|
1
|
|
|
import numbers |
|
2
|
|
|
import warnings |
|
3
|
|
|
from typing import Union |
|
4
|
|
|
from copy import deepcopy |
|
5
|
|
|
from abc import ABC, abstractmethod |
|
6
|
|
|
|
|
7
|
|
|
import torch |
|
|
|
|
|
|
8
|
|
|
import SimpleITK as sitk |
|
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
from .. import TypeData, INTENSITY, DATA |
|
11
|
|
|
from ..data.image import Image |
|
12
|
|
|
from ..data.subject import Subject |
|
13
|
|
|
from ..data.dataset import ImagesDataset |
|
14
|
|
|
from ..utils import nib_to_sitk, sitk_to_nib |
|
15
|
|
|
from .interpolation import Interpolation |
|
16
|
|
|
|
|
17
|
|
|
|
|
18
|
|
|
class Transform(ABC): |
|
19
|
|
|
"""Abstract class for all TorchIO transforms. |
|
20
|
|
|
|
|
21
|
|
|
All classes used to transform a sample from an |
|
22
|
|
|
:py:class:`~torchio.ImagesDataset` should subclass it. |
|
23
|
|
|
All subclasses should overwrite |
|
24
|
|
|
:py:meth:`torchio.tranforms.Transform.apply_transform`, |
|
25
|
|
|
which takes a sample, applies some transformation and returns the result. |
|
26
|
|
|
|
|
27
|
|
|
Args: |
|
28
|
|
|
p: Probability that this transform will be applied. |
|
29
|
|
|
""" |
|
30
|
|
|
def __init__(self, p: float = 1): |
|
31
|
|
|
self.probability = self.parse_probability(p) |
|
32
|
|
|
|
|
33
|
|
|
def __call__(self, data: Union[Subject, torch.Tensor]): |
|
34
|
|
|
"""Transform a sample and return the result. |
|
35
|
|
|
|
|
36
|
|
|
Args: |
|
37
|
|
|
data: Instance of :py:class:`~torchio.Subject` or 4D |
|
38
|
|
|
:py:class:`torch.Tensor` with dimensions :math:`(C, D, H, W)`, |
|
39
|
|
|
where :math:`C` is the number of channels and :math:`D, H, W` |
|
40
|
|
|
are the spatial dimensions. If the input is a tensor, the affine |
|
41
|
|
|
matrix is an identity and a tensor will be also returned. |
|
42
|
|
|
""" |
|
43
|
|
|
if isinstance(data, torch.Tensor): |
|
44
|
|
|
is_tensor = True |
|
45
|
|
|
sample = self.parse_tensor(data) |
|
46
|
|
|
else: |
|
47
|
|
|
is_tensor = False |
|
48
|
|
|
sample = data |
|
49
|
|
|
self.parse_sample(sample) |
|
50
|
|
|
if torch.rand(1).item() > self.probability: |
|
51
|
|
|
return sample |
|
52
|
|
|
sample = deepcopy(sample) |
|
53
|
|
|
transformed = self.apply_transform(sample) |
|
54
|
|
|
if is_tensor: |
|
55
|
|
|
num_channels = len(data) |
|
56
|
|
|
images = [ |
|
57
|
|
|
transformed[f'channel_{i}'][DATA] |
|
58
|
|
|
for i in range(num_channels) |
|
59
|
|
|
] |
|
60
|
|
|
transformed = torch.cat(images) |
|
61
|
|
|
return transformed |
|
62
|
|
|
|
|
63
|
|
|
@abstractmethod |
|
64
|
|
|
def apply_transform(self, sample: Subject): |
|
65
|
|
|
raise NotImplementedError |
|
66
|
|
|
|
|
67
|
|
|
@staticmethod |
|
68
|
|
|
def parse_probability(probability: float) -> float: |
|
69
|
|
|
is_number = isinstance(probability, numbers.Number) |
|
70
|
|
|
if not (is_number and 0 <= probability <= 1): |
|
71
|
|
|
message = ( |
|
72
|
|
|
'Probability must be a number in [0, 1],' |
|
73
|
|
|
f' not {probability}' |
|
74
|
|
|
) |
|
75
|
|
|
raise ValueError(message) |
|
76
|
|
|
return probability |
|
77
|
|
|
|
|
78
|
|
|
@staticmethod |
|
79
|
|
|
def parse_sample(sample: Subject) -> None: |
|
80
|
|
|
if not isinstance(sample, Subject) or not sample.is_sample: |
|
81
|
|
|
message = ( |
|
82
|
|
|
'Inputs to transforms must be instances of torchio.Subject' |
|
83
|
|
|
f' generated by a torchio.ImagesDataset, not "{type(sample)}"' |
|
84
|
|
|
) |
|
85
|
|
|
raise RuntimeError(message) |
|
86
|
|
|
|
|
87
|
|
|
def parse_tensor(self, tensor: torch.Tensor) -> Subject: |
|
88
|
|
|
num_dimensions = tensor.dim() |
|
89
|
|
|
if num_dimensions != 4: |
|
90
|
|
|
message = ( |
|
91
|
|
|
'The input tensor must have 4 dimensions (channels, i, j, k),' |
|
92
|
|
|
f' but has {num_dimensions}: {tensor.shape}' |
|
93
|
|
|
) |
|
94
|
|
|
raise RuntimeError(message) |
|
95
|
|
|
return self._get_subject_from_tensor(tensor) |
|
96
|
|
|
|
|
97
|
|
|
@staticmethod |
|
98
|
|
|
def parse_interpolation(interpolation: str) -> Interpolation: |
|
99
|
|
|
if isinstance(interpolation, Interpolation): |
|
100
|
|
|
message = ( |
|
101
|
|
|
'Interpolation of type torchio.Interpolation' |
|
102
|
|
|
' is deprecated, please use a string instead' |
|
103
|
|
|
) |
|
104
|
|
|
warnings.warn(message, FutureWarning) |
|
105
|
|
|
elif isinstance(interpolation, str): |
|
106
|
|
|
interpolation = interpolation.lower() |
|
107
|
|
|
supported_values = [key.name.lower() for key in Interpolation] |
|
108
|
|
|
if interpolation in supported_values: |
|
109
|
|
|
interpolation = getattr(Interpolation, interpolation.upper()) |
|
110
|
|
|
else: |
|
111
|
|
|
message = ( |
|
112
|
|
|
f'Interpolation "{interpolation}" is not among' |
|
113
|
|
|
f' the supported values: {supported_values}' |
|
114
|
|
|
) |
|
115
|
|
|
raise AttributeError(message) |
|
116
|
|
|
else: |
|
117
|
|
|
message = ( |
|
118
|
|
|
'image_interpolation must be a string,' |
|
119
|
|
|
f' not {type(interpolation)}' |
|
120
|
|
|
) |
|
121
|
|
|
raise TypeError(message) |
|
122
|
|
|
return interpolation |
|
123
|
|
|
|
|
124
|
|
|
@staticmethod |
|
125
|
|
|
def _get_subject_from_tensor(tensor: torch.Tensor) -> Subject: |
|
126
|
|
|
subject_dict = {} |
|
127
|
|
|
for channel_index, channel_tensor in enumerate(tensor): |
|
128
|
|
|
name = f'channel_{channel_index}' |
|
129
|
|
|
image = Image(tensor=channel_tensor, type=INTENSITY) |
|
130
|
|
|
subject_dict[name] = image |
|
131
|
|
|
subject = Subject(subject_dict) |
|
132
|
|
|
dataset = ImagesDataset([subject]) |
|
133
|
|
|
sample = dataset[0] |
|
134
|
|
|
return sample |
|
135
|
|
|
|
|
136
|
|
|
@staticmethod |
|
137
|
|
|
def nib_to_sitk(data: TypeData, affine: TypeData): |
|
138
|
|
|
return nib_to_sitk(data, affine) |
|
139
|
|
|
|
|
140
|
|
|
@staticmethod |
|
141
|
|
|
def sitk_to_nib(image: sitk.Image): |
|
142
|
|
|
return sitk_to_nib(image) |
|
143
|
|
|
|
|
144
|
|
|
@property |
|
145
|
|
|
def name(self): |
|
146
|
|
|
return self.__class__.__name__ |
|
147
|
|
|
|
fepegar /
torchio
Loading history...