def cvxhull_area(self):
'''
Calculate the convexhull area such that:
A=0.5*Sumfrom 0 to N-1 of {xn+1*yn-xn*yn+1}
Pn(xn,yn) and PN=P0
see http://en.wikipedia.org/wiki/Polygon#Area_and_centroid
'''
#print "cvxhull called"
binIm=self.particuleImage>0
area=np.sum(binIm[:,:]==True)
print "area",area
#print binIm.dtype
contour=mahotas.bwperim(binIm)
#print "contour",contour.dtype
pointlist=mahotas.polygon.convexhull(contour)
N=len(pointlist)
fP=pointlist[0]
#duplicate the first point P0
#at the end such that PointN=Point0
pointlist.append(fP)
s=0
#compute the sum from 0 to N-1
for i in range(0,N-1):
cx=pointlist[i][0]#x of the current point
cy=pointlist[i][1]#y of the current point
#print "Point",i," x=",cx," y=",cy
nx=pointlist[i+1][0]#x of the next point
ny=pointlist[i+1][1]#y of the next point
#print "Point suiv",i+1," x=",nx," y=",ny
det=nx*cy-cx*ny
s=s+det
#print "det:",det," S=",s
CvxhParticleArea_ratio=area/(0.5*abs(s))
return 0.5*abs(s),CvxhParticleArea_ratio
def _hull_computations(imageproc,imagehull = None):
# Just share code between the two functions below
if imagehull is None:
imagehull = convexhull(imageproc > 0)
Ahull = _bwarea(imagehull)
Phull = _bwarea(bwperim(imagehull))
cofy,cofx = center_of_mass(imagehull)
hull_mu00 = imgcentmoments(imagehull,0,0,cofy,cofx)
hull_mu11 = imgcentmoments(imagehull,1,1,cofy,cofx)
hull_mu02 = imgcentmoments(imagehull,0,2,cofy,cofx)
hull_mu20 = imgcentmoments(imagehull,2,0,cofy,cofx)
# Parameters of the 'image ellipse'
# (the constant intensity ellipse with the same mass and
# second order moments as the original image.)
# From Prokop, RJ, and Reeves, AP. 1992. CVGIP: Graphical
# Models and Image Processing 54(5):438-460
hull_semimajor = sqrt((2 * (hull_mu20 + hull_mu02 + \
sqrt((hull_mu20 - hull_mu02)**2 + \
4 * hull_mu11**2)))/hull_mu00)
hull_semiminor = sqrt((2 * (hull_mu20 + hull_mu02 - \
sqrt((hull_mu20 - hull_mu02)**2 + \
4 * hull_mu11**2)))/hull_mu00)
return imagehull,Ahull, Phull, hull_semimajor, hull_semiminor
def dna_size_shape(labeled, scale=1., minarea=float('-inf'), maxarea=float('+inf'), minroundness=-1.):
'''
positives = dna_size_shape(dnamasks, scale=1., minarea=-Inf, maxarea=+Inf, minroundness=0)
Only accepts DNA objects that fulfill the following criterion:
* are greater than minarea
* are smaller than maxarea
* are rounder than minroundness
'''
nr_objects = labeled .max()
positives = np.zeros(nr_objects+1, np.bool)
minarea /= scale
maxarea /= scale
for obj in xrange(1,nr_objects+1):
objimg = croptobbox(labeled == obj)
area = objimg.sum()
if area > maxarea or area < minarea:
continue
hull = convexhull(objimg)
hullArea = hull.sum()
hullPerim = bwperim(hull).sum()
roundness = hullPerim**2/(4*np.pi*hullArea)
if roundness < minroundness:
continue
positives[obj] = 1
return positives
def phase_stretch_transform(img, LPF, S, W, threshold_min, threshold_max, flag):
L = 0.5
x = np.linspace(-L, L, img.shape[0])
y = np.linspace(-L, L, img.shape[1])
[X1, Y1] = (np.meshgrid(x, y))
X = X1.T
Y = Y1.T
theta, rho = cart2pol(X, Y)
orig = ((np.fft.fft2(img)))
expo = np.fft.fftshift(np.exp(-np.power((np.divide(rho, math.sqrt((LPF ** 2) / np.log(2)))), 2)))
orig_filtered = np.real(np.fft.ifft2((np.multiply(orig, expo))))
PST_Kernel_1 = np.multiply(np.dot(rho, W), np.arctan(np.dot(rho, W))) - 0.5 * np.log(
1 + np.power(np.dot(rho, W), 2))
PST_Kernel = PST_Kernel_1 / np.max(PST_Kernel_1) * S
temp = np.multiply(np.fft.fftshift(np.exp(-1j * PST_Kernel)), np.fft.fft2(orig_filtered))
orig_filtered_PST = np.fft.ifft2(temp)
PHI_features = np.angle(orig_filtered_PST)
if flag == 0:
out = PHI_features
else:
features = np.zeros((PHI_features.shape[0], PHI_features.shape[1]))
features[PHI_features > threshold_max] = 1
features[PHI_features < threshold_min] = 1
features[img < (np.amax(img) / 20)] = 0
out = features
out = mh.thin(out, 1)
out = mh.bwperim(out, 4)
out = mh.thin(out, 1)
out = mh.erode(out, np.ones((1, 1)))
return out, PST_Kernel
def find_perimeter(bim):
pim = mahotas.bwperim(bim)
pts = numpy.array(numpy.where(pim))
if len(pts):
return pts.T[:, ::-1]
else:
return []
def cvxhull_area(self):
'''
Calculate the convexhull area such that:
A=0.5*Sumfrom 0 to N-1 of {xn+1*yn-xn*yn+1}
Pn(xn,yn) and PN=P0
see http://en.wikipedia.org/wiki/Polygon#Area_and_centroid
'''
#print "cvxhull called"
binIm = self.particuleImage > 0
area = np.sum(binIm[:, :] == True)
print "area", area
#print binIm.dtype
contour = mahotas.bwperim(binIm)
#print "contour",contour.dtype
pointlist = mahotas.polygon.convexhull(contour)
N = len(pointlist)
fP = pointlist[0]
#duplicate the first point P0
#at the end such that PointN=Point0
pointlist.append(fP)
s = 0
#compute the sum from 0 to N-1
for i in range(0, N - 1):
cx = pointlist[i][0] #x of the current point
cy = pointlist[i][1] #y of the current point
#print "Point",i," x=",cx," y=",cy
nx = pointlist[i + 1][0] #x of the next point
ny = pointlist[i + 1][1] #y of the next point
#print "Point suiv",i+1," x=",nx," y=",ny
det = nx * cy - cx * ny
s = s + det
#print "det:",det," S=",s
CvxhParticleArea_ratio = area / (0.5 * abs(s))
return 0.5 * abs(s), CvxhParticleArea_ratio
def pred_to_mask(pred, wsi=None, perim=False):
"""
given a prediction logit
of size [# classes, width, height]
& corresponding wsi image
return the mask embedded onto
this wsi (as np array)
"""
str_elem = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (10, 10))
pred = threshold_probs(pred)
'save image'
pred = 255 * (np.eye(args.num_classes)[pred][..., 1:]).astype(np.uint8)
wsi = np.zeros_like(pred) if wsi is None else np.array(wsi.copy())
for cj in range(args.num_classes - 1):
rgbcolor = [0, 0, 0]
rgbcolor[cj] = 255
if perim:
pred[..., cj] = bwperim(pred[..., cj])
pred[..., cj] = cv2.dilate(pred[..., cj], str_elem, iterations=1)
wsi[pred[..., cj] > 0, :] = rgbcolor
del pred
return wsi
def extract(self):
'''Extracts neighbour features.
Returns
-------
pandas.DataFrame
extracted feature values for each object in `label_image`
'''
# Create an empty dataset in case no objects were detected
logger.info('extract neighbour features')
features = list()
for obj in self.object_ids:
pad = max(self.neighbour_distance, self.touching_distance)
object_image = self._get_bbox_containing_neighbours(obj, pad)
# dilate the current object
object_image_dilate = mh.dilate(object_image == obj,
Bc=mh.disk(
self.neighbour_distance))
# mask the corresponding region of the label image
object_image_mask = np.copy(object_image)
object_image_mask[object_image_dilate == 0] = 0
object_image_mask[object_image == obj] = 0
neighbour_ids = np.unique(object_image_mask)
unique_values = neighbour_ids[np.nonzero(neighbour_ids)].tolist()
neighbour_count = len(unique_values)
# save these unique values as a string
if neighbour_count == 0:
neighbour_string = '.'
else:
neighbour_string = '.'.join(str(x) for x in unique_values)
# create an inverted image of the surrounding cells
neighbours = np.zeros_like(object_image)
for n in unique_values:
neighbours += mh.dilate(object_image == n)
# calculate the distance from each pixel of object to neighbours
dist = ndi.morphology.distance_transform_edt(
np.invert(neighbours > 0))
# select perimeter pixels whose distance to neighbours is
# less than threshold touching distance
perimeter_image = mh.bwperim(object_image == obj)
dist[perimeter_image == 0] = 0
dist[dist > self.touching_distance] = 0
fraction_touching = np.count_nonzero(dist) / float(
np.count_nonzero(perimeter_image))
values = [neighbour_count, neighbour_string, fraction_touching]
features.append(values)
return pd.DataFrame(features,
columns=self.names,
index=self.object_ids)
def pst_algorithm(image, LPF, Phase_strength, Warp_strength, Threshold_min,
Threshold_max, Morph_flag):
L = 0.5
x = np.linspace(-L, L, image.shape[0])
y = np.linspace(-L, L, image.shape[1])
[X1, Y1] = (np.meshgrid(x, y))
X = X1.T
Y = Y1.T
[THETA, RHO] = [np.arctan2(Y, X),
np.hypot(X, Y)] # cartesian to polar coordinates
# Apply localization kernel to the original image to reduce noise
Image_orig_f = np.fft.fft2(image)
expo = np.fft.fftshift(
np.exp(-np.power((np.divide(RHO, math.sqrt((LPF**2) /
np.log(2)))), 2)))
Image_orig_filtered = np.real(
np.fft.ifft2((np.multiply(Image_orig_f, expo))))
# Constructing the PST Kernel
PST_Kernel_1 = np.multiply(
np.dot(RHO, Warp_strength), np.arctan(np.dot(RHO, Warp_strength))
) - 0.5 * np.log(1 + np.power(np.dot(RHO, Warp_strength), 2))
PST_Kernel = PST_Kernel_1 / np.max(PST_Kernel_1) * Phase_strength
# Apply the PST Kernel
temp = np.multiply(np.fft.fftshift(np.exp(-1j * PST_Kernel)),
np.fft.fft2(Image_orig_filtered))
Image_orig_filtered_PST = np.fft.ifft2(temp)
# Calculate phase of the transformed image
PHI_features = np.angle(Image_orig_filtered_PST)
if Morph_flag == 0:
out = PHI_features
out = (out / np.max(out)) * 3
else:
# find image sharp transitions by thresholding the phase
features = np.zeros((PHI_features.shape[0], PHI_features.shape[1]))
features[PHI_features > Threshold_max] = 1 # Bi-threshold decision
features[
PHI_features <
Threshold_min] = 1 # as the output phase has both positive and negative values
features[image < (
np.amax(image) / 20
)] = 0 # Removing edges in the very dark areas of the image (noise)
# apply binary morphological operations to clean the transformed image
out = features
out = mh.thin(out, 1)
out = mh.bwperim(out, 4)
out = mh.thin(out, 1)
out = mh.erode(out, np.ones((1, 1)))
return out
def phase_stretch_transform(img, LPF, S, W, Threshold_min, Threshold_max,
flag):
L = 0.5
x = np.linspace(-L, L, img.shape[0])
y = np.linspace(-L, L, img.shape[1])
X, Y = np.meshgrid(x, y)
p, q = X.T, y.T
theta, rho = cart2pol(p, q)
# 接下来对PST公式从右至左依次实现，
# 对输入图像进行快速傅里叶变换,
orig = np.fft.fft2(img)
# 实现L[p, q]
expo = np.fft.fftshift(
np.exp(-np.power((np.divide(rho, math.sqrt((LPF**2) /
np.log(2)))), 2)))
# 对图像进行平滑处理，
orig_filtered = np.real(np.fft.ifft2((np.multiply(orig, expo))))
# 实现相位核，
PST_Kernel_1 = np.multiply(np.dot(rho, W), np.arctan(np.dot(
rho, W))) - 0.5 * np.log(1 + np.power(np.dot(rho, W), 2))
PST_Kernel = PST_Kernel_1 / np.max(PST_Kernel_1) * S
# 将前面实现的部分与相位核做乘积，
temp = np.multiply(np.fft.fftshift(np.exp(-1j * PST_Kernel)),
np.fft.fft2(orig_filtered))
# 对图像进行逆快速傅里叶变换，
orig_filtered_PST = np.fft.ifft2(temp)
# 进行角运算，得到变换图像的相位，
PHI_features = np.angle(orig_filtered_PST)
if flag == 0:
out = PHI_features
else:
# 对图像进行阈值化处理，
features = np.zeros((PHI_features.shape[0], PHI_features.shape[1]))
features[PHI_features > Threshold_max] = 1
features[PHI_features < Threshold_min] = 1
features[img < (np.amax(img) / 20)] = 0
# 应用二进制形态学操作来清除转换后的图像,
out = features
out = mh.thin(out, 1)
out = mh.bwperim(out, 4)
out = mh.thin(out, 1)
out = mh.erode(out, np.ones((1, 1)))
return out, PST_Kernel
def find_perimeter(labeled, seeds):
"""
return [[x,y],[x,y]...]
"""
rpts = None
for seed in seeds:
bim = to_binary(labeled, seed)
pim = mahotas.bwperim(bim)
pts = numpy.array(numpy.where(pim)).astype(numpy.float64)
if len(pts):
if rpts is None:
rpts = pts
else:
rpts = numpy.hstack((rpts, pts))
if rpts is None:
return []
else:
return rpts.T[:, ::-1] # swap dimensions to get to (x, y)
def get_key_points_for_patch(params):
'''
patches will only have image level
label (no segmentation mask);
hence no need for region
level key points. instead, generate
a roughly uniform
point set.
'''
y_max = params.dimensions[1] // 4**params.scan_level
x_max = params.dimensions[0] // 4**params.scan_level
mask = np.zeros((y_max, x_max), dtype=np.uint8)
y_min = 32
x_min = 32
mask[y_min:y_max - y_min, x_min:x_max - x_min] = 1
perim = bwperim(mask)
perim_coords = np.transpose(np.where(perim))[:, ::-1] # (x,y) pairs
skip = np.maximum(2, perim_coords.shape[0] // params.num_perim_points)
perim_coords = perim_coords[::skip, :]
kernel = np.ones((10, 10), np.uint8)
_, center_pts, _, _ = get_key_points(cv2.erode(mask, kernel, iterations=1),
1, params.num_center_points,
params.num_center_points)
center_pts -= [params.tile_w // 2, params.tile_h // 2]
perim_coords -= [params.tile_w // 2, params.tile_h // 2]
return {
'cnt_xy': center_pts,
'perim_xy': perim_coords,
'scan_level': params.scan_level
}
def detect_blobs(image, mask, threshold, min_area, deblend_nthresh=500,
deblend_cont=0):
'''Detects blobs in `image` using an implementation of
`SExtractor <http://www.astromatic.net/software/sextractor>`_ [1].
Parameters
----------
image: numpy.ndarray[Union[numpy.uint8, numpy.uint16]]
grayscale image in which blobs should be detected
mask: numpy.ndarray[numpy.bool]
binary image that masks pixel regions in which no blobs should be
detected
threshold: int, optional
factor by which pixel values must be above background
to be considered part of a blob (default: ``5``)
min_area: int, optional
minimal size of a blob
deblend_ntresh: int, optional
number of deblending thresholds (default: ``500``)
deblend_cont: int, optional
minimum contrast ratio for deblending (default: ``0``)
Returns
-------
Tuple[numpy.ndarray[numpy.int32]]
detected blobs and the corresponding centroids
References
----------
.. [1] Bertin, E. & Arnouts, S. 1996: SExtractor: Software for source
extraction, Astronomy & Astrophysics Supplement 317, 393
'''
sep.set_extract_pixstack(10**7)
img = image.astype('float')
# We pad the image with mirrored pixels to prevent border artifacts.
pad = 50
left = img[:, 1:pad]
right = img[:, -pad:-1]
detect_img = np.c_[np.fliplr(left), img, np.fliplr(right)]
upper = detect_img[1:pad, :]
lower = detect_img[-pad:-1, :]
detect_img = np.r_[np.flipud(upper), detect_img, np.flipud(lower)]
logger.info('detect blobs via thresholding and deblending')
detection, blobs = sep.extract(
detect_img, threshold,
minarea=min_area, segmentation_map=True,
deblend_nthresh=deblend_nthresh, deblend_cont=deblend_cont,
filter_kernel=None, clean=False
)
centroids = np.zeros(detect_img.shape, dtype=np.int32)
y = detection['y'].astype(int)
x = detection['x'].astype(int)
# WTF? In rare cases object coorindates lie outside of the image.
n = len(detection)
y[y > detect_img.shape[0]] = detect_img.shape[0]
x[x > detect_img.shape[1]] = detect_img.shape[1]
centroids[y, x] = np.arange(1, n + 1)
# Remove the padded border pixels
blobs = blobs[pad-1:-(pad-1), pad-1:-(pad-1)].copy()
centroids = centroids[pad-1:-(pad-1), pad-1:-(pad-1)].copy()
# Blobs detected outside of regions of interest are discarded.
blobs[mask > 0] = 0
blobs[mh.bwperim(np.invert(mask)) > 0] = 0
mh.labeled.relabel(blobs, inplace=True)
# We need to ensure that centroids are labeled the same way as blobs.
centroids[centroids > 0] = blobs[centroids > 0]
return (blobs, centroids)
def Bfx_basicint(I, R, *args):
"""
X, Xn = Bfx_basicint(I,R,options)
Toolbox: Balu
Basic intensity features
X is the features vector, Xn is the list feature names (see Example to see how it works).
Reference:
Kumar, A.; Pang, G.K.H. (2002): Defect detection in textured materials
using Gabor filters. IEEE Transactions on Industry Applications,
38(2):425-440.
Example:
import numpy as np
from mahotas.colors import rgb2gray
from balu.ImageProcessing import Bim_segbalu
from balu.ImagesAndData import balu_imageload
options = {'show': True, 'mask': 5} % Gauss mask for gradient computation and display results
I = balu_imageload(('testimg1.jpg'); % input image
R,_,_ = Bim_segbalu(I); % segmentation
X, Xn = Bfx_basicint(I,R,options) % basic intenisty features
See also Bfx_haralick, Bfx_clp, Bfx_gabor, Bfx_fourier, Bfx_dct, Bfx_lbp.
(c) D.Mery, PUC-DCC, 2010
http://dmery.ing.puc.cl
With collaboration from:
Jose Miguel Arrieta Ramos ([email protected]) -> Translated implementation into python (2017)
"""
if len(args) == 0:
options = {'show': False}
else:
options = args[0]
if options['show']:
print('--- extracting Basic intensity features...')
if 'mask' not in options:
options['mask'] = 15
E = bwperim(R, n=4)
ii = R == 1
jj = np.where(R.ravel() == 0)[0]
kk = E == 1
I = I.astype(float)
I1, _, _ = Bim_d1(I, options['mask'])
I2 = Bim_d2(I)
if len(jj) > 0:
C = np.mean(np.abs(I1[kk]))
else:
C = -1
J = I[ii]
G = np.mean(J)
S = np.std(J)
K = st.kurtosis(J, fisher=False)
Sk = st.skew(J)
D = np.mean(I2[ii])
X = np.array([[G, S, K, Sk, D, C]])
Xn = [
'Intensity Mean', 'Intensity StdDev', 'Intensity Kurtosis',
'Intensity Skewness', 'Mean Laplacian', 'Mean Boundary Gradient'
]
return X, Xn
def _compare_slow(img):
for n in (4, 8):
assert np.all(_slow_bwperim(img, n) == bwperim(img, n))
im = np.uint8(heatmap/255 >= 0.9)
im = cv2.morphologyEx(
im,
cv2.MORPH_OPEN,
kernel=np.ones((30, 30))
)
heatmap_orig = heatmap
heatmap = heatmap * im
heatmap = np.repeat(heatmap[..., np.newaxis], 3, 2)
tb_perim = cv2.morphologyEx(
bwperim(chull(im)).astype(np.uint8),
cv2.MORPH_DILATE,
kernel=np.ones((20, 20))
)
overlay = 0.65 * wsi + 0.35 * heatmap
yy, xx = np.where(tb_perim)
overlay[yy,xx,...]=0
overlay = np.uint8(overlay)
overlay_image = Image.fromarray(overlay).resize((x//4, y//4))
overlay_image.save('overlay_tumor_bed.png')
wsi = Image.fromarray(wsi).resize((x//4, y//4))
wsi.save('wsi.png')
def detect_blobs(image, mask, threshold, min_area, deblend_nthresh=500,
deblend_cont=0):
'''Detects blobs in `image` using an implementation of
`SExtractor <http://www.astromatic.net/software/sextractor>`_ [1].
Parameters
----------
image: numpy.ndarray[Union[numpy.uint8, numpy.uint16]]
grayscale image in which blobs should be detected
mask: numpy.ndarray[numpy.bool]
binary image that masks pixel regions in which no blobs should be
detected
threshold: int, optional
factor by which pixel values must be above background
to be considered part of a blob (default: ``5``)
min_area: int, optional
minimal size of a blob
deblend_ntresh: int, optional
number of deblending thresholds (default: ``500``)
deblend_cont: int, optional
minimum contrast ratio for deblending (default: ``0``)
Returns
-------
Tuple[numpy.ndarray[numpy.int32]]
detected blobs and the corresponding centroids
References
----------
.. [1] Bertin, E. & Arnouts, S. 1996: SExtractor: Software for source
extraction, Astronomy & Astrophysics Supplement 317, 393
'''
sep.set_extract_pixstack(10**7)
img = image.astype('float')
# We pad the image with mirrored pixels to prevent border artifacts.
pad = 50
left = img[:, 1:pad]
right = img[:, -pad:-1]
detect_img = np.c_[np.fliplr(left), img, np.fliplr(right)]
upper = detect_img[1:pad, :]
lower = detect_img[-pad:-1, :]
detect_img = np.r_[np.flipud(upper), detect_img, np.flipud(lower)]
logger.info('detect blobs via thresholding and deblending')
detection, blobs = sep.extract(
detect_img, threshold,
minarea=min_area, segmentation_map=True,
deblend_nthresh=deblend_nthresh, deblend_cont=deblend_cont,
filter_kernel=None, clean=False
)
centroids = np.zeros(detect_img.shape, dtype=np.int32)
y = detection['y'].astype(int)
x = detection['x'].astype(int)
# WTF? In rare cases object coorindates lie outside of the image.
n = len(detection)
y[y > detect_img.shape[0]] = detect_img.shape[0]
x[x > detect_img.shape[1]] = detect_img.shape[1]
centroids[y, x] = np.arange(1, n + 1)
# Remove the padded border pixels
blobs = blobs[pad-1:-(pad-1), pad-1:-(pad-1)].copy()
centroids = centroids[pad-1:-(pad-1), pad-1:-(pad-1)].copy()
# Blobs detected outside of regions of interest are discarded.
blobs[mask > 0] = 0
blobs[mh.bwperim(np.invert(mask)) > 0] = 0
mh.labeled.relabel(blobs, inplace=True)
# We need to ensure that centroids are labeled the same way as blobs.
centroids[centroids > 0] = blobs[centroids > 0]
return (blobs, centroids)
def _compare_slow(img):
for n in (4,8):
assert np.all(_slow_bwperim(img, n) == bwperim(img, n))
def predict_wsis(model, dataset, ep):
'''
given directory svs_path,
current model goes through each
wsi (svs) and generates a
prediction mask embedded onto
wsi.
sequential: tiles images and
generates batches of size 1
parallel: uses preallocated
dataset and batches>>1
'''
os.makedirs('{}/{}'.format(args.val_save_pth, ep), exist_ok=True)
model.eval()
with torch.no_grad():
' go through each svs and make a pred. mask'
ious_tb = 0
for key in dataset.wsis:
'create prediction template'
pred = np.zeros(
(args.num_classes, *dataset.wsis[key]['iterator'].dataset.scan.
level_dimensions[args.scan_level][::-1]),
dtype=np.float)
'slide over wsi'
for batch_x, batch_y, batch_image in dataset.wsis[key]['iterator']:
batch_image = batch_image.cuda()
pred_src = model(batch_image)
if args.scan_resize != 1:
pred_src = torch.nn.functional.interpolate(
pred_src, (args.tile_h * args.scan_resize,
args.tile_w * args.scan_resize))
pred_src = pred_src.cpu().numpy()
for bj in range(batch_image.size(0)):
tile_x, tile_y = int(batch_x[bj]), int(batch_y[bj])
pred[:, tile_y:tile_y + dataset.params.ph,
tile_x:tile_x + dataset.params.pw] += pred_src[bj]
'post process wsi (throw out non-foreground tissue)'
scan = openslide.OpenSlide(dataset.wsis[key]['wsipath'])
mask = Image.open(dataset.wsis[key]['wsipath'] +
'_find_nuclei.png')
mask = np.asarray(mask)
'downsample pred'
pred_ = np.zeros(
(args.num_classes, *scan.level_dimensions[2][::-1]))
for ij in range(args.num_classes):
pred_[ij, ...] = cv2.resize(pred[ij, ...],
scan.level_dimensions[2])
pred = pred_
del pred_
'calculate score'
if os.path.exists(dataset.wsis[key]['wsipath'] + '_mask.png'):
gt = Image.open(dataset.wsis[key]['wsipath'] + '_mask.png')
gt = gt.resize(pred.shape[1:][::-1])
gt = np.array(gt)
p = np.argmax(pred, 0)
'''
get tumor bed
erosion: to remove small possibly
miss-regions
'''
# tb = (p.astype(np.uint8) == 3).astype(np.uint8)
tb = (p.astype(np.uint8) >= 2).astype(np.uint8)
tb = cv2.morphologyEx(tb, cv2.MORPH_OPEN,
np.ones((20, 20), dtype=np.uint8))
tb_pred = chull(tb)
tb = bwperim(tb_pred).astype(np.uint8)
tb = cv2.dilate(tb,
np.ones((20, 20), dtype=np.uint8),
iterations=1)
tb = np.nonzero(tb)
'''
use gt tumor bed
'''
tb_pth = dataset.wsis[key]['wsipath'] + '_tumor_bed.png'
if os.path.exists(tb_pth):
tb_gt = Image.open(tb_pth).convert('L')
tb_gt = (np.array(tb_gt) > 0).astype(np.uint8)
iou_tb = (tb_gt *
tb_pred).sum() / (args.epsilon +
(tb_gt | tb_pred).sum())
ious_tb += iou_tb
acc = (p == gt)
acc = acc[gt > 0]
acc = np.mean(acc)
s = 1 - np.sum(np.abs(p - gt)) / np.sum(
np.maximum(np.abs(gt - 0), np.abs(gt - 3.0)) *
(1 - (1 - (p > 0)) * (1 - gt > 0)))
p = mask * p
acc_masked = (p == gt)
acc_masked = acc_masked[gt > 0]
acc_masked = np.mean(acc_masked)
s_masked = 1 - np.sum(np.abs(p - gt)) / np.sum(
np.maximum(np.abs(gt - 0), np.abs(gt - 3.0)) *
(1 - (1 - (p > 0)) * (1 - gt > 0)))
'only detect foreground vs background'
iou_fg = ((p > 0) * (gt > 0)).sum() / (args.epsilon +
((p > 0) |
(gt > 0)).sum())
print('{}, '
'{:.3f}({:.3f}),'
' {:.3f}({:.3f}),'
' {:.3f},'
' tb iou: {:.3f} '.format(
dataset.wsis[key]['wsipath'].split('/')[-1],
s_masked, s, acc_masked, acc, iou_fg,
iou_tb if os.path.exists(tb_pth) else -1))
del p
del gt
'save color mask'
pred_image = np.expand_dims(mask,
-1) * preprocessing.pred_to_mask(pred)
pred_image[tb] = [255, 255, 255]
pred_image = Image.fromarray(pred_image)
pred_image.resize((dataset.wsis[key]['scan'].level_dimensions[2][0] // 2,
dataset.wsis[key]['scan'].level_dimensions[2][1] // 2)). \
save('{}/{}/{}_{}.png'.format(args.val_save_pth, ep, key, args.tile_stride_w))
del pred
del pred_image
print('Average tb iou: {:.3f}'.format(ious_tb / len(dataset.wsis)))
model.train()
def PST(I,
LPF=0.21,
Phase_strength=0.48,
Warp_strength=12.14,
Threshold_min=-1,
Threshold_max=0.0019,
Morph_flag=1):
# I: image
# Gaussian Low Pass Filter
# LPF = 0.21
# PST parameters:
# Phase_strength = 0.48
# Warp_strength = 12.14
# Thresholding parameters (for post processing after the edge is computed)
# Threshold_min = -1
# Threshold_max = 0.0019
# To compute analog edge, set Morph_flag = 0 and to compute digital edge, set Morph_flag = 1
# Morph_flag = 1
I_initial = I
if (len(I.shape) == 3):
I = I.mean(axis=2)
L = 0.5
x = np.linspace(-L, L, I.shape[0])
y = np.linspace(-L, L, I.shape[1])
[X1, Y1] = (np.meshgrid(x, y))
X = X1.T
Y = Y1.T
[THETA, RHO] = cart2pol(X, Y)
# Apply localization kernel to the original image to reduce noise
Image_orig_f = ((np.fft.fft2(I)))
expo = np.fft.fftshift(
np.exp(-np.power((np.divide(RHO, math.sqrt((LPF**2) /
np.log(2)))), 2)))
Image_orig_filtered = np.real(
np.fft.ifft2((np.multiply(Image_orig_f, expo))))
# Constructing the PST Kernel
PST_Kernel_1 = np.multiply(
np.dot(RHO, Warp_strength), np.arctan(np.dot(RHO, Warp_strength))
) - 0.5 * np.log(1 + np.power(np.dot(RHO, Warp_strength), 2))
PST_Kernel = PST_Kernel_1 / np.max(PST_Kernel_1) * Phase_strength
# Apply the PST Kernel
temp = np.multiply(np.fft.fftshift(np.exp(-1j * PST_Kernel)),
np.fft.fft2(Image_orig_filtered))
Image_orig_filtered_PST = np.fft.ifft2(temp)
# Calculate phase of the transformed image
PHI_features = np.angle(Image_orig_filtered_PST)
if Morph_flag == 0:
out = PHI_features
return out
else:
# find image sharp transitions by thresholding the phase
features = np.zeros((PHI_features.shape[0], PHI_features.shape[1]))
features[PHI_features > Threshold_max] = 1 # Bi-threshold decision
features[
PHI_features <
Threshold_min] = 1 # as the output phase has both positive and negative values
features[I < (
np.amax(I) / 20
)] = 0 # Removing edges in the very dark areas of the image (noise)
# apply binary morphological operations to clean the transformed image
out = features
out = mh.thin(out, 1)
out = mh.bwperim(out, 4)
out = mh.thin(out, 1)
out = mh.erode(out, np.ones((1, 1)))
Overlay = mh.overlay(I, out)
return (out, Overlay)
labels = labels.resize((x, y))
image = np.asarray(image)
labels = np.asarray(labels)
metadata = {}
for tile_id in range(labels.max()):
label_patch = labels == tile_id
n, center_pts, out_image, foreground_indices = regiontools.get_key_points(
label_patch, us_kmeans, dhr.HR_NUM_CNT_SAMPLES, dhr.HR_NUM_CNT_SAMPLES)
perim_coords = np.zeros([0, 2])
if dhr.HR_NUM_PERIM_SAMPLES > 0:
perim = bwperim(label_patch)
perim_coords = np.transpose(np.where(perim))[:, ::-1] # (x,y) pairs
skip = np.maximum(2, perim_coords.shape[0] // dhr.HR_NUM_PERIM_SAMPLES)
perim_coords = perim_coords[::skip, :]
metadata[tile_id] = {
'cnt_xy': center_pts,
'perim_xy': perim_coords,
'wsipath': svspth,
'scan_level': scan_level,
'foreground_indices': foreground_indices,
'tile_id': tile_id,
}
'''
evaluation stage
'''
expo = np.fft.fftshift(np.exp(-np.power((np.divide(rho, math.sqrt((LPF ** 2) / np.log(2)))), 2)))
orig_filtered = np.real(np.fft.ifft2((np.multiply(orig, expo))))
PST_Kernel_1 = np.multiply(np.dot(rho, W), np.arctan(np.dot(rho, W))) - 0.5 * np.log(1 + np.power(np.dot(rho, W), 2))
PST_Kernel = PST_Kernel_1 / np.max(PST_Kernel_1) * S
temp = np.multiply(np.fft.fftshift(np.exp(-1j * PST_Kernel)), np.fft.fft2(orig_filtered))
temp = np.multiply(np.fft.fftshift(np.exp(-1j * PST_Kernel)), np.fft.fft2(Image_orig_filtered))
orig_filtered_PST = np.fft.ifft2(temp)
PHI_features = np.angle(Image_orig_filtered_PST)
features = np.zeros((PHI_features.shape[0], PHI_features.shape[1]))
features[PHI_features > Threshold_max] = 1
features[PHI_features < Threshold_min] = 1
features[I < (np.amax(I) / 20)] = 0
out = features
out = mh.thin(out, 1)
out = mh.bwperim(out, 4)
out = mh.thin(out, 1)
out = mh.erode(out, np.ones((1, 1)))
def phase_stretch_transform(img, LPF, S, W, threshold_min, threshold_max, flag):
L = 0.5
x = np.linspace(-L, L, img.shape[0])
y = np.linspace(-L, L, img.shape[1])
[X1, Y1] = (np.meshgrid(x, y))
X = X1.T
Y = Y1.T
theta, rho = cart2pol(X, Y)
orig = ((np.fft.fft2(img)))
expo = np.fft.fftshift(np.exp(-np.power((np.divide(rho, math.sqrt((LPF ** 2) / np.log(2)))), 2)))
orig_filtered = np.real(np.fft.ifft2((np.multiply(orig, expo))))
PST_Kernel_1 = np.multiply(np.dot(rho, W), np.arctan(np.dot(rho, W))) - 0.5 * np.log(
def test_n5():
img = np.zeros((8, 8), np.bool)
img[3:7, 3:7] = 1
bwperim(img, 5)
for imageNumber in range(len(filenames)):
cellID.append(str(uuid.uuid4()))
filename="file.csv"
df = pd.DataFrame (cellID, columns=["CellID"]) # assign unique ID for each cell
df.to_csv(filename, index=False)
# finding a list of pixels in cell boundries
cell_boundaries_list=[]
filenames = list(glob.glob('./cropped_and_resized/*')) #load image
for idx,filename in enumerate(filenames):
targetCell = skimage.io.imread(filenames[idx])
dimension = targetCell.shape
# calculating cell boundaries
boolean_border = mh.bwperim(targetCell, n=4)
cell_boundaries = np.ravel_multi_index(np.nonzero(boolean_border),dimension)
# formatting the calculated cell boundaries
CB_string=map(str, cell_boundaries) # converting all the elements in the array to string
for x,y in enumerate(CB_string): # add ", " to all the elements so that when they are combined, they are still distinctly recognizable
CB_string[x]=CB_string[x]+' '
boundaries=''.join(CB_string) # combine all elements in an array to form a single string
# adding the calculation to the cell_boundaries_list
cell_boundaries_list.append(boundaries)
csv = "file.csv"
df = utils.read_csv(csv)
print df.shape
df['cell boundary']=pd.Series (cell_boundaries_list)
df['FileName']=pd.Series (filenames)
df.to_csv(csv, index=False)
file="part3.png"
complete_path=os.path.join(workdir,file)
if __name__ == "__main__":
im=readmagick.readimg(complete_path)
im0=np.copy(im)
hip=pymorph.subm(im,nd.gaussian_filter(im,5))
hip0=np.copy(hip)
im=mahotas.thin(im)
#im=mahotas.bwperim(im>0)
hip=mahotas.thin(hip)
#print im.dtype#print uint16
p1=particle(im)
print "particle 1",p1.cvxhull_area()
p2=particle(hip)
contour=mahotas.bwperim(im>0)
p3=particle(contour)
print "particle 1 hi pass",p2.cvxhull_area()
theta1,rosedesvents,VImage=p1.orientationByErosion(5)
theta2,rdv,VHip=p2.orientationByErosion(5)
x=rosedesvents[0,:]
y=rosedesvents[1,:]
xh=rdv[0,:]
yh=rdv[1,:]
pylab.subplot(321,frameon=False, xticks=[], yticks=[])
pylab.gray()
pylab.imshow(im0)
pylab.subplot(322,frameon=False, xticks=[], yticks=[])
pylab.imshow(hip0)
pylab.subplot(323)
pylab.title("Resistance to vertical erosion")
file = "part3.png"
complete_path = os.path.join(workdir, file)
if __name__ == "__main__":
im = readmagick.readimg(complete_path)
im0 = np.copy(im)
hip = pymorph.subm(im, nd.gaussian_filter(im, 5))
hip0 = np.copy(hip)
im = mahotas.thin(im)
#im=mahotas.bwperim(im>0)
hip = mahotas.thin(hip)
#print im.dtype#print uint16
p1 = particle(im)
print "particle 1", p1.cvxhull_area()
p2 = particle(hip)
contour = mahotas.bwperim(im > 0)
p3 = particle(contour)
print "particle 1 hi pass", p2.cvxhull_area()
theta1, rosedesvents, VImage = p1.orientationByErosion(5)
theta2, rdv, VHip = p2.orientationByErosion(5)
x = rosedesvents[0, :]
y = rosedesvents[1, :]
xh = rdv[0, :]
yh = rdv[1, :]
pylab.subplot(321, frameon=False, xticks=[], yticks=[])
pylab.gray()
pylab.imshow(im0)
pylab.subplot(322, frameon=False, xticks=[], yticks=[])
pylab.imshow(hip0)
pylab.subplot(323)
pylab.title("Resistance to vertical erosion")
cx, cy = centers[tile_id, :]
label_patch = labels == tile_id
area = np.count_nonzero(label_patch)
num_clusters = dhr.HR_NUM_CNT_SAMPLES #6 + int(area / (0.01 * gt.size))
if (w * h)/gt.size <= 0.005:
n, cnt_pts, out_image, foreground_indices = regiontools.get_key_points(label_patch, us_kmeans, dhr.HR_NUM_CNT_SAMPLES, dhr.HR_NUM_CNT_SAMPLES)
if n is not None:
label_patch = Image.fromarray(label_patch.astype(np.uint8))
x, y = label_patch.size
label_patch = label_patch.resize((x // us_kmeans, y // us_kmeans))
perim = bwperim(np.asarray(label_patch))
coords_ = np.transpose(np.where(perim))[:, ::-1] # (x,y) pairs
cvh = concaveHull(coords_, 3)
coords = esp(cvh, dhr.HR_NUM_PERIM_SAMPLES) * us_kmeans
plt.plot(cnt_pts[:, 0], cnt_pts[:, 1], 'wo', ms=ms)
plt.plot(coords[:, 0], coords[:, 1], 'ko', ms=ms)
else:
min_center_large = int((w * h)/(gt.size * 0.01))
min_center_large = np.maximum(min_center_large, 5)
min_center_large = num_clusters
n, cnt_pts, out_image, foreground_indices = regiontools.get_key_points(label_patch, us_kmeans, min_center_large, min_center_large)
label_patch = labels == tile_id
area = np.count_nonzero(label_patch)
num_clusters = 2 + int(area / (0.01 * labels.size))
n, center_pts, out_image, foreground_indices = regiontools.get_key_points(
label_patch, us_kmeans, num_clusters, num_clusters)
'get width & height'
indices = np.where(label_patch)
h = 1 + indices[0].max() - indices[0].min()
w = 1 + indices[1].max() - indices[1].min()
if n is not None and (w * h) / labels.size <= 0.05:
perim = bwperim(label_patch)
perim_coords = np.transpose(np.where(perim))[:, ::-1] # (x,y) pairs
skip = np.maximum(2, perim_coords.shape[0] // dhr.HR_NUM_PERIM_SAMPLES)
perim_coords = perim_coords[::skip, :]
metadata[patch_id] = {
'cnt_xy': center_pts,
'perim_xy': perim_coords,
'wsipath': svspth,
'scan_level': scan_level,
'foreground_indices': foreground_indices,
'tile_id': patch_id,
}
patch_id = patch_id + 1
elif n is not None:
