def generate_training_data(number_of_samples, num_noise_sample_percentage=0.0):
number_of_noise_samples = int(number_of_samples *
num_noise_sample_percentage)
number_of_circles = number_of_samples - number_of_noise_samples
growth_rate = np.min(GEOMETRY.volume_shape) / number_of_circles / 2
data = np.empty((number_of_samples, ) + tuple(GEOMETRY.sinogram_shape))
labels = np.empty((number_of_samples, ) + tuple(GEOMETRY.volume_shape))
with tf.Session() as sess:
# build growing circles until volume is filled
for i in range(number_of_circles):
pos = GEOMETRY.volume_shape // 2 # middle
radius = int(growth_rate * (i + 1))
value = 1.0 #np.random.uniform(0.01, 1.0) # here one could also try random values for each circle
labels[i] = primitives_2d.circle(GEOMETRY.volume_shape, pos,
radius, value)
# project it
data[i] = generate_sinogram_parallel_2d(labels[i], GEOMETRY)
# add some noise only phantoms
for i in range(number_of_samples - number_of_noise_samples,
number_of_samples):
labels[i] = np.random.uniform(0.0, 1.0, GEOMETRY.volume_shape)
data[i] = generate_sinogram_parallel_2d(labels[i], GEOMETRY)
return data, labels
def get_test_data(number_of_samples=1):
data = np.empty((number_of_samples, ) + tuple(GEOMETRY.sinogram_shape))
labels = np.empty((number_of_samples, ) + tuple(GEOMETRY.volume_shape))
with tf.Session() as sess:
# get shepp logan 2d
if number_of_samples == 1:
labels[0] = shepp_logan.shepp_logan_enhanced(GEOMETRY.volume_shape)
data[0] = generate_sinogram_parallel_2d(labels[0], GEOMETRY)
# every slice of shepp logan 3d with number_of_samples as Z-dimension as own image
else:
labels = shepp_logan.shepp_logan_3d((number_of_samples, ) +
tuple(GEOMETRY.sinogram_shape))
for i in range(number_of_samples):
data[i] = generate_sinogram_parallel_2d(labels[i], GEOMETRY)
return data, labels
def get_test_cupping_data():
data = np.empty((1, ) + tuple(GEOMETRY.sinogram_shape), dtype=np.float32)
labels = np.empty((1, ) + tuple(GEOMETRY.volume_shape), dtype=np.float32)
labels[0] = primitives_2d.circle(GEOMETRY.volume_shape,
GEOMETRY.volume_shape // 2,
np.min(GEOMETRY.volume_shape // 2))
data[0] = generate_sinogram_parallel_2d(np.expand_dims(labels[0], axis=0),
GEOMETRY)
return data, labels
def get_test_cupping_data():
data = np.empty((1, ) + tuple(GEOMETRY.sinogram_shape))
labels = np.empty((1, ) + tuple(GEOMETRY.volume_shape))
with tf.Session() as sess:
labels[0] = primitives_2d.circle(GEOMETRY.volume_shape,
GEOMETRY.volume_shape // 2,
np.min(GEOMETRY.volume_shape // 2))
data[0] = generate_sinogram_parallel_2d(labels[0], GEOMETRY)
return data, labels
def generate_training_data():
label_list = []
input_data_list = []
max_radius = np.min(GEOMETRY.volume_shape) // 2
center_pos = [(GEOMETRY.volume_shape[0]-1)//2, (GEOMETRY.volume_shape[1]-1)//2]
#Compute phantom
for n in range(9, max_radius, 2):
#Add batch dimension with dim == 1 for sinogram generation
phantom = np.expand_dims( primitives_2d.circle(GEOMETRY.volume_shape, center_pos, n), axis = 0)
label_list.append(phantom)
#Create sinogram data
with tf.Session() as sess:
for phantom in label_list:
sinogram = generate_sinogram_parallel_2d(phantom, GEOMETRY)
input_data_list.append(sinogram)
#Remove batch dimension
return np.squeeze(np.asarray(input_data_list)), np.squeeze(np.asarray(label_list))
def generate_validation_data(number_of_samples):
growth_rate = np.min(GEOMETRY.volume_shape) / number_of_samples / 2
data = np.empty((number_of_samples,) + tuple(GEOMETRY.sinogram_shape))
labels = np.empty((number_of_samples,) + tuple(GEOMETRY.volume_shape))
with tf.Session() as sess:
# build growing rectangles
for i in range(number_of_samples):
size = np.array([growth_rate * (i+1), growth_rate * (i+1)], dtype=np.int32)
pos = GEOMETRY.volume_shape//2 - size//2 # middle
value = np.random.uniform(0.01, 1.0)
labels[i] = primitives_2d.rect(GEOMETRY.volume_shape, pos, size, value)
# project it
data[i] = generate_sinogram_parallel_2d(np.expand_dims(labels[i],axis=0), GEOMETRY)
return data, labels
def get_test_data(number_of_samples=1):
data = np.empty((number_of_samples, ) + tuple(GEOMETRY.sinogram_shape),
dtype=np.float32)
labels = np.empty((number_of_samples, ) + tuple(GEOMETRY.volume_shape),
dtype=np.float32)
# get shepp logan 2d
if number_of_samples == 1:
labels[0] = shepp_logan.shepp_logan_enhanced(GEOMETRY.volume_shape)
data[0] = generate_sinogram(np.expand_dims(labels[0], axis=0),
parallel_projection2d, GEOMETRY)
# every slice of shepp logan 3d with number_of_samples as Z-dimension as own image
else:
labels = shepp_logan.shepp_logan_3d((number_of_samples, ) +
tuple(GEOMETRY.sinogram_shape))
for i in range(number_of_samples):
data[i] = generate_sinogram_parallel_2d(
np.expand_dims(labels[i], axis=0), GEOMETRY)
return data, labels