|
1
|
|
|
import copy |
|
2
|
|
|
from typing import Sequence, Optional, Callable, Iterable |
|
3
|
|
|
|
|
4
|
|
|
from torch.utils.data import Dataset |
|
5
|
|
|
|
|
6
|
|
|
from .subject import Subject |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
class SubjectsDataset(Dataset): |
|
10
|
|
|
"""Base TorchIO dataset. |
|
11
|
|
|
|
|
12
|
|
|
Reader of 3D medical images that directly inherits from the PyTorch |
|
13
|
|
|
:class:`~torch.utils.data.Dataset`. It can be used with a PyTorch |
|
14
|
|
|
:class:`~torch.utils.data.DataLoader` for efficient loading and |
|
15
|
|
|
augmentation. It receives a list of instances of :class:`~torchio.Subject` |
|
16
|
|
|
and an optional transform applied to the volumes after loading. |
|
17
|
|
|
|
|
18
|
|
|
Args: |
|
19
|
|
|
subjects: List of instances of :class:`~torchio.Subject`. |
|
20
|
|
|
transform: An instance of :class:`~torchio.transforms.Transform` |
|
21
|
|
|
that will be applied to each subject. |
|
22
|
|
|
load_getitem: Load all subject images before returning it in |
|
23
|
|
|
:meth:`__getitem__`. Set it to ``False`` if some of the images will |
|
24
|
|
|
not be needed during training. |
|
25
|
|
|
|
|
26
|
|
|
Example: |
|
27
|
|
|
>>> import torchio as tio |
|
28
|
|
|
>>> subject_a = tio.Subject( |
|
29
|
|
|
... t1=tio.ScalarImage('t1.nrrd',), |
|
30
|
|
|
... t2=tio.ScalarImage('t2.mha',), |
|
31
|
|
|
... label=tio.LabelMap('t1_seg.nii.gz'), |
|
32
|
|
|
... age=31, |
|
33
|
|
|
... name='Fernando Perez', |
|
34
|
|
|
... ) |
|
35
|
|
|
>>> subject_b = tio.Subject( |
|
36
|
|
|
... t1=tio.ScalarImage('colin27_t1_tal_lin.minc',), |
|
37
|
|
|
... t2=tio.ScalarImage('colin27_t2_tal_lin_dicom',), |
|
38
|
|
|
... label=tio.LabelMap('colin27_seg1.nii.gz'), |
|
39
|
|
|
... age=56, |
|
40
|
|
|
... name='Colin Holmes', |
|
41
|
|
|
... ) |
|
42
|
|
|
>>> subjects_list = [subject_a, subject_b] |
|
43
|
|
|
>>> transforms = [ |
|
44
|
|
|
... tio.RescaleIntensity((0, 1)), |
|
45
|
|
|
... tio.RandomAffine(), |
|
46
|
|
|
... ] |
|
47
|
|
|
>>> transform = tio.Compose(transforms) |
|
48
|
|
|
>>> subjects_dataset = tio.SubjectsDataset(subjects_list, transform=transform) |
|
49
|
|
|
>>> subject = subjects_dataset[0] |
|
50
|
|
|
|
|
51
|
|
|
.. _NiBabel: https://nipy.org/nibabel/#nibabel |
|
52
|
|
|
.. _SimpleITK: https://itk.org/Wiki/ITK/FAQ#What_3D_file_formats_can_ITK_import_and_export.3F |
|
53
|
|
|
.. _DICOM: https://www.dicomstandard.org/ |
|
54
|
|
|
.. _affine matrix: https://nipy.org/nibabel/coordinate_systems.html |
|
55
|
|
|
|
|
56
|
|
|
.. tip:: To quickly iterate over the subjects without loading the images, |
|
57
|
|
|
use :meth:`dry_iter()`. |
|
58
|
|
|
""" # noqa: E501 |
|
59
|
|
|
|
|
60
|
|
|
def __init__( |
|
61
|
|
|
self, |
|
62
|
|
|
subjects: Sequence[Subject], |
|
63
|
|
|
transform: Optional[Callable] = None, |
|
64
|
|
|
load_getitem: bool = True, |
|
65
|
|
|
): |
|
66
|
|
|
self._parse_subjects_list(subjects) |
|
67
|
|
|
self._subjects = subjects |
|
68
|
|
|
self._transform: Optional[Callable] |
|
69
|
|
|
self.set_transform(transform) |
|
70
|
|
|
self.load_getitem = load_getitem |
|
71
|
|
|
|
|
72
|
|
|
def __len__(self): |
|
73
|
|
|
return len(self._subjects) |
|
74
|
|
|
|
|
75
|
|
|
def __getitem__(self, index: int) -> Subject: |
|
76
|
|
|
if not isinstance(index, int): |
|
77
|
|
|
raise ValueError(f'Index "{index}" must be int, not {type(index)}') |
|
78
|
|
|
subject = self._subjects[index] |
|
79
|
|
|
subject = copy.deepcopy(subject) # cheap since images not loaded yet |
|
80
|
|
|
if self.load_getitem: |
|
81
|
|
|
subject.load() |
|
82
|
|
|
|
|
83
|
|
|
# Apply transform (this is usually the bottleneck) |
|
84
|
|
|
if self._transform is not None: |
|
85
|
|
|
subject = self._transform(subject) |
|
86
|
|
|
return subject |
|
87
|
|
|
|
|
88
|
|
|
def dry_iter(self): |
|
89
|
|
|
"""Return the internal list of subjects. |
|
90
|
|
|
|
|
91
|
|
|
This can be used to iterate over the subjects without loading the data |
|
92
|
|
|
and applying any transforms:: |
|
93
|
|
|
|
|
94
|
|
|
>>> names = [subject.name for subject in dataset.dry_iter()] |
|
95
|
|
|
""" |
|
96
|
|
|
return self._subjects |
|
97
|
|
|
|
|
98
|
|
|
def set_transform(self, transform: Optional[Callable]) -> None: |
|
99
|
|
|
"""Set the :attr:`transform` attribute. |
|
100
|
|
|
|
|
101
|
|
|
Args: |
|
102
|
|
|
transform: Callable object, typically an subclass of |
|
103
|
|
|
:class:`torchio.transforms.Transform`. |
|
104
|
|
|
""" |
|
105
|
|
|
if transform is not None and not callable(transform): |
|
106
|
|
|
message = ( |
|
107
|
|
|
'The transform must be a callable object,' |
|
108
|
|
|
f' but it has type {type(transform)}' |
|
109
|
|
|
) |
|
110
|
|
|
raise ValueError(message) |
|
111
|
|
|
self._transform = transform |
|
112
|
|
|
|
|
113
|
|
|
@staticmethod |
|
114
|
|
|
def _parse_subjects_list(subjects_list: Iterable[Subject]) -> None: |
|
115
|
|
|
# Check that it's an iterable |
|
116
|
|
|
try: |
|
117
|
|
|
iter(subjects_list) |
|
118
|
|
|
except TypeError as e: |
|
119
|
|
|
message = ( |
|
120
|
|
|
f'Subject list must be an iterable, not {type(subjects_list)}' |
|
121
|
|
|
) |
|
122
|
|
|
raise TypeError(message) from e |
|
123
|
|
|
|
|
124
|
|
|
# Check that it's not empty |
|
125
|
|
|
if not subjects_list: |
|
126
|
|
|
raise ValueError('Subjects list is empty') |
|
127
|
|
|
|
|
128
|
|
|
# Check each element |
|
129
|
|
|
for subject in subjects_list: |
|
130
|
|
|
if not isinstance(subject, Subject): |
|
131
|
|
|
message = ( |
|
132
|
|
|
'Subjects list must contain instances of torchio.Subject,' |
|
133
|
|
|
f' not "{type(subject)}"' |
|
134
|
|
|
) |
|
135
|
|
|
raise TypeError(message) |
|
136
|
|
|
|
fepegar /
torchio
