def __init__(self):
self.num_channels = 1
self.datagen = ImageDataGenerator(rotation_range=[5., 5., 5.],
zoom_range=[0.9, 1.1],
width_shift_range=0.1,
height_shift_range=0.1,
depth_shift_range=0.1,
horizontal_flip=False,
fill_mode='constant',
cval=0.)
class Preprocess(object):
def __init__(self):
self.num_channels = 1
self.datagen = ImageDataGenerator(rotation_range=[5., 5., 5.],
zoom_range=[0.9, 1.1],
width_shift_range=0.1,
height_shift_range=0.1,
depth_shift_range=0.1,
horizontal_flip=False,
fill_mode='constant',
cval=0.)
def sample_z_norm(self, img):
for self.num_channel in range(self.num_channels):
img[..., self.num_channel] -= np.mean(img[..., self.num_channel])
img[..., self.num_channel] /= np.std(img[..., self.num_channel])
# print('sample z norm img shape :',img.shape)
return img
def simple_preprocess_img(self, img):
img = np.expand_dims(img, -1)
# print('preprocessed img shape :',img.shape)
return img
def simple_sample_z_norm(self, img):
for self.num_channel in range(self.num_channels):
img[..., self.num_channel] -= np.mean(img[..., self.num_channel])
img[..., self.num_channel] /= np.std(img[..., self.num_channel])
return img
def simple_preprocess_mask(self, mask):
#print('before preprocessd mask shape :',mask.shape)
# value, count = np.unique(mask, return_counts=True)
# print('image value : ' + str(value) + ' image Count : ' + str(count))
mask_dummy = np.zeros(mask.shape + (3, ))
mask_dummy[:, :, :, 0][mask == 1] = 1.
mask_dummy[:, :, :, 1][mask == 2] = 1.
mask_dummy[:, :, :, 2][mask == 3] = 1.
#print('processed mask shape :',mask_dummy.shape)
# value, count = np.unique(mask_dummy[:, :, :, :], return_counts=True)
# print('image value : ' + str(value) + ' image Count : ' + str(count))
return mask_dummy
def whole_merge_mask(self, mask):
mask = np.sum(mask, axis=-1)
mask = np.expand_dims(mask, axis=-1)
# print('merged mask shape :',mask.shape)
return mask
def resize_whole(self, img_mask):
image = img_mask[0].copy()
label = img_mask[1].copy()
max_size = 180
check_size = (image.shape[0] / max_size) * (
image.shape[1] / max_size) * (image.shape[2] / max_size)
if check_size > 1.:
reduction_factor = (
(max_size * max_size * max_size) /
(image.shape[0] * image.shape[1] * image.shape[2]))**(1 / 3)
resize_factor = (reduction_factor, reduction_factor,
reduction_factor)
image = ndimage.zoom(image,
resize_factor + (1., ),
order=0,
mode='constant',
cval=0.0)
label = ndimage.zoom(label,
resize_factor + (1., ),
order=0,
mode='constant',
cval=0.0)
image = np.expand_dims(image, 0)
label = np.expand_dims(label, 0)
return [image, label]
def crop_to_patch(self, img_mask):
mask_dummy = img_mask[1].copy()
mask_dummy = np.sum(mask_dummy, axis=-1)
coordinates = np.argwhere(mask_dummy != 0)
start_end = [[np.min(coordinates[:, 0]),
np.max(coordinates[:, 0])],
[np.min(coordinates[:, 1]),
np.max(coordinates[:, 1])],
[np.min(coordinates[:, 2]),
np.max(coordinates[:, 2])]]
x_start = start_end[0][0]
x_end = start_end[0][1]
y_start = start_end[1][0]
y_end = start_end[1][1]
z_start = start_end[2][0]
z_end = start_end[2][1]
# define margin for minimum size
min_size = 24
if x_end - x_start < min_size:
x_margin = math.ceil((min_size - (x_end - x_start)) / 2)
if x_start - x_margin < 0:
x_start = 0
else:
x_start -= x_margin
if x_end + x_margin > mask_dummy.shape[0]:
x_end = mask_dummy.shape[0]
else:
x_end += x_margin
if y_end - y_start < min_size:
y_margin = math.ceil((min_size - (y_end - y_start)) / 2)
if y_start - y_margin < 0:
y_start = 0
else:
y_start -= y_margin
if y_end + y_margin > mask_dummy.shape[1]:
y_end = mask_dummy.shape[1]
else:
y_end += y_margin
if z_end - z_start < min_size:
z_margin = math.ceil((min_size - (z_end - z_start)) / 2)
if z_start - z_margin < 0:
z_start = 0
else:
z_start -= z_margin
if z_end + z_margin > mask_dummy.shape[2]:
z_end = mask_dummy.shape[2]
else:
z_end += z_margin
# define margin 20% of object size
x_margin = math.ceil((x_end - x_start) * 0.2)
y_margin = math.ceil((y_end - y_start) * 0.2)
z_margin = math.ceil((z_end - z_start) * 0.2)
if x_start - x_margin < 0:
x_start = 0
else:
x_start -= x_margin
if x_end + x_margin > mask_dummy.shape[0]:
x_end = mask_dummy.shape[0]
else:
x_end += x_margin
if y_start - y_margin < 0:
y_start = 0
else:
y_start -= y_margin
if y_end + y_margin > mask_dummy.shape[1]:
y_end = mask_dummy.shape[1]
else:
y_end += y_margin
if z_start - z_margin < 0:
z_start = 0
else:
z_start -= z_margin
if z_end + z_margin > mask_dummy.shape[2]:
z_end = mask_dummy.shape[2]
else:
z_end += z_margin
# print("x~z start end point :",x_start,x_end,y_start,y_end,z_start,z_end)
# print("crop image shape :",img_mask[0].shape)
# print("crop mask shape :",img_mask[1].shape)
img_mask[0] = img_mask[0][x_start:x_end, y_start:y_end,
z_start:z_end, :]
img_mask[1] = img_mask[1][x_start:x_end, y_start:y_end,
z_start:z_end, :]
return img_mask
def resize_patch(self, img_mask):
crop_image = img_mask[0].copy()
crop_label = img_mask[1].copy()
resize_factor = ()
for num_axis in range(3):
if crop_image.shape[num_axis] < 24:
resize_factor += (24. / crop_image.shape[num_axis], )
else:
resize_factor += (1., )
if resize_factor != (1., 1., 1.):
crop_image = ndimage.zoom(crop_image,
resize_factor + (1., ),
order=0,
mode='constant',
cval=0.0)
crop_label = ndimage.zoom(crop_label,
resize_factor + (1., ),
order=0,
mode='constant',
cval=0.0)
max_size = 180
check_size = (crop_image.shape[0] / max_size) * (
crop_image.shape[1] / max_size) * (crop_image.shape[2] / max_size)
if check_size > 1.:
reduction_factor = ((max_size * max_size * max_size) /
(crop_image.shape[0] * crop_image.shape[1] *
crop_image.shape[2]))**(1 / 3)
resize_factor = (reduction_factor, reduction_factor,
reduction_factor)
crop_image = ndimage.zoom(crop_image,
resize_factor + (1., ),
order=0,
mode='constant',
cval=0.0)
crop_label = ndimage.zoom(crop_label,
resize_factor + (1., ),
order=0,
mode='constant',
cval=0.0)
crop_image = np.expand_dims(crop_image, 0)
crop_label = np.expand_dims(crop_label, 0)
return [crop_image, crop_label]
def data_aug(self, img_mask):
seed = np.random.randint(0, 100, 1)[0]
for img in self.datagen.flow(img_mask[0], batch_size=1, seed=seed):
break
for mask in self.datagen.flow(img_mask[1], batch_size=1, seed=seed):
break
print('-------------------------------------------------')
print('input image shape :', img.shape)
print('input label shape :', mask.shape)
print('-------------------------------------------------')
return [img, mask]