def extract_HRF(db_folder, expert=1):
"""
Full images with primitive optic disc segmentation (as a circle).
2336 x 3504 original, 3504 x 3504 after preprocessing.
Accepted values for `expert`: 1, 2.
Required schema:
db_folder/
Healthy/
{:02}_h.jpg (number from 01 to 15)
Glaucomatous/
{:02}_h.jpg (number from 01 to 15)
optic_disk_centers_expert_A.csv
optic_disk_centers_expert_B.csv
"""
orig_resolution = (2336, 3504)
result_resolution = (3504, 3504)
if expert == 1:
expert_letter = 'A'
elif expert == 2:
expert_letter = 'B'
anot_df = pd.read_csv(os.path.join(
db_folder, 'optic_disk_centers_expert_{}.csv'.format(expert_letter)),
index_col=0)
X_all, Y_all, file_codes_all, is_ill = [], [], [], []
for pic_type in ('Healthy', 'Glaucomatous'):
X, file_names = imh.load_set(os.path.join(db_folder, pic_type))
X_all.extend(X)
rel_file_names = [os.path.split(fn)[-1] for fn in file_names]
file_codes = [fn[:fn.rfind('.')] for fn in rel_file_names]
file_codes_all.extend(file_codes)
for i in xrange(len(X)):
record_str = file_codes[i]
if expert == 2:
record_str = record_str.replace('_', '')
anot_record = anot_df.loc[record_str]
od_center = (anot_record['Pap. Center x'],
anot_record['Pap. Center y'])
#od_center = (anot_record['vessel orig. x'], anot_record['vessel orig. y'])
od_radius = anot_record['disk diameter'] / 2
segmn_img = np.zeros(orig_resolution, dtype=np.uint8)
cv2.circle(segmn_img, od_center, od_radius, color=1, thickness=-1)
Y_all.append(segmn_img)
is_ill.append(HEALTHY if pic_type ==
'Healthy' else GLAUCOMA_OR_SUSPECT)
for i in xrange(len(X_all)):
side = result_resolution[0]
X_all[i] = imh.resize_image_to_square(X_all[i], side, pad_cval=0)
Y_all[i] = imh.resize_image_to_square(Y_all[i], side, pad_cval=0)
Y_all[i] = Y_all[i].reshape(Y_all[i].shape + (1, ))
return X_all, Y_all, file_codes_all, is_ill
def extract_DRIONS_DB(db_folder, expert=1):
"""
Full images with polygonal optic disc segmentation from 2 experts.
400 x 600 original, 560 x 560 after post-processing.
Images have unwanted text at the left, so it needs to be dropped out.
Accepted values for `expert`: 1, 2.
Required schema:
db_folder/
images/
image_{:03}.jpg
experts_annotation/
anotExpert1_{:03}.txt
anotExpert2_{:03}.txt
"""
orig_resolution = (400, 600)
left_cut_thr = 40
result_resolution = (orig_resolution[1] - left_cut_thr,
orig_resolution[1] - left_cut_thr)
X, filenames = imh.load_set(os.path.join(db_folder, 'images'))
file_codes = [fn[-7:-4] for fn in filenames]
Y = []
for i, code in enumerate(file_codes):
anot_filename = os.path.join(
db_folder, 'experts_annotation',
'anotExpert{}_{}.txt'.format(expert, code))
with open(anot_filename) as anot_fin:
coords = anot_fin.readlines()
coords = map(
lambda s: map(lambda x: int(round(float(x))), s.split(' , ')),
coords)
coords = np.array(coords)
segm_img = np.zeros(orig_resolution, dtype=np.uint8)
cv2.fillPoly(segm_img, coords.reshape((1, ) + coords.shape), color=1)
Y.append(segm_img)
for i in xrange(len(X)):
side = result_resolution[0]
X[i] = imh.resize_image_to_square(X[i][:, left_cut_thr:],
side,
pad_cval=0)
Y[i] = imh.resize_image_to_square(Y[i][:, left_cut_thr:],
side,
pad_cval=0)
Y[i] = Y[i].reshape(Y[i].shape + (1, ))
return X, Y, file_codes
def extract_DRISHTI_GS_test(db_folder):
"""
Full images with optic disc and optic cup segmentation.
Average segmentation and "softmap" segmentation image are given.
51 images of various resolution close to 2040 x 1740.
Data set is split into training and test sets. Groundtruth is available for training set only.
This function returns Test set only.
Required schema:
db_folder/
Drishti-GS1_files/
Test/
Images/
drishtiGS_{:03}.png # numbers overlap with train
"""
result_resolution = (2040, 2040)
set_path = os.path.join(db_folder, 'Drishti-GS1_files', 'Test')
images_path = os.path.join(set_path, 'Images')
X_all, file_names = imh.load_set(images_path)
rel_file_names = [os.path.split(fn)[-1] for fn in file_names]
rel_file_names_wo_ext = [fn[:fn.rfind('.')] for fn in rel_file_names]
file_codes = ['Test' + fn[fn.find('_'):] for fn in rel_file_names_wo_ext]
for i in xrange(len(X_all)):
side = result_resolution[0]
X_all[i] = imh.resize_image_to_square(X_all[i], side, pad_cval=0)
return X_all, file_codes
def extract_RIM_ONE_v2(db_folder):
"""
Cropped (to optic disc region) images with polygonal optic disc segmentation.
380 x 394 original, 394 x 394 after post-processing.
Required schema:
db_folder/
Normal/
im{:03}.jpg (number from 0 to 255)
im{:03}_gs.txt
Glaucoma and glaucoma suspicious/
im{:03}.jpg (number from 256 to 455)
im{:03}_gs.txt
"""
orig_resolution = (380, 394)
result_resolution = (394, 394)
X_all, Y_all, filecodes_all, is_ill = [], [], [], []
for pic_type in ('Normal', 'Glaucoma and glaucoma suspicious'):
X, filenames = imh.load_set(os.path.join(db_folder, pic_type))
file_codes = [fn[-7:-4] for fn in filenames]
Y = []
for i, code in enumerate(file_codes):
anot_filename = os.path.join(db_folder, pic_type,
'Im{}-gs.txt'.format(code))
with open(anot_filename) as anot_fin:
lines = anot_fin.readlines()
'''
# polygonal segmentation
coords = lines[1:lines.index('Ellipse parameters\r\n')]
coords = np.array(map(int, coords))
if coords.size % 2 != 0:
raise imh.ImLibException('n_coords % 2 != 0')
coords = coords.reshape((coords.size / 2, 2))
coords = coords[1:] # optic disc center point is included in annotation for some reason
segm_img = np.zeros(orig_resolution, dtype=np.uint8)
cv2.fillPoly(segm_img, coords.reshape((1,) + coords.shape), color=1)
'''
'''
# ellipse segmentation
coords = lines[lines.index('Ellipse parameters\r\n') + 1:]
coords = map(int, coords)
i0, j0, a, b, angle = coords
a /= 2
b /= 2
segm_img = np.zeros(orig_resolution, dtype=np.uint8)
cv2.ellipse(segm_img, (i0, j0), (a, b), angle, 0, 360, color=1, thickness=-1)
'''
# acquiring segmentation from pre-computed image
segm_img = imh.load_image(
os.path.join(db_folder, pic_type + ' segmentation',
'Im{}-gs_mask.jpg'.format(code)))
Y.append(segm_img)
is_ill.append(HEALTHY if pic_type ==
'Normal' else GLAUCOMA_OR_SUSPECT)
for i in xrange(len(X)):
side = result_resolution[0]
X[i] = imh.resize_image_to_square(X[i], side, pad_cval=0)
Y[i] = imh.resize_image_to_square(Y[i], side, pad_cval=0)
Y[i] = Y[i].reshape(Y[i].shape + (1, ))
X_all.extend(X)
Y_all.extend(Y)
filecodes_all.extend(file_codes)
return X_all, Y_all, filecodes_all, is_ill
def extract_DRISHTI_GS_train(db_folder, return_disc=True, return_cup=True):
"""
Full images with optic disc and optic cup segmentation.
Average segmentation and "softmap" segmentation image are given.
50 images of various resolution close to 2040 x 1740.
Data set is split into training and test sets. Groundtruth is available for training set only.
This function returns Training set only.
Required schema:
db_folder/
Drishti-GS1_files/
Training/
Images/
drishtiGS_{:03}.png # some numbers are omitted, like 001, 003, 004, ...
GT/
drishtiGS_{:03}/
drishtiGS_{:03}_cdrValues.txt
AvgBoundary/
drishtiGS_{:03}_ODAvgBoundary.txt
drishtiGS_{:03}_CupAvgBoundary.txt
drishtiGS_{:03}_diskCenter.txt
SoftMap/
drishtiGS_{:03}_ODsegSoftmap.png
drishtiGS_{:03}_cupsegSoftmap.png
"""
result_resolution = (2040, 2040)
disc_all, cup_all, file_codes_all = [], [], []
set_path = os.path.join(db_folder, 'Drishti-GS1_files', 'Training')
images_path = os.path.join(set_path, 'Images')
X_all, file_names = imh.load_set(images_path)
rel_file_names = [os.path.split(fn)[-1] for fn in file_names]
rel_file_names_wo_ext = [fn[:fn.rfind('.')] for fn in rel_file_names]
file_codes = [
'Training' + fn[fn.find('_'):] for fn in rel_file_names_wo_ext
]
file_codes_all.extend(file_codes)
for fn in rel_file_names_wo_ext:
if return_disc:
disc_segmn = imh.load_image(
os.path.join(set_path, 'GT', fn, 'SoftMap',
fn + '_ODsegSoftmap.png'))
disc_all.append(disc_segmn)
if return_cup:
cup_segmn = imh.load_image(
os.path.join(set_path, 'GT', fn, 'SoftMap',
fn + '_cupsegSoftmap.png'))
cup_all.append(cup_segmn)
for i in xrange(len(X_all)):
side = result_resolution[0]
X_all[i] = imh.resize_image_to_square(X_all[i], side, pad_cval=0)
if return_disc:
disc_all[i] = imh.resize_image_to_square(disc_all[i],
side,
pad_cval=0)
disc_all[i] = disc_all[i].reshape(disc_all[i].shape + (1, ))
if return_cup:
cup_all[i] = imh.resize_image_to_square(cup_all[i],
side,
pad_cval=0)
cup_all[i] = cup_all[i].reshape(cup_all[i].shape + (1, ))
if return_disc:
if return_cup:
return X_all, disc_all, cup_all, file_codes_all
return X_all, disc_all, file_codes_all
if return_cup:
return X_all, cup_all, file_codes_all
return X_all, file_codes_all
def extract_RIM_ONE_v3(db_folder,
expert='avg',
return_disc=True,
return_cup=True):
"""
Cropped (to optic disc region, and a little more by vertical axis) images
with polygonal optic disc segmentation. 1424 x 2144 original, 1424 x 1424 after post-processing.
Images are two-channel (stereo) --- caught from 2 angles.
But segmentation is given for only one view (see L/R letter in file name for clarification).
So only one view of two is chosen.
Accepted values for `expert`: 1, 2, 'avg'.
Required schema:
db_folder/
Healthy/
Stereo Images/
N-{}-[L,R].jpg (number is without leading zeros, from 1 to 92)
(image cannot be used as is. it is two-part image, divided by vertical border)
Expert1_masks/
N-{}-[L,R]-Cup-exp1.png (4 files for one image number and L/R characteristic)
N-{}-[L,R]-Cup-exp1.txt
N-{}-[L,R]-Disc-exp1.png
N-{}-[L,R]-Disc-exp1.txt
Expert2_masks/
N-{}-[L,R]-Cup-exp2.png (4 files for one image number and L/R characteristic)
N-{}-[L,R]-Cup-exp2.txt
N-{}-[L,R]-Disc-exp2.png
N-{}-[L,R]-Disc-exp2.txt
Average_masks/
N-{}-[L,R]-Cup-Avg.png (4 files for one image number and L/R characteristic)
N-{}-[L,R]-Cup-Avg.txt
N-{}-[L,R]-Disc-Avg.png
N-{}-[L,R]-Disc-Avg.txt
Glaucoma and suspects/
(...) (the same as for Healthy, but images start with G not N)
"""
orig_resolution = (1424, 2144)
result_resolution = (1424, 1424)
if expert == 1:
expert_folder = 'Expert1_masks'
suffix = 'exp1'
elif expert == 2:
expert_folder = 'Expert2_masks'
suffix = 'exp2'
elif expert == 'avg':
expert_folder = 'Average_masks'
suffix = 'Avg'
else:
raise imh.ImLibException('value for "expert" argument not understood')
X_all, disc_all, cup_all, file_codes_all, is_ill = [], [], [], [], []
for pic_type in ('Healthy', 'Glaucoma and suspects'):
X, file_names = imh.load_set(
os.path.join(db_folder, pic_type, 'Stereo Images'))
X_all.extend(X)
rel_file_names = [os.path.split(fn)[-1] for fn in file_names]
file_codes = [fn[:fn.rfind('.')] for fn in rel_file_names]
file_codes_all.extend(file_codes)
for fc in file_codes:
if return_disc:
disc_segmn = imh.load_image(
os.path.join(db_folder, pic_type, expert_folder,
'{}-Disc-{}.png'.format(fc, suffix)))
disc_all.append(disc_segmn)
if return_cup:
cup_segmn = imh.load_image(
os.path.join(db_folder, pic_type, expert_folder,
'{}-Cup-{}.png'.format(fc, suffix)))
cup_all.append(cup_segmn)
is_ill.append(HEALTHY if pic_type ==
'Healthy' else GLAUCOMA_OR_SUSPECT)
for i in xrange(len(X_all)):
side = result_resolution[0]
if file_codes_all[i][-1] == 'L':
X_all[i] = X_all[i][:, :orig_resolution[1] / 2]
elif file_codes_all[i][-1] == 'R':
X_all[i] = X_all[i][:, orig_resolution[1] / 2:]
if return_disc:
disc_all[i] = disc_all[i][:, :orig_resolution[1] / 2]
if return_cup:
cup_all[i] = cup_all[i][:, :orig_resolution[1] / 2]
else:
raise imh.ImLibException(
'image {} has no L/R characteristic'.format(file_codes_all[i]))
X_all[i] = imh.resize_image_to_square(X_all[i], side, pad_cval=0)
if return_disc:
disc_all[i] = imh.resize_image_to_square(disc_all[i],
side,
pad_cval=0)
disc_all[i] = disc_all[i].reshape(disc_all[i].shape + (1, ))
if return_cup:
cup_all[i] = imh.resize_image_to_square(cup_all[i],
side,
pad_cval=0)
cup_all[i] = cup_all[i].reshape(cup_all[i].shape + (1, ))
if return_disc:
if return_cup:
return X_all, disc_all, cup_all, file_codes_all, is_ill
return X_all, disc_all, file_codes_all, is_ill
if return_cup:
return X_all, cup_all, file_codes_all, is_ill
return X_all, file_codes_all, is_ill
def train(self, epochs, batch_size=128, save_interval=50):
# Load the dataset
# (X_train, _), (_, _) = mnist.load_data()
# data = np.load('mnist.npz')
# X_train=data['x_train.npy']
# print(X_train.shape)
side = 128 #28
filelist = glob.glob('data/images_001/images/*.png')
XX = np.empty([side, side, 1])
for fname in filelist:
X_t = np.array(Image.open(fname))
if X_t.ndim == 2:
X_t = imh.resize_image_to_square(X_t, side, pad_cval=0)
X_train = X_t.reshape(side, side, 1)
else:
# print(X_t.shape)
X_train = X_t[:, :, 1]
X_train = imh.resize_image_to_square(X_train, side, pad_cval=0)
X_train = X_train.reshape(side, side, 1)
XX = np.concatenate((XX, X_train), axis=2)
X_train = XX.reshape(XX.shape[2], XX.shape[0], XX.shape[1])
X_train = X_train.astype(np.uint8)
# print(X_train.shape)
# from tensorflow.examples.tutorials.mnist import input_data
# mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
# (X_train, _), (_, _) = mnist.load_data()
# Rescale -1 to 1
X_train = (X_train.astype(np.float32) - 127.5) / 127.5
X_train = np.expand_dims(X_train, axis=3)
half_batch = int(batch_size / 2)
for epoch in range(epochs):
# ---------------------
# Train Discriminator
# ---------------------
# Select a random half batch of images
idx = np.random.randint(0, X_train.shape[0], half_batch)
imgs = X_train[idx]
noise = np.random.normal(0, 1, (half_batch, 100))
# Generate a half batch of new images
gen_imgs = self.generator.predict(noise)
# Train the discriminator
d_loss_real = self.discriminator.train_on_batch(
imgs, np.ones((half_batch, 1)))
d_loss_fake = self.discriminator.train_on_batch(
gen_imgs, np.zeros((half_batch, 1)))
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
# ---------------------
# Train Generator
# ---------------------
noise = np.random.normal(0, 1, (batch_size, 100))
# The generator wants the discriminator to label the generated samples
# as valid (ones)
valid_y = np.array([1] * batch_size)
# Train the generator
g_loss = self.combined.train_on_batch(noise, valid_y)
# Plot the progress
print("%d [D loss: %f, acc.: %.2f%%] [G loss: %f]" %
(epoch, d_loss[0], 100 * d_loss[1], g_loss))
# If at save interval => save generated image samples
if epoch % save_interval == 0:
self.save_imgs(epoch)