0]
image_dist_wshed, image_dist_wshed_lines = mahotas.cwatershed(
image_distance, distance_markers, eight_conn, return_lines=True)
background_markers = image_dist_wshed_lines - image_black_white
all_markers = np.logical_or(foreground_markers, background_markers)
# impose a min on the gradient image. assumes int64
gradmag2 = imimposemin(gradmag.astype(int), all_markers, eight_conn)
# call watershed
segmented_cells, segmented_cell_lines = mahotas.cwatershed(
gradmag2, mahotas.label(all_markers)[0], eight_conn, return_lines=True)
segmented_cells -= 1
# seperate watershed regions
segmented_cells[gradientMagnitudue(segmented_cells) > 0] = 0
return segmented_cells > 0, segmented_cells
def watershedSegment(image, diskSize=20):
def gradientMagnitudue(image):
sobel_x = nd.sobel(image.astype('double'), 0)
sobel_y = nd.sobel(image.astype('double'), 1)
return np.sqrt((sobel_x * sobel_x) + (sobel_y * sobel_y))
def imimposemin(image, mask, connectivity):
fm = image.copy()
fm[mask] = -9223372036854775800
fm[np.logical_not(mask)] = 9223372036854775800
fp1 = image + 1
g = np.minimum(fp1, fm)
j = infrec(fm, g)
return j
def infrec(f, g, Bc=None):
if Bc is None: Bc = pymorph.secross()
n = f.size
return fast_conditional_dilate(f, g, Bc, n);
def fast_conditional_dilate(f, g, Bc=None, n=1):
if Bc is None:
Bc = pymorph.secross()
f = pymorph.intersec(f,g)
for i in xrange(n):
prev = f
f = pymorph.intersec(mahotas.dilate(f, Bc), g)
if pymorph.isequal(f, prev):
break
return f
gradmag = gradientMagnitudue(image)
## compute foreground markers
# open image to create flat regions at cell centers
se_disk = pymorph.sedisk(diskSize)
image_opened = mahotas.open(image, se_disk);
# define foreground markers as regional maxes of cells
# this step is slow!
foreground_markers = mahotas.regmax(image_opened)
## compute background markers
# Threshold the image, cast it to the right datatype, and then calculate the distance image
image_black_white = image_opened > mahotas.otsu(image_opened)
image_black_white = image_black_white.astype('uint16')
# note the inversion here- a key difference from the matlab algorithm
# matlab distance is to nearest non-zero pixel
# python distance is to nearest 0 pixel
image_distance = pymorph.to_uint16(nd.distance_transform_edt(np.logical_not(image_black_white)))
eight_conn = pymorph.sebox()
distance_markers = mahotas.label(mahotas.regmin(image_distance, eight_conn))[0]
image_dist_wshed, image_dist_wshed_lines = mahotas.cwatershed(image_distance, distance_markers, eight_conn, return_lines=True)
background_markers = image_dist_wshed_lines - image_black_white
all_markers = np.logical_or(foreground_markers, background_markers)
# impose a min on the gradient image. assumes int64
gradmag2 = imimposemin(gradmag.astype(int), all_markers, eight_conn)
# call watershed
segmented_cells, segmented_cell_lines = mahotas.cwatershed(gradmag2, mahotas.label(all_markers)[0], eight_conn, return_lines=True)
segmented_cells -= 1
# seperate watershed regions
segmented_cells[gradientMagnitudue(segmented_cells) > 0] = 0
return segmented_cells > 0, segmented_cells
im16 = mh.gaussian_filter(image, 16)
thresh = mh.thresholding.otsu(im16.astype(np.uint8))
threshed = (im16 > thresh)
plt.figure()
plt.imshow(threshed)
plt.title('threholded image (after blurring)')
print('Otsu threshold after blurring is {}.'.format(thresh))
mh.imsave('thresholded16.png', threshed.astype(np.uint8) * 255)
plt.show()
image = mh.imread('SimpleImageDataset/building05.jpg')
image = mh.colors.rgb2grey(image, dtype=np.uint8)
th = mh.thresholding.otsu(image)
print('Otsu threshold is {0}'.format(thresh))
otsubin = (image > thresh)
print('Saving thresholded image (with Otsu threshold) to otsu-threshold.jpeg')
mh.imsave('otsu-threshold.jpeg', otsubin.astype(np.uint8) * 255)
otsubin = mh.open(otsubin, np.ones((15, 15)))
mh.imsave('otsu-closed.jpeg', otsubin.astype(np.uint8) * 255)
th = mh.thresholding.rc(image)
print('Ridley-Calvard threshold is {0}'.format(thresh))
print(
'Saving thresholded image (with Ridley-Calvard threshold) to rc-threshold.jpeg'
)
mh.imsave('rc-threshold.jpeg', (image > thresh).astype(np.uint8) * 255)
def save_open_image(self, threshold, name):
image = mh.open(self.get_binary_image(threshold), np.ones((15, 15)))
mh.imsave(name, image.astype(np.uint8) * 255)
def extract_polygons(self, y_offset, x_offset):
'''Creates a polygon representation for each segmented object.
The coordinates of the polygon contours are relative to the global map,
i.e. an offset is added to the :class:`Site <tmlib.models.site.Site>`.
Parameters
----------
y_offset: int
global vertical offset that needs to be subtracted from
*y*-coordinates (*y*-axis is inverted)
x_offset: int
global horizontal offset that needs to be added to *x*-coordinates
Returns
-------
Generator[Tuple[Union[int, shapely.geometry.polygon.Polygon]]]
label and geometry for each segmented object
'''
bboxes = mh.labeled.bbox(self.array)
# We set border pixels to zero to get closed contours for
# border objects. This may cause problems for very small objects
# at the border, because they may get lost.
# We recreate them later on (see below).
plane = self.array.copy()
plane[0, :] = 0
plane[-1, :] = 0
plane[:, 0] = 0
plane[:, -1] = 0
for label in np.unique(plane[plane > 0]):
bbox = bboxes[label]
obj_im = self._get_bbox_image(plane, bbox)
logger.debug('find contour for object #%d', label)
# We could do this for all objects at once, but doing it on the
# bounding box for each object individually ensures that we get the
# correct number of objects and that polygons are in the
# correct order, i.e. sorted according to their corresponding label.
mask = obj_im == label
if np.sum(mask > 0) > 1:
# We need to remove single pixel extensions on the border of
# objects because they can lead to polygon self-intersections.
# However, this should only be done if the object is larger
# than 1 pixel.
mask = mh.open(mask)
# NOTE: OpenCV returns x, y coordinates. This means one would need
# to flip the axis for numpy-based indexing (y,x coordinates).
_, contours, hierarchy = cv2.findContours(
(mask).astype(np.uint8) * 255,
cv2.RETR_CCOMP, # two-level hierarchy (holes)
cv2.CHAIN_APPROX_NONE
)
if len(contours) == 0:
logger.warn('no contours identified for object #%d', label)
# This is most likely an object that does not extend
# beyond the line of border pixels.
# To ensure a correct number of objects we represent
# it by the smallest possible valid polygon.
coords = np.array(np.where(plane == label)).T
y, x = np.mean(coords, axis=0).astype(int)
shell = np.array([
[x-1, x+1, x+1, x-1, x-1],
[y-1, y-1, y+1, y+1, y-1]
]).T
holes = None
elif len(contours) > 1:
# It may happens that more than one contour is
# identified per object, for example if the object
# has holes, i.e. enclosed background pixels.
logger.debug(
'%d contours identified for object #%d',
len(contours), label
)
holes = list()
for i in range(len(contours)):
child_idx = hierarchy[0][i][2]
parent_idx = hierarchy[0][i][3]
# There should only be two levels with one
# contour each.
if parent_idx >= 0:
shell = np.squeeze(contours[parent_idx])
elif child_idx >= 0:
holes.append(np.squeeze(contours[child_idx]))
else:
# Same hierarchy level. This shouldn't happen.
# Take only the largest one.
lengths = [len(c) for c in contours]
idx = lengths.index(np.max(lengths))
shell = np.squeeze(contours[idx])
break
else:
shell = np.squeeze(contours[0])
holes = None
if shell.ndim < 2 or shell.shape[0] < 3:
logger.warn('polygon doesn\'t have enough coordinates')
# In case the contour cannot be represented as a
# valid polygon we create a little square to not loose
# the object.
y, x = np.array(mask.shape) / 2
# Create a closed ring with coordinates sorted
# counter-clockwise
shell = np.array([
[x-1, x+1, x+1, x-1, x-1],
[y-1, y-1, y+1, y+1, y-1]
]).T
# Add offset required due to alignment and cropping and
# invert the y-axis as required by Openlayers.
add_y = y_offset + bbox[0] - 1
add_x = x_offset + bbox[2] - 1
shell[:, 0] = shell[:, 0] + add_x
shell[:, 1] = -1 * (shell[:, 1] + add_y)
if holes is not None:
for i in range(len(holes)):
holes[i][:, 0] = holes[i][:, 0] + add_x
holes[i][:, 1] = -1 * (holes[i][:, 1] + add_y)
poly = shapely.geometry.Polygon(shell, holes)
if not poly.is_valid:
logger.warn(
'invalid polygon for object #%d - trying to fix it',
label
)
# In some cases there may be invalid intersections
# that can be fixed with the buffer trick.
poly = poly.buffer(0)
if not poly.is_valid:
raise ValueError(
'Polygon of object #%d is invalid.' % label
)
if isinstance(poly, shapely.geometry.MultiPolygon):
logger.warn(
'object #%d has multiple polygons - '
'take largest', label
)
# Repair may create multiple polygons.
# We take the largest and discard the smaller ones.
areas = [g.area for g in poly.geoms]
index = areas.index(np.max(areas))
poly = poly.geoms[index]
yield (int(label), poly)
