|
1
|
|
|
from typing import Union, Tuple, Optional, List |
|
2
|
|
|
import torch |
|
3
|
|
|
import numpy as np |
|
4
|
|
|
from ....torchio import DATA |
|
5
|
|
|
from ....data.subject import Subject |
|
6
|
|
|
from ....utils import to_tuple |
|
7
|
|
|
from ... import SpatialTransform |
|
8
|
|
|
from .. import RandomTransform |
|
9
|
|
|
|
|
10
|
|
|
|
|
11
|
|
|
class RandomFlip(RandomTransform, SpatialTransform): |
|
12
|
|
|
"""Reverse the order of elements in an image along the given axes. |
|
13
|
|
|
|
|
14
|
|
|
Args: |
|
15
|
|
|
axes: Index or tuple of indices of the spatial dimensions along which |
|
16
|
|
|
the image might be flipped. If they are integers, they must be in |
|
17
|
|
|
``(0, 1, 2)``. Anatomical labels may also be used, such as |
|
18
|
|
|
``'Left'``, ``'Right'``, ``'Anterior'``, ``'Posterior'``, |
|
19
|
|
|
``'Inferior'``, ``'Superior'``, ``'Height'`` and ``'Width'``, |
|
20
|
|
|
``'AP'`` (antero-posterior), ``'lr'`` (lateral), ``'w'`` (width) or |
|
21
|
|
|
``'i'`` (inferior). Only the first letter of the string will be |
|
22
|
|
|
used. If the image is 2D, ``'Height'`` and ``'Width'`` may be |
|
23
|
|
|
used. |
|
24
|
|
|
flip_probability: Probability that the image will be flipped. This is |
|
25
|
|
|
computed on a per-axis basis. |
|
26
|
|
|
p: Probability that this transform will be applied. |
|
27
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
28
|
|
|
|
|
29
|
|
|
Example: |
|
30
|
|
|
>>> import torchio as tio |
|
31
|
|
|
>>> fpg = tio.datasets.FPG() |
|
32
|
|
|
>>> flip = tio.RandomFlip(axes=('LR',)) # flip along lateral axis only |
|
33
|
|
|
|
|
34
|
|
|
.. tip:: It is handy to specify the axes as anatomical labels when the image |
|
35
|
|
|
orientation is not known. |
|
36
|
|
|
""" |
|
37
|
|
|
|
|
38
|
|
|
def __init__( |
|
39
|
|
|
self, |
|
40
|
|
|
axes: Union[int, Tuple[int, ...]] = 0, |
|
41
|
|
|
flip_probability: float = 0.5, |
|
42
|
|
|
p: float = 1, |
|
43
|
|
|
keys: Optional[List[str]] = None, |
|
44
|
|
|
): |
|
45
|
|
|
super().__init__(p=p, keys=keys) |
|
46
|
|
|
self.axes = _parse_axes(axes) |
|
47
|
|
|
self.flip_probability = self.parse_probability(flip_probability) |
|
48
|
|
|
|
|
49
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
50
|
|
|
potential_axes = _ensure_axes_indices(subject, self.axes) |
|
51
|
|
|
axes_to_flip_hot = self.get_params(self.flip_probability) |
|
52
|
|
|
for i in range(3): |
|
53
|
|
|
if i not in potential_axes: |
|
54
|
|
|
axes_to_flip_hot[i] = False |
|
55
|
|
|
axes, = np.where(axes_to_flip_hot) |
|
56
|
|
|
transform = Flip(axes=axes.tolist()) |
|
57
|
|
|
transformed = transform(subject) |
|
58
|
|
|
return transformed |
|
59
|
|
|
|
|
60
|
|
|
@staticmethod |
|
61
|
|
|
def get_params(probability: float) -> List[bool]: |
|
62
|
|
|
return (probability > torch.rand(3)).tolist() |
|
63
|
|
|
|
|
64
|
|
|
|
|
65
|
|
|
class Flip(SpatialTransform): |
|
66
|
|
|
"""Reverse the order of elements in an image along the given axes. |
|
67
|
|
|
|
|
68
|
|
|
Args: |
|
69
|
|
|
axes: Index or tuple of indices of the spatial dimensions along which |
|
70
|
|
|
the image will be flipped. See |
|
71
|
|
|
:py:class:`~torchio.transforms.augmentation.spatial.random_flip.RandomFlip` |
|
72
|
|
|
for more information. |
|
73
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
74
|
|
|
|
|
75
|
|
|
.. tip:: It is handy to specify the axes as anatomical labels when the image |
|
76
|
|
|
orientation is not known. |
|
77
|
|
|
""" |
|
78
|
|
|
|
|
79
|
|
|
def __init__(self, axes, keys: Optional[List[str]] = None): |
|
80
|
|
|
super().__init__(keys=keys) |
|
81
|
|
|
self.axes = _parse_axes(axes) |
|
82
|
|
|
|
|
83
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
84
|
|
|
axes = _ensure_axes_indices(subject, self.axes) |
|
85
|
|
|
for image in self.get_images(subject): |
|
86
|
|
|
_flip_image(image, axes) |
|
87
|
|
|
return subject |
|
88
|
|
|
|
|
89
|
|
|
@staticmethod |
|
90
|
|
|
def is_invertible(): |
|
91
|
|
|
return True |
|
92
|
|
|
|
|
93
|
|
|
def inverse(self): |
|
94
|
|
|
return self |
|
95
|
|
|
|
|
96
|
|
|
|
|
97
|
|
|
def _parse_axes(axes: Union[int, Tuple[int, ...]]): |
|
98
|
|
|
axes_tuple = to_tuple(axes) |
|
99
|
|
|
for axis in axes_tuple: |
|
100
|
|
|
is_int = isinstance(axis, int) |
|
101
|
|
|
is_string = isinstance(axis, str) |
|
102
|
|
|
valid_number = is_int and axis in (0, 1, 2) |
|
103
|
|
|
if not is_string and not valid_number: |
|
104
|
|
|
message = ( |
|
105
|
|
|
f'All axes must be 0, 1 or 2, but found "{axis}"' |
|
106
|
|
|
f' with type {type(axis)}' |
|
107
|
|
|
) |
|
108
|
|
|
raise ValueError(message) |
|
109
|
|
|
return axes_tuple |
|
110
|
|
|
|
|
111
|
|
|
|
|
112
|
|
|
def _ensure_axes_indices(subject, axes): |
|
113
|
|
|
if any(isinstance(n, str) for n in axes): |
|
114
|
|
|
subject.check_consistent_orientation() |
|
115
|
|
|
image = subject.get_first_image() |
|
116
|
|
|
axes = sorted(3 + image.axis_name_to_index(n) for n in axes) |
|
117
|
|
|
return axes |
|
118
|
|
|
|
|
119
|
|
|
|
|
120
|
|
|
def _flip_image(image, axes): |
|
121
|
|
|
spatial_axes = np.array(axes) + 1 |
|
122
|
|
|
data = image.numpy() |
|
123
|
|
|
data = np.flip(data, axis=spatial_axes) |
|
124
|
|
|
data = data.copy() # remove negative strides |
|
125
|
|
|
data = torch.from_numpy(data) |
|
126
|
|
|
image[DATA] = data |
|
127
|
|
|
|
fepegar /
torchio
