def get_validation_augmentation(augmentation_key):
"""
Validation data augmentations. Usually, just cropping.
"""
transform_dict = {
"aug1": [
bg.RandomCropTransform(crop_size=(80, 160, 160))
],
"aug2": [
bg.RandomCropTransform(crop_size=(96, 160, 160))
],
"aug3": [
bg.RandomCropTransform(crop_size=(96, 160, 160))
],
"aug4": [
bg.RandomCropTransform(crop_size=(96, 160, 160))
],
"aug5": [
ROICropTransform(crop_size=(96, 160, 160))
],
"aug6": [
bg.RandomCropTransform(crop_size=(192, 192))
],
"aug7": [
bg.CenterCropTransform(crop_size=(256, 256))
],
"tu_only2d": [
bg.CenterCropTransform(crop_size=(256, 256)),
bgut.RemoveLabelTransform(1, 0)
],
"tu_only2d2": [
bg.CenterCropTransform(crop_size=(256, 256)),
MultiClassToBinaryTransform(roi_label="2", remove_label="1"),
],
"tu_only2d3": [
bg.CenterCropTransform(crop_size=(256, 256)),
MultiClassToBinaryTransform(roi_label="2", remove_label="1"),
],
}
test_transform = transform_dict[augmentation_key]
print(f"\nTest/Validation Transforms: {test_transform}")
return bg.Compose(test_transform)
def get_preprocessing(rgb: bool = False):
"""
Construct preprocessing transform
Args:
rgb (bool): Whether or not to return the input with three channels
or just single (grayscale)
Return:
transform: albumentations.Compose
"""
_transform = [
bgct.ClipValueRange(min=-79, max=304),
bgsnt.MeanStdNormalizationTransform(mean=101, std=76.9,
per_channel=False),
bgut.RemoveLabelTransform(-1, 0),
bg.NumpyToTensor(),
]
if rgb:
# insert right before converting to a torch tensor
_transform.insert(-2, RepeatChannelsTransform(num_repeats=3))
print(f"\nPreprocessing Transforms: {_transform}")
return bg.Compose(_transform)
def get_training_augmentation(augmentation_key="aug1"):
default_angle = (-15. / 360 * 2. * np.pi, 15. / 360 * 2. * np.pi)
aug1_spatial_kwargs = {"patch_size": (80, 160, 160),
"patch_center_dist_from_border": (30, 30, 30),
"do_elastic_deform": True,
"alpha": (0., 900.),
"sigma": (9., 13.),
"do_rotation": True,
"angle_x": default_angle,
"angle_y": default_angle,
"angle_z": default_angle,
"do_scale": True,
"scale": (0.85, 1.25),
"border_mode_data": "constant",
"order_data": 3,
"random_crop": True,
"p_el_per_sample": 0.2,
"p_scale_per_sample": 0.2,
"p_rot_per_sample": 0.2
}
transform_dict = {
"aug1": [
bg.SpatialTransform(**aug1_spatial_kwargs),
bg.MirrorTransform(axes=(0, 1, 2)),
bg.GammaTransform(gamma_range=(0.7, 1.5),
invert_image=False,
per_channel=True,
retain_stats=True,
p_per_sample=0.3),
],
}
# aug2
aug2_spatial_kwargs = deepcopy(aug1_spatial_kwargs)
aug2_spatial_kwargs["patch_size"] = (96, 160, 160)
transform_dict["aug2"] = [bg.SpatialTransform(**aug2_spatial_kwargs),] \
+ transform_dict["aug1"][1:]
# aug3
aug3_spatial_kwargs = deepcopy(aug2_spatial_kwargs)
aug3_spatial_kwargs["random_crop"] = False
new_transforms = [bg.SpatialTransform(**aug3_spatial_kwargs),
bg.BrightnessTransform(mu=101, sigma=76.9,
p_per_sample=0.3),]
# spatial, mirror, gamma, brightness
transform_dict["aug3"] = [new_transforms[0]] + transform_dict["aug1"][1:] \
+ [new_transforms[1]]
# aug4
# roicrop, spatial, mirror, gamma, brightness
transform_dict["aug4"] = [ROICropTransform(crop_size=(96, 160, 160)),] + \
transform_dict["aug3"]
# aug5
# roicrop, spatial, mirror, gamma, brightness
aug5_spatial_kwargs = deepcopy(aug3_spatial_kwargs)
aug5_spatial_kwargs["patch_center_dist_from_border"] = None
aug5_spatial_kwargs["border_cval_data"] = 0
new_transforms = [ROICropTransform(crop_size=(96, 160, 160)),
bg.SpatialTransform(**aug5_spatial_kwargs),]
# RemoveLabelTransform added to preprocessing
transform_dict["aug5"] = new_transforms + transform_dict["aug4"][2:]
# 2D Transforms
# spatial, mirror, gamma, brightness
aug6_spatial_kwargs = deepcopy(aug3_spatial_kwargs)
aug6_spatial_kwargs["patch_size"] = (192, 192)
transforms_2d = [bg.SpatialTransform(**aug6_spatial_kwargs),
bg.MirrorTransform(axes=(0, 1)),]
transform_dict["aug6"] = transforms_2d + transform_dict["aug3"][2:]
# spatial, mirror, gamma, brightness
aug7_spatial_kwargs = deepcopy(aug3_spatial_kwargs)
aug7_spatial_kwargs["patch_size"] = (256, 256)
transforms_2d = [bg.SpatialTransform(**aug7_spatial_kwargs),
bg.MirrorTransform(axes=(0, 1)),]
transform_dict["aug7"] = transforms_2d + transform_dict["aug3"][2:]
# spatial, mirror, gamma, brightness, removelabel
tu_only = [MultiClassToBinaryTransform(roi_label="2", remove_label="1")]
transform_dict["tu_only2d"] = transform_dict["aug7"] + tu_only
tu_only2d2_spatial_kwargs = deepcopy(aug7_spatial_kwargs)
tu_only2d2_spatial_kwargs["p_scale_per_sample"] = 0.5
tu_only2d2_spatial_kwargs["p_rot_per_sample"] = 0.5
new_t = [bg.SpatialTransform(**tu_only2d2_spatial_kwargs),
bg.BrightnessTransform(mu=101, sigma=76.9, p_per_sample=0.5)]
transform_dict["tu_only2d2"] = [new_t[0]] + transform_dict["tu_only2d"][1:-2] \
+ [new_t[1]] + [transform_dict["tu_only2d"][-1]]
# adding RandomResizedCropTransform to the end
# spatial, mirror, gamma, brightness, random_resized_crop, removelabel
rand_resized_t = RandomResizedCropTransform(target_size=(256, 256),
p_per_sample=0.33)
transform_dict["tu_only2d3"] = deepcopy(transform_dict["tu_only2d2"])
transform_dict["tu_only2d3"].insert(-2, rand_resized_t)
train_transform = transform_dict[augmentation_key]
print(f"Train Transforms: {train_transform}")
return bg.Compose(train_transform)