|
1
|
|
|
import warnings |
|
2
|
|
|
from typing import Union, Tuple, Optional, List, Sequence |
|
3
|
|
|
|
|
4
|
|
|
import torch |
|
5
|
|
|
|
|
6
|
|
|
from ....torchio import TypeRangeFloat |
|
7
|
|
|
from ....data.subject import Subject |
|
8
|
|
|
from ....utils import to_tuple |
|
9
|
|
|
from ... import SpatialTransform |
|
10
|
|
|
from .. import RandomTransform |
|
11
|
|
|
from ...preprocessing import Resample |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class RandomAnisotropy(RandomTransform, SpatialTransform): |
|
15
|
|
|
r"""Downsample an image along an axis and upsample to initial space. |
|
16
|
|
|
|
|
17
|
|
|
This transform simulates an image that has been acquired using anisotropic |
|
18
|
|
|
spacing and resampled back to a more standard spacing. |
|
19
|
|
|
|
|
20
|
|
|
Args: |
|
21
|
|
|
axes: Axis or tuple of axes along which the image will be downsampled. |
|
22
|
|
|
downsampling: Downsampling factor :math:`m \gt 1`. If a tuple |
|
23
|
|
|
:math:`(a, b)` is provided then :math:`m \sim \mathcal{U}(a, b)`. |
|
24
|
|
|
image_interpolation: Image interpolation used to upsample the image back |
|
25
|
|
|
to its initial spacing. Downsampling is performed using nearest |
|
26
|
|
|
neighbor interpolation. See :ref:`Interpolation` for supported |
|
27
|
|
|
interpolation types. |
|
28
|
|
|
p: Probability that this transform will be applied. |
|
29
|
|
|
keys: See :class:`~torchio.transforms.Transform`. |
|
30
|
|
|
|
|
31
|
|
|
Example: |
|
32
|
|
|
>>> import torchio as tio |
|
33
|
|
|
>>> transform = tio.RandomAnisotropy(axes=1, downsampling=2.) # Multiply spacing of second axis by 2 |
|
34
|
|
|
>>> transform = tio.RandomAnisotropy( |
|
35
|
|
|
... axes=(0, 1, 2), |
|
36
|
|
|
... downsampling=(2, 5), |
|
37
|
|
|
... ) # Multiply spacing of one of the 3 axes by a factor randomly chosen in [2, 5] |
|
38
|
|
|
>>> colin = tio.datasets.Colin27() |
|
39
|
|
|
>>> transformed = transform(colin) # images have now anisotropic spacing |
|
40
|
|
|
""" |
|
41
|
|
|
|
|
42
|
|
|
def __init__( |
|
43
|
|
|
self, |
|
44
|
|
|
axes: Union[int, Tuple[int, ...]] = (0, 1, 2), |
|
45
|
|
|
downsampling: TypeRangeFloat = (1.5, 5), |
|
46
|
|
|
image_interpolation: str = 'linear', |
|
47
|
|
|
p: float = 1, |
|
48
|
|
|
keys: Optional[Sequence[str]] = None, |
|
49
|
|
|
): |
|
50
|
|
|
super().__init__(p=p, keys=keys) |
|
51
|
|
|
self.axes = self.parse_axes(axes) |
|
52
|
|
|
self.downsampling_range = self.parse_range( |
|
53
|
|
|
downsampling, 'downsampling', min_constraint=1) |
|
54
|
|
|
self.image_interpolation = self.parse_interpolation(image_interpolation) |
|
55
|
|
|
|
|
56
|
|
|
def get_params( |
|
57
|
|
|
self, |
|
58
|
|
|
axes: Tuple[int, ...], |
|
59
|
|
|
downsampling_range: Tuple[float, float], |
|
60
|
|
|
is_2d: bool, |
|
61
|
|
|
) -> List[bool]: |
|
62
|
|
|
axis = axes[torch.randint(0, len(axes), (1,))] |
|
63
|
|
|
downsampling = self.sample_uniform(*downsampling_range).item() |
|
64
|
|
|
return axis, downsampling |
|
65
|
|
|
|
|
66
|
|
|
@staticmethod |
|
67
|
|
|
def parse_axes(axes: Union[int, Tuple[int, ...]]): |
|
68
|
|
|
axes_tuple = to_tuple(axes) |
|
69
|
|
|
for axis in axes_tuple: |
|
70
|
|
|
is_int = isinstance(axis, int) |
|
71
|
|
|
if not is_int or axis not in (0, 1, 2): |
|
72
|
|
|
raise ValueError('All axes must be 0, 1 or 2') |
|
73
|
|
|
return axes_tuple |
|
74
|
|
|
|
|
75
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
76
|
|
|
is_2d = subject.get_first_image().is_2d() |
|
77
|
|
|
if is_2d and 2 in self.axes: |
|
78
|
|
|
warnings.warn(f'Input image is 2D, but "2" is in axes: {self.axes}') |
|
79
|
|
|
self.axes = list(self.axes) |
|
80
|
|
|
self.axes.remove(2) |
|
81
|
|
|
axis, downsampling = self.get_params( |
|
82
|
|
|
self.axes, |
|
83
|
|
|
self.downsampling_range, |
|
84
|
|
|
is_2d, |
|
85
|
|
|
) |
|
86
|
|
|
target_spacing = list(subject.spacing) |
|
87
|
|
|
target_spacing[axis] *= downsampling |
|
88
|
|
|
downsample = Resample( |
|
89
|
|
|
tuple(target_spacing), |
|
90
|
|
|
image_interpolation='nearest', |
|
91
|
|
|
) |
|
92
|
|
|
downsampled = downsample(subject) |
|
93
|
|
|
upsample = Resample( |
|
94
|
|
|
subject.get_first_image(), |
|
95
|
|
|
image_interpolation=self.image_interpolation, |
|
96
|
|
|
) |
|
97
|
|
|
upsampled = upsample(downsampled) |
|
98
|
|
|
return upsampled |
|
99
|
|
|
|
fepegar /
torchio
