def test_is_uint8_image():
R = np.array([[0., 55., 250.], [1., 3., 7.], [2., 4., 7.]])
G = np.array([[0., 55., 250.], [1., 3., 7.], [2., 4., 7.]])
B = np.array([[0., 55., 250.], [1., 3., 7.], [2., 4., 7.]])
I = np.array([R, G, B])
assert not is_uint8_image(I)
I = I.astype(np.uint8)
assert is_uint8_image(I)
I = I / 255
assert not is_uint8_image(I)
def get_stain_matrix(I, luminosity_threshold=0.8, regularizer=0.1):
"""
Stain matrix estimation via method of:
A. Vahadane et al. 'Structure-Preserving Color Normalization and Sparse Stain Separation for Histological Images'
:param I: Image RGB uint8.
:param luminosity_threshold:
:param regularizer:
:return:
"""
assert is_uint8_image(I), "Image should be RGB uint8."
# convert to OD and ignore background
tissue_mask = LuminosityThresholdTissueLocator.get_tissue_mask(
I, luminosity_threshold=luminosity_threshold).reshape((-1, ))
OD = convert_RGB_to_OD(I).reshape((-1, 3))
OD = OD[tissue_mask]
# do the dictionary learning
dictionary = spams.trainDL(X=OD.T,
K=2,
lambda1=regularizer,
mode=2,
modeD=0,
posAlpha=True,
posD=True,
verbose=False).T
# order H and E.
# H on first row.
if dictionary[0, 0] < dictionary[1, 0]:
dictionary = dictionary[[1, 0], :]
return normalize_matrix_rows(dictionary)
def get_stain_matrix(I, luminosity_threshold=0.8, angular_percentile=99):
"""
Stain matrix estimation via method of:
M. Macenko et al. 'A method for normalizing histology slides for quantitative analysis'
:param I: Image RGB uint8.
:param luminosity_threshold:
:param angular_percentile:
:return:
"""
assert is_uint8_image(I), "Image should be RGB uint8."
# Convert to OD and ignore background
tissue_mask = LuminosityThresholdTissueLocator.get_tissue_mask(
I, luminosity_threshold=luminosity_threshold).reshape((-1, ))
OD = convert_RGB_to_OD(I).reshape((-1, 3))
OD = OD[tissue_mask]
# Yoni's edit!!
if tissue_mask.sum(
) <= 1: # If the tissue mask has 0 or 1 pixel the linalg does not work. We will return an untransformed image
return False
# Eigenvectors of cov in OD space (orthogonal as cov symmetric)
_, V = np.linalg.eigh(np.cov(OD, rowvar=False))
# The two principle eigenvectors
V = V[:, [2, 1]]
# Make sure vectors are pointing the right way
if V[0, 0] < 0: V[:, 0] *= -1
if V[0, 1] < 0: V[:, 1] *= -1
# Project on this basis.
That = np.dot(OD, V)
# Angular coordinates with repect to the prinicple, orthogonal eigenvectors
phi = np.arctan2(That[:, 1], That[:, 0])
# Min and max angles
minPhi = np.percentile(phi, 100 - angular_percentile)
maxPhi = np.percentile(phi, angular_percentile)
# the two principle colors
v1 = np.dot(V, np.array([np.cos(minPhi), np.sin(minPhi)]))
v2 = np.dot(V, np.array([np.cos(maxPhi), np.sin(maxPhi)]))
# Order of H and E.
# H first row.
if v1[0] > v2[0]:
HE = np.array([v1, v2])
else:
HE = np.array([v2, v1])
return normalize_matrix_rows(HE)
def lab_split(I):
"""
Convert from RGB uint8 to LAB and split into channels.
:param I: Image RGB uint8.
:return:
"""
assert is_uint8_image(I), "Should be a RGB uint8 image"
I = cv.cvtColor(I, cv.COLOR_RGB2LAB)
I_float = I.astype(np.float32)
I1, I2, I3 = cv.split(I_float)
I1 /= 2.55 # should now be in range [0,100]
I2 -= 128.0 # should now be in range [-127,127]
I3 -= 128.0 # should now be in range [-127,127]
return I1, I2, I3
def get_mean_std(self, I):
"""
Get mean and standard deviation of each channel.
:param I: Image RGB uint8.
:return:
"""
assert is_uint8_image(I), "Should be a RGB uint8 image"
I1, I2, I3 = self.lab_split(I)
m1, sd1 = cv.meanStdDev(I1)
m2, sd2 = cv.meanStdDev(I2)
m3, sd3 = cv.meanStdDev(I3)
means = m1, m2, m3
stds = sd1, sd2, sd3
return means, stds
def standardize(I, percentile=95):
"""
Transform image I to standard brightness.
Modifies the luminosity channel such that a fixed percentile is saturated.
:param I: Image uint8 RGB.
:param percentile: Percentile for luminosity saturation. At least (100 - percentile)% of pixels should be fully luminous (white).
:return: Image uint8 RGB with standardized brightness.
"""
assert is_uint8_image(I), "Image should be RGB uint8."
I_LAB = cv.cvtColor(I, cv.COLOR_RGB2LAB)
L_float = I_LAB[:, :, 0].astype(float)
p = np.percentile(L_float, percentile)
I_LAB[:, :, 0] = np.clip(255 * L_float / p, 0, 255).astype(np.uint8)
I = cv.cvtColor(I_LAB, cv.COLOR_LAB2RGB)
return I
def get_mean_std(self, I, mask):
"""
Get mean and standard deviation of each channel.
:param I: Image RGB uint8.
:return:
"""
assert is_uint8_image(I), "Should be a RGB uint8 image"
I1, I2, I3 = self.lab_split(I)
m1 = np.mean(I1[mask])
sd1 = np.std(I1[mask])
m2 = np.mean(I2[mask])
sd2 = np.std(I2[mask])
m3 = np.mean(I3[mask])
sd3 = np.std(I3[mask])
means = m1, m2, m3
stds = sd1, sd2, sd3,
return means, stds
def get_tissue_mask(I, luminosity_threshold=0.8):
"""
Get a binary mask where true denotes pixels with a luminosity less than the specified threshold.
Typically we use to identify tissue in the image and exclude the bright white background.
:param I: RGB uint 8 image.
:param luminosity_threshold: Luminosity threshold.
:return: Binary mask.
"""
assert is_uint8_image(I), "Image should be RGB uint8."
I_LAB = cv.cvtColor(I, cv.COLOR_RGB2LAB)
L = I_LAB[:, :, 0] / 255.0 # Convert to range [0,1].
mask = L < luminosity_threshold
# Check it's not empty
if mask.sum() == 0:
raise TissueMaskException("Empty tissue mask computed")
return mask
def test_is_uint8_image_is_false_for_floats(self):
rgb = np.random.uniform(0, 256, [2, 7, 3])
get = is_uint8_image(rgb)
self.assertFalse(get)
def test_is_uint8_image_is_false_for_ints_outside_0_255(self):
rgb = np.random.randint(0, 300, [5, 2, 3])
get = is_uint8_image(rgb)
self.assertFalse(get)
def test_is_uint8_image_is_true_for_ints_in_0_255(self):
rgb = np.random.randint(0, 256, [5, 4, 3]).astype(np.uint8)
get = is_uint8_image(rgb)
self.assertTrue(get)
