1
|
|
|
from typing import Optional, List, Sequence, Union |
2
|
|
|
|
3
|
|
|
import torch |
4
|
|
|
import numpy as np |
5
|
|
|
import nibabel as nib |
6
|
|
|
import SimpleITK as sitk |
7
|
|
|
|
8
|
|
|
from ..typing import TypeData |
9
|
|
|
from ..data.subject import Subject |
10
|
|
|
from ..data.image import Image, ScalarImage |
11
|
|
|
from ..data.io import nib_to_sitk, sitk_to_nib |
12
|
|
|
|
13
|
|
|
|
14
|
|
|
TypeTransformInput = Union[ |
15
|
|
|
Subject, |
16
|
|
|
Image, |
17
|
|
|
torch.Tensor, |
18
|
|
|
np.ndarray, |
19
|
|
|
sitk.Image, |
20
|
|
|
dict, |
21
|
|
|
nib.Nifti1Image, |
22
|
|
|
] |
23
|
|
|
|
24
|
|
|
|
25
|
|
|
class DataParser: |
26
|
|
|
def __init__( |
27
|
|
|
self, |
28
|
|
|
data: TypeTransformInput, |
29
|
|
|
keys: Optional[Sequence[str]] = None, |
30
|
|
|
): |
31
|
|
|
self.data = data |
32
|
|
|
self.keys = keys |
33
|
|
|
self.default_image_name = 'default_image_name' |
34
|
|
|
self.is_tensor = False |
35
|
|
|
self.is_array = False |
36
|
|
|
self.is_dict = False |
37
|
|
|
self.is_image = False |
38
|
|
|
self.is_sitk = False |
39
|
|
|
self.is_nib = False |
40
|
|
|
|
41
|
|
|
def get_subject(self): |
42
|
|
|
if isinstance(self.data, nib.Nifti1Image): |
43
|
|
|
tensor = self.data.get_fdata(dtype=np.float32) |
44
|
|
|
data = ScalarImage(tensor=tensor, affine=self.data.affine) |
45
|
|
|
subject = self._get_subject_from_image(data) |
46
|
|
|
self.is_nib = True |
47
|
|
|
elif isinstance(self.data, (np.ndarray, torch.Tensor)): |
48
|
|
|
subject = self._parse_tensor(self.data) |
49
|
|
|
self.is_array = isinstance(self.data, np.ndarray) |
50
|
|
|
self.is_tensor = True |
51
|
|
|
elif isinstance(self.data, Image): |
52
|
|
|
subject = self._get_subject_from_image(self.data) |
53
|
|
|
self.is_image = True |
54
|
|
|
elif isinstance(self.data, Subject): |
55
|
|
|
subject = self.data |
56
|
|
|
elif isinstance(self.data, sitk.Image): |
57
|
|
|
subject = self._get_subject_from_sitk_image(self.data) |
58
|
|
|
self.is_sitk = True |
59
|
|
|
elif isinstance(self.data, dict): # e.g. Eisen or MONAI dicts |
60
|
|
|
if self.keys is None: |
61
|
|
|
message = ( |
62
|
|
|
'If the input is a dictionary, a value for "include" must' |
63
|
|
|
' be specified when instantiating the transform. See the' |
64
|
|
|
' docs for Transform:' |
65
|
|
|
' https://torchio.readthedocs.io/transforms/transforms.html#torchio.transforms.Transform' # noqa: E501 |
66
|
|
|
) |
67
|
|
|
raise RuntimeError(message) |
68
|
|
|
subject = self._get_subject_from_dict(self.data, self.keys) |
69
|
|
|
self.is_dict = True |
70
|
|
|
else: |
71
|
|
|
raise ValueError(f'Input type not recognized: {type(self.data)}') |
72
|
|
|
self._parse_subject(subject) |
73
|
|
|
return subject |
74
|
|
|
|
75
|
|
|
def get_output(self, transformed): |
76
|
|
|
if self.is_tensor or self.is_sitk: |
77
|
|
|
image = transformed[self.default_image_name] |
78
|
|
|
transformed = image.data |
79
|
|
|
if self.is_array: |
80
|
|
|
transformed = transformed.numpy() |
81
|
|
|
elif self.is_sitk: |
82
|
|
|
transformed = nib_to_sitk(image.data, image.affine) |
83
|
|
|
elif self.is_image: |
84
|
|
|
transformed = transformed[self.default_image_name] |
85
|
|
|
elif self.is_dict: |
86
|
|
|
transformed = dict(transformed) |
87
|
|
|
for key, value in transformed.items(): |
88
|
|
|
if isinstance(value, Image): |
89
|
|
|
transformed[key] = value.data |
90
|
|
|
elif self.is_nib: |
91
|
|
|
image = transformed[self.default_image_name] |
92
|
|
|
data = image.data |
93
|
|
|
if len(data) > 1: |
94
|
|
|
message = ( |
95
|
|
|
'Multichannel images not supported for input of type' |
96
|
|
|
' nibabel.nifti.Nifti1Image' |
97
|
|
|
) |
98
|
|
|
raise RuntimeError(message) |
99
|
|
|
transformed = nib.Nifti1Image(data[0].numpy(), image.affine) |
100
|
|
|
return transformed |
101
|
|
|
|
102
|
|
|
@staticmethod |
103
|
|
|
def _parse_subject(subject: Subject) -> None: |
104
|
|
|
if not isinstance(subject, Subject): |
105
|
|
|
message = ( |
106
|
|
|
'Input to a transform must be a tensor or an instance' |
107
|
|
|
f' of torchio.Subject, not "{type(subject)}"' |
108
|
|
|
) |
109
|
|
|
raise RuntimeError(message) |
110
|
|
|
|
111
|
|
|
def _parse_tensor(self, data: TypeData) -> Subject: |
112
|
|
|
if data.ndim != 4: |
113
|
|
|
message = ( |
114
|
|
|
'The input must be a 4D tensor with dimensions' |
115
|
|
|
f' (channels, x, y, z) but it has shape {tuple(data.shape)}.' |
116
|
|
|
' Tips: if it is a volume, please add the channels dimension;' |
117
|
|
|
' if it is 2D, also add a dimension of size 1 for the z axis' |
118
|
|
|
) |
119
|
|
|
raise ValueError(message) |
120
|
|
|
return self._get_subject_from_tensor(data) |
121
|
|
|
|
122
|
|
|
def _get_subject_from_tensor(self, tensor: torch.Tensor) -> Subject: |
123
|
|
|
image = ScalarImage(tensor=tensor) |
124
|
|
|
return self._get_subject_from_image(image) |
125
|
|
|
|
126
|
|
|
def _get_subject_from_image(self, image: Image) -> Subject: |
127
|
|
|
subject = Subject({self.default_image_name: image}) |
128
|
|
|
return subject |
129
|
|
|
|
130
|
|
|
@staticmethod |
131
|
|
|
def _get_subject_from_dict( |
132
|
|
|
data: dict, |
133
|
|
|
image_keys: List[str], |
134
|
|
|
) -> Subject: |
135
|
|
|
subject_dict = {} |
136
|
|
|
for key, value in data.items(): |
137
|
|
|
if key in image_keys: |
138
|
|
|
value = ScalarImage(tensor=value) |
139
|
|
|
subject_dict[key] = value |
140
|
|
|
return Subject(subject_dict) |
141
|
|
|
|
142
|
|
|
def _get_subject_from_sitk_image(self, image): |
143
|
|
|
tensor, affine = sitk_to_nib(image) |
144
|
|
|
image = ScalarImage(tensor=tensor, affine=affine) |
145
|
|
|
return self._get_subject_from_image(image) |
146
|
|
|
|