|
1
|
|
|
from collections import defaultdict |
|
2
|
|
|
from typing import Union, Tuple, Optional, List, Dict |
|
3
|
|
|
|
|
4
|
|
|
import torch |
|
5
|
|
|
import numpy as np |
|
6
|
|
|
import scipy.ndimage as ndi |
|
7
|
|
|
|
|
8
|
|
|
from ....utils import to_tuple |
|
9
|
|
|
from ....torchio import DATA, TypeData, TypeTripletFloat, TypeSextetFloat |
|
10
|
|
|
from ....data.subject import Subject |
|
11
|
|
|
from ... import IntensityTransform |
|
12
|
|
|
from .. import RandomTransform |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
class RandomBlur(RandomTransform, IntensityTransform): |
|
16
|
|
|
r"""Blur an image using a random-sized Gaussian filter. |
|
17
|
|
|
|
|
18
|
|
|
Args: |
|
19
|
|
|
std: Tuple :math:`(a_1, b_1, a_2, b_2, a_3, b_3)` representing the |
|
20
|
|
|
ranges (in mm) of the standard deviations |
|
21
|
|
|
:math:`(\sigma_1, \sigma_2, \sigma_3)` of the Gaussian kernels used |
|
22
|
|
|
to blur the image along each axis, where |
|
23
|
|
|
:math:`\sigma_i \sim \mathcal{U}(a_i, b_i)`. |
|
24
|
|
|
If two values :math:`(a, b)` are provided, |
|
25
|
|
|
then :math:`\sigma_i \sim \mathcal{U}(a, b)`. |
|
26
|
|
|
If only one value :math:`x` is provided, |
|
27
|
|
|
then :math:`\sigma_i \sim \mathcal{U}(0, x)`. |
|
28
|
|
|
If three values :math:`(x_1, x_2, x_3)` are provided, |
|
29
|
|
|
then :math:`\sigma_i \sim \mathcal{U}(0, x_i)`. |
|
30
|
|
|
p: Probability that this transform will be applied. |
|
31
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
32
|
|
|
""" |
|
33
|
|
|
def __init__( |
|
34
|
|
|
self, |
|
35
|
|
|
std: Union[float, Tuple[float, float]] = (0, 2), |
|
36
|
|
|
p: float = 1, |
|
37
|
|
|
keys: Optional[List[str]] = None, |
|
38
|
|
|
): |
|
39
|
|
|
super().__init__(p=p, keys=keys) |
|
40
|
|
|
self.std_ranges = self.parse_params(std, None, 'std', min_constraint=0) |
|
41
|
|
|
|
|
42
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
43
|
|
|
arguments = defaultdict(dict) |
|
44
|
|
|
for name, image in self.get_images_dict(subject).items(): |
|
45
|
|
|
stds = [self.get_params(self.std_ranges) for _ in image.data] |
|
46
|
|
|
arguments['std'][name] = stds |
|
47
|
|
|
transform = Blur(**arguments) |
|
48
|
|
|
transformed = transform(subject) |
|
49
|
|
|
return transformed |
|
50
|
|
|
|
|
51
|
|
|
def get_params(self, std_ranges: TypeSextetFloat) -> TypeTripletFloat: |
|
52
|
|
|
std = self.sample_uniform_sextet(std_ranges) |
|
53
|
|
|
return std |
|
54
|
|
|
|
|
55
|
|
|
|
|
56
|
|
|
class Blur(IntensityTransform): |
|
57
|
|
|
r"""Blur an image using a Gaussian filter. |
|
58
|
|
|
|
|
59
|
|
|
Args: |
|
60
|
|
|
std: Tuple :math:`(\sigma_1, \sigma_2, \sigma_3)` representing the |
|
61
|
|
|
the standard deviations (in mm) of the standard deviations |
|
62
|
|
|
of the Gaussian kernels used to blur the image along each axis. |
|
63
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
64
|
|
|
""" |
|
65
|
|
|
def __init__( |
|
66
|
|
|
self, |
|
67
|
|
|
std: Union[TypeTripletFloat, Dict[str, TypeTripletFloat]], |
|
68
|
|
|
keys: Optional[List[str]] = None, |
|
69
|
|
|
): |
|
70
|
|
|
super().__init__(keys=keys) |
|
71
|
|
|
self.std = std |
|
72
|
|
|
self.args_names = ('std',) |
|
73
|
|
|
|
|
74
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
75
|
|
|
std = self.std |
|
76
|
|
|
for name, image in self.get_images_dict(subject).items(): |
|
77
|
|
|
if self.arguments_are_dict(): |
|
78
|
|
|
std = self.std[name] |
|
79
|
|
|
stds = to_tuple(std, length=len(image.data)) |
|
80
|
|
|
transformed_tensors = [] |
|
81
|
|
|
for std, tensor in zip(stds, image.data): |
|
82
|
|
|
transformed_tensor = blur( |
|
83
|
|
|
tensor, |
|
84
|
|
|
image.spacing, |
|
85
|
|
|
std, |
|
86
|
|
|
) |
|
87
|
|
|
transformed_tensors.append(transformed_tensor) |
|
88
|
|
|
image[DATA] = torch.stack(transformed_tensors) |
|
89
|
|
|
return subject |
|
90
|
|
|
|
|
91
|
|
|
|
|
92
|
|
|
def blur( |
|
93
|
|
|
data: TypeData, |
|
94
|
|
|
spacing: TypeTripletFloat, |
|
95
|
|
|
std_voxel: TypeTripletFloat, |
|
96
|
|
|
) -> torch.Tensor: |
|
97
|
|
|
assert data.ndim == 3 |
|
98
|
|
|
std_physical = np.array(std_voxel) / np.array(spacing) |
|
99
|
|
|
blurred = ndi.gaussian_filter(data, std_physical) |
|
100
|
|
|
tensor = torch.from_numpy(blurred) |
|
101
|
|
|
return tensor |
|
102
|
|
|
|
fepegar /
torchio
