def create_profile():
lm = Landmarks('Data/Landmarks/original/landmarks1-1.txt')
img = rg.load()
img = rg.enhance(img[0])
grad_img = rg.togradient_sobel(img)
profile = Profile(img, grad_img, lm, 30, 40)
Plotter.plot_profile(grad_img, profile)
def create_profile():
lm = Landmarks('Data/Landmarks/original/landmarks1-1.txt')
img = rg.load()
img = rg.enhance(img[0])
grad_img = rg.togradient_sobel(img)
profile = Profile(img, grad_img, lm, 30, 40)
Plotter.plot_profile(grad_img, profile)
def build_grey_level_model():
lms = landmarks.load(2)
images = rg.load()
images = [rg.enhance(img) for img in images]
gimages = [rg.togradient_sobel(img) for img in images]
glm = GreyLevelModel()
glm.build(images, gimages, lms, 10, 20)
Plotter.plot_grey_level_model(glm, gimages)
def build_grey_level_model():
lms = landmarks.load(2)
images = rg.load()
images = [rg.enhance(img) for img in images]
gimages = [rg.togradient_sobel(img) for img in images]
glm = GreyLevelModel()
glm.build(images, gimages, lms, 10, 20)
Plotter.plot_grey_level_model(glm, gimages)
def auto_init():
imgs = rg.load()
for radiograph in range(0, 14):
img = imgs[radiograph]
X = []
for tooth in range(1, 9):
model = IncisorModel.load(tooth)
X.append(ai.init(model, img))
Plotter.plot_landmarks_on_image(X, img, show=False, save=True,
title='Autoinit/%02d' % (radiograph,))
def build_database(path):
"""Interactively select ROIs on the dental radiographs for creating the
appearance model used for the automatic initialization.
Args:
path (str): The location of the dental radiographs.
"""
images = rg.load(path+"/", indices=range(1, 31))
create_database(images)
def train_and_save():
# parameters
k = 10
imgs = rg.load()
for tooth in range(1, 9):
lms = landmarks.load_mirrored(tooth)
model = IncisorModel(tooth)
model.train(lms, imgs, k)
model.save()
def train_and_save():
# parameters
k = 10
imgs = rg.load()
for tooth in range(1, 9):
lms = landmarks.load_mirrored(tooth)
model = IncisorModel(tooth)
model.train(lms, imgs, k)
model.save()
def fit(radiograph_index, incisor_ind, fit_mode, k, m, out):
"""Find the region of an incisor, using the given parameters.
Args:
radiograph_index (int): The index of the dental radiograph to fit on.
incisor_ind (int): The index of the incisor to fit.
fit mode (AUTO|MANUAL): Wheter to ask for a manual initial fit, or try
to find one automatically.
k (int): Number of pixels either side of point to represent in grey model.
m (int): Number of sample points either side of current point for search.
out (str): The location to store the result.
"""
# leave-one-out
train_indices = range(0, 14)
train_indices.remove(radiograph_index-1)
lms = landmarks.load_mirrored(incisor_ind)
test_lm = lms[radiograph_index-1]
train_lms = [lms[index] for index in train_indices]
imgs = rg.load()
test_img = imgs[radiograph_index-1]
train_imgs = [imgs[index] for index in train_indices]
# train
model = IncisorModel(incisor_ind)
model.train(train_lms, train_imgs, k)
# fit
X = model.estimate_fit(test_img, fit_mode)
X = model.fit(X, test_img, m)
# evaluate
## show live
Plotter.plot_landmarks_on_image([test_lm, X], test_img, wait=False)
## save image with tooth circled
img = test_img.copy()
colors = [(255, 0, 0), (0, 255, 0)]
for ind, lms in enumerate([test_lm, X]):
points = lms.as_matrix()
for i in range(len(points) - 1):
cv2.line(img, (int(points[i, 0]), int(points[i, 1])),
(int(points[i + 1, 0]), int(points[i + 1, 1])),
colors[ind])
cv2.imwrite('%s/%02d-%d.png' % (out, radiograph_index, incisor_ind,), img)
## save tooth region segmented
height, width, _ = test_img.shape
image2 = np.zeros((height, width), np.int8)
mask = np.array([X.points], dtype=np.int32)
cv2.fillPoly(image2, [mask], 255)
maskimage2 = cv2.inRange(image2, 1, 255)
segmented = cv2.bitwise_and(test_img, test_img, mask=maskimage2)
cv2.imwrite('%s/%02d-%d-segmented.png' % (out, radiograph_index, incisor_ind,), segmented)
def auto_init():
imgs = rg.load()
for radiograph in range(0, 14):
img = imgs[radiograph]
X = []
for tooth in range(1, 9):
model = IncisorModel.load(tooth)
X.append(ai.init(model, img))
Plotter.plot_landmarks_on_image(X,
img,
show=False,
save=True,
title='Autoinit/%02d' % (radiograph, ))
def preprocess(path):
"""Apply an adaptive median filter to the radiographs at the given path and
stores the results in a separate ``filtered`` folder.
Args:
path (str): The location of the dental radiographs.
"""
images = rg.load(path+"/")
for ind, img in enumerate(images):
print "Preprocessing radiographs ["+str(ind+1)+"/14]"
img = rg.adaptive_median(img, 3, 5)
cv2.imwrite("%s/filtered/%02d.tif" % (path, ind+1,), img)
def auto_fit():
# parameters
m = 15
imgs = rg.load()
for radiograph in range(0, 14):
img = imgs[radiograph]
Xs = []
for tooth in range(1, 9):
model = IncisorModel.load(tooth)
X = model.estimate_fit(img, MODE_FIT_AUTO)
X = model.fit(X, img, m)
Xs.append(X)
Plotter.plot_landmarks_on_image(Xs, img, show=False, save=True,
title='Autofit/%02d' % (radiograph,))
def auto_fit():
# parameters
m = 15
imgs = rg.load()
for radiograph in range(0, 14):
img = imgs[radiograph]
Xs = []
for tooth in range(1, 9):
model = IncisorModel.load(tooth)
X = model.estimate_fit(img, MODE_FIT_AUTO)
X = model.fit(X, img, m)
Xs.append(X)
Plotter.plot_landmarks_on_image(Xs,
img,
show=False,
save=True,
title='Autofit/%02d' % (radiograph, ))
def run_experiment(mode):
"""Test the fit quality for all training examples.
Args:
mode: MODE_FIT_AUTO or MODE_FIT_MANUAL
"""
imgs = rg.load()
for radiograph in range(0, 14):
Xs = []
for tooth in range(1, 9):
X = test_fit(radiograph, tooth, mode)
Xs.append(X)
img = imgs[radiograph]
Plotter.plot_landmarks_on_image(Xs,
img,
show=False,
save=True,
title='Results/%02d' %
(radiograph + 1, ))
def init(model, img, show=False):
"""Find an initial estimate for the model in the given image.
Args:
model (Landmarks): The shape which should be fitted.
img: The dental radiograph on which the shape should be fitted.
Returns:
Landmarks: An initial estimate for the position of the model in the given image.
"""
# Are we fitting a lower or an upper incisor?
tooth = model.incisor_nr
is_upper = tooth < 5
if is_upper:
# UPPER: Avg. height: 314.714285714, Avg. width: 381.380952381
width = 380
height = 315
else:
# LOWER: Avg. height: 259.518518519, Avg. width: 281.518518519
width = 280
height = 260
# Create the appearance model for the four upper/lower teeth
radiographs = rg.load()
data = load_database(radiographs, is_upper, width, height)
[_, evecs, mean] = pca(data, 5)
# Visualize the appearance model
# cv2.imshow('img',np.hstack( (mean.reshape(height,width),
# normalize(evecs[:,0].reshape(height,width)),
# normalize(evecs[:,1].reshape(height,width)),
# normalize(evecs[:,2].reshape(height,width)))
# ).astype(np.uint8))
# cv2.waitKey(0)
# Find the jaw split
jaw_split = split_jaws(img)
# Find the region of the radiograph that matches best with the appearance model
img = rg.enhance(img)
[(a, b), (c, d)], _ = find_bbox(mean, evecs, img, width, height, is_upper, jaw_split, show=show)
# Assume all teeth have more or less the same width
ind = tooth if tooth < 5 else tooth - 4
bbox = [(a +(ind-1)*(c-a)/4, b), (a +(ind)*(c-a)/4, d)]
center = np.mean(bbox, axis=0)
# Plot a bounding box around the estimated region of the requested tooth
if show:
Plotter.plot_autoinit(img, bbox, 0, jaw_split, wait=True)
# Position the mean shape of the requested incisor inside the bbox
template = model.asm.mean_shape.scale_to_bbox(bbox).translate(center)
# The position of the lower incisors is further improved using the grey-level model
if is_upper:
X = template
else:
X = template
# X = fit_template(template, model, img)
# Show the final result
if show:
Plotter.plot_landmarks_on_image([X], img)
# Return the estimated position of the shape's landmark points
return X
def AM_filter():
img = rg.load()[0]
filt_img = rg.adaptive_median(img, 3, 5)
cv2.imwrite('Data/Radiographs/filtered/01.tif', filt_img)
def init(model, img, show=False):
"""Find an initial estimate for the model in the given image.
Args:
model (Landmarks): The shape which should be fitted.
img: The dental radiograph on which the shape should be fitted.
Returns:
Landmarks: An initial estimate for the position of the model in the given image.
"""
# Are we fitting a lower or an upper incisor?
tooth = model.incisor_nr
is_upper = tooth < 5
if is_upper:
# UPPER: Avg. height: 314.714285714, Avg. width: 381.380952381
width = 380
height = 315
else:
# LOWER: Avg. height: 259.518518519, Avg. width: 281.518518519
width = 280
height = 260
# Create the appearance model for the four upper/lower teeth
radiographs = rg.load()
data = load_database(radiographs, is_upper, width, height)
[_, evecs, mean] = pca(data, 5)
# Visualize the appearance model
# cv2.imshow('img',np.hstack( (mean.reshape(height,width),
# normalize(evecs[:,0].reshape(height,width)),
# normalize(evecs[:,1].reshape(height,width)),
# normalize(evecs[:,2].reshape(height,width)))
# ).astype(np.uint8))
# cv2.waitKey(0)
# Find the jaw split
jaw_split = split_jaws(img)
# Find the region of the radiograph that matches best with the appearance model
img = rg.enhance(img)
[(a, b), (c, d)], _ = find_bbox(mean,
evecs,
img,
width,
height,
is_upper,
jaw_split,
show=show)
# Assume all teeth have more or less the same width
ind = tooth if tooth < 5 else tooth - 4
bbox = [(a + (ind - 1) * (c - a) / 4, b), (a + (ind) * (c - a) / 4, d)]
center = np.mean(bbox, axis=0)
# Plot a bounding box around the estimated region of the requested tooth
if show:
Plotter.plot_autoinit(img, bbox, 0, jaw_split, wait=True)
# Position the mean shape of the requested incisor inside the bbox
template = model.asm.mean_shape.scale_to_bbox(bbox).translate(center)
# The position of the lower incisors is further improved using the grey-level model
if is_upper:
X = template
else:
X = template
# X = fit_template(template, model, img)
# Show the final result
if show:
Plotter.plot_landmarks_on_image([X], img)
# Return the estimated position of the shape's landmark points
return X
center = np.mean(bbox, axis=0)
# Plot a bounding box around the estimated region of the requested tooth
if show:
Plotter.plot_autoinit(img, bbox, 0, jaw_split, wait=True)
# Position the mean shape of the requested incisor inside the bbox
template = model.asm.mean_shape.scale_to_bbox(bbox).translate(center)
# The position of the lower incisors is further improved using the grey-level model
if is_upper:
X = template
else:
X = template
# X = fit_template(template, model, img)
# Show the final result
if show:
Plotter.plot_landmarks_on_image([X], img)
# Return the estimated position of the shape's landmark points
return X
if __name__ == '__main__':
# create_database()
from incisor_model import IncisorModel
model = IncisorModel.load(5)
img = rg.load()[7]
init(model, img, show=False)
def preprocess():
img = rg.load()[0]
img = rg.enhance(img)
grad = rg.togradient_sobel(img)
lms = landmarks.load_all_incisors_of_example(1)
Plotter.plot_landmarks_on_image(lms, grad)
def AM_filter():
img = rg.load()[0]
filt_img = rg.adaptive_median(img, 3, 5)
cv2.imwrite('Data/Radiographs/filtered/01.tif', filt_img)
"""
def gaussian_function(sigma, u):
return 1 / (math.sqrt(2 * math.pi) * sigma) * math.e**-(u**2 /
(2 * sigma**2))
if filter_length is None:
#determine the length of the filter
filter_length = math.ceil(sigma * 5)
#make the length odd
filter_length = 2 * (int(filter_length) / 2) + 1
#make sure sigma is a float
sigma = float(sigma)
#create the filter
result = np.asarray([
gaussian_function(sigma, u)
for u in range(-(filter_length / 2), filter_length / 2 + 1, 1)
])
result = result / result.sum()
#return the filter
return result
if __name__ == '__main__':
# test it on all radiographs
imgs = rg.load()
for i in range(0, 14):
split_jaws(imgs[i], 50, True)
def test_fit(test_index, incisor_ind, fit_mode):
"""Test the fit quality of one incisor using leave-one-out analysis.
Args:
test_index (int): The index of the radiograph to fit on. [0-13]
incisor_ind (int): The index of the incisor to fit. [1-8]
fit_mode: MODE_FIT_AUTO or MODE_FIT_MANUAL
Returns:
Landmarks: The fitted model.
"""
# parameters
k = 10
m = 15
# leave-one-out
train_indices = range(0, 14)
train_indices.remove(test_index)
lms = landmarks.load_mirrored(incisor_ind)
test_lm = lms[test_index]
train_lms = [lms[index] for index in train_indices]
imgs = rg.load()
test_img = imgs[test_index]
train_imgs = [imgs[index] for index in train_indices]
# train
model = IncisorModel(incisor_ind)
model.train(train_lms, train_imgs, k)
# test
X = model.estimate_fit(test_img, fit_mode)
X = model.fit(X, test_img, m)
# evaluate
Plotter.plot_landmarks_on_image([test_lm, X], test_img, wait=False)
fit = fit_quality(test_lm, X)
# save result
## landmark annotated
img = test_img.copy()
colors = [(255, 0, 0), (0, 255, 0)]
for ind, lms in enumerate([test_lm, X]):
points = lms.as_matrix()
for i in range(len(points) - 1):
cv2.line(img, (int(points[i, 0]), int(points[i, 1])),
(int(points[i + 1, 0]), int(points[i + 1, 1])),
colors[ind])
cv2.imwrite('Plot/Results/%02d-%d.png' % (
test_index + 1,
incisor_ind,
), img)
## tooth region segmented
height, width, _ = test_img.shape
image2 = np.zeros((height, width), np.int8)
mask = np.array([X.points], dtype=np.int32)
cv2.fillPoly(image2, [mask], 255)
maskimage2 = cv2.inRange(image2, 1, 255)
out = cv2.bitwise_and(test_img, test_img, mask=maskimage2)
cv2.imwrite(
'Plot/Results/Segmentations/%02d-%d.png' % (
test_index + 1,
incisor_ind - 1,
), out)
truth = cv2.imread(
'Data/Segmentations/%02d-%d.png' % (
test_index + 1,
incisor_ind - 1,
), 0)
(_, truth) = cv2.threshold(truth, 128, 255,
cv2.THRESH_BINARY | cv2.THRESH_OTSU)
fit = cv2.imread(
'Plot/Results/Segmentations/%02d-%d.png' % (
test_index + 1,
incisor_ind - 1,
), 0)
(_, fit) = cv2.threshold(fit, 128, 255,
cv2.THRESH_BINARY | cv2.THRESH_OTSU)
fit2 = fit_quality2(truth, fit)
## fit quality
results = np.loadtxt('experiment_results.csv', delimiter=',')
results[test_index, incisor_ind - 1] = fit2
np.savetxt('experiment_results.csv', results, fmt='%10.2f', delimiter=',')
return X
# Plot a bounding box around the estimated region of the requested tooth
if show:
Plotter.plot_autoinit(img, bbox, 0, jaw_split, wait=True)
# Position the mean shape of the requested incisor inside the bbox
template = model.asm.mean_shape.scale_to_bbox(bbox).translate(center)
# The position of the lower incisors is further improved using the grey-level model
if is_upper:
X = template
else:
X = template
# X = fit_template(template, model, img)
# Show the final result
if show:
Plotter.plot_landmarks_on_image([X], img)
# Return the estimated position of the shape's landmark points
return X
if __name__ == '__main__':
# create_database()
from incisor_model import IncisorModel
model = IncisorModel.load(5)
img = rg.load()[7]
init(model, img, show=False)
A 1-D numpy array of odd length, containing the symmetric, discrete
approximation of a Gaussian with sigma Summation of the array-values
must be equal to one.
"""
def gaussian_function(sigma, u):
return 1/(math.sqrt(2*math.pi)*sigma)*math.e**-(u**2/(2*sigma**2))
if filter_length is None:
#determine the length of the filter
filter_length = math.ceil(sigma*5)
#make the length odd
filter_length = 2*(int(filter_length)/2) + 1
#make sure sigma is a float
sigma = float(sigma)
#create the filter
result = np.asarray([gaussian_function(sigma, u) for u in range(-(filter_length/2), filter_length/2 + 1, 1)])
result = result / result.sum()
#return the filter
return result
if __name__ == '__main__':
# test it on all radiographs
imgs = rg.load()
for i in range(0, 14):
split_jaws(imgs[i], 50, True)
def preprocess():
img = rg.load()[0]
img = rg.enhance(img)
grad = rg.togradient_sobel(img)
lms = landmarks.load_all_incisors_of_example(1)
Plotter.plot_landmarks_on_image(lms, grad)