|
1
|
|
|
import warnings |
|
2
|
|
|
from pathlib import Path |
|
3
|
|
|
from typing import Any, Dict, Tuple, Optional |
|
4
|
|
|
|
|
5
|
|
|
import torch |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
from ..torchio import TypePath, DATA, TYPE, AFFINE, PATH, STEM, INTENSITY |
|
9
|
|
|
from .io import read_image |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class Image(dict): |
|
13
|
|
|
r"""Class to store information about an image. |
|
14
|
|
|
|
|
15
|
|
|
Args: |
|
16
|
|
|
path: Path to a file that can be read by |
|
17
|
|
|
:mod:`SimpleITK` or :mod:`nibabel` or to a directory containing |
|
18
|
|
|
DICOM files. |
|
19
|
|
|
type: Type of image, such as :attr:`torchio.INTENSITY` or |
|
20
|
|
|
:attr:`torchio.LABEL`. This will be used by the transforms to |
|
21
|
|
|
decide whether to apply an operation, or which interpolation to use |
|
22
|
|
|
when resampling. |
|
23
|
|
|
tensor: If :attr:`path` is not given, :attr:`tensor` must be a 4D |
|
24
|
|
|
:py:class:`torch.Tensor` with dimensions :math:`(C, D, H, W)`, |
|
25
|
|
|
where :math:`C` is the number of channels and :math:`D, H, W` |
|
26
|
|
|
are the spatial dimensions. |
|
27
|
|
|
affine: If :attr:`path` is not given, :attr:`affine` must be a |
|
28
|
|
|
:math:`4 \times 4` NumPy array. If ``None``, :attr:`affine` is an |
|
29
|
|
|
identity matrix. |
|
30
|
|
|
**kwargs: Items that will be added to image dictionary within the |
|
31
|
|
|
subject sample. |
|
32
|
|
|
""" |
|
33
|
|
|
|
|
34
|
|
|
def __init__( |
|
35
|
|
|
self, |
|
36
|
|
|
path: Optional[TypePath] = None, |
|
37
|
|
|
type: str = INTENSITY, |
|
38
|
|
|
tensor: Optional[torch.Tensor] = None, |
|
39
|
|
|
affine: Optional[torch.Tensor] = None, |
|
40
|
|
|
**kwargs: Dict[str, Any], |
|
41
|
|
|
): |
|
42
|
|
|
if path is None and tensor is None: |
|
43
|
|
|
raise ValueError('A value for path or tensor must be given') |
|
44
|
|
|
if path is not None: |
|
45
|
|
|
if tensor is not None or affine is not None: |
|
46
|
|
|
message = 'If a path is given, tensor and affine must be None' |
|
47
|
|
|
raise ValueError(message) |
|
48
|
|
|
self.tensor = self.parse_tensor(tensor) |
|
49
|
|
|
self.affine = self.parse_affine(affine) |
|
50
|
|
|
if self.affine is None: |
|
51
|
|
|
self.affine = np.eye(4) |
|
52
|
|
|
for key in (DATA, AFFINE, TYPE, PATH, STEM): |
|
53
|
|
|
if key in kwargs: |
|
54
|
|
|
raise ValueError(f'Key {key} is reserved. Use a different one') |
|
55
|
|
|
|
|
56
|
|
|
super().__init__(**kwargs) |
|
57
|
|
|
self.path = self._parse_path(path) |
|
58
|
|
|
self.type = type |
|
59
|
|
|
self.is_sample = False # set to True by ImagesDataset |
|
60
|
|
|
|
|
61
|
|
|
@staticmethod |
|
62
|
|
|
def _parse_path(path: TypePath) -> Path: |
|
63
|
|
|
if path is None: |
|
64
|
|
|
return |
|
65
|
|
|
try: |
|
66
|
|
|
path = Path(path).expanduser() |
|
67
|
|
|
except TypeError: |
|
68
|
|
|
message = f'Conversion to path not possible for variable: {path}' |
|
69
|
|
|
raise TypeError(message) |
|
70
|
|
|
if not (path.is_file() or path.is_dir()): # might be a dir with DICOM |
|
71
|
|
|
raise FileNotFoundError(f'File not found: {path}') |
|
72
|
|
|
return path |
|
73
|
|
|
|
|
74
|
|
|
@staticmethod |
|
75
|
|
|
def parse_tensor(tensor: torch.Tensor) -> torch.Tensor: |
|
76
|
|
|
if tensor is None: |
|
77
|
|
|
return None |
|
78
|
|
|
num_dimensions = tensor.dim() |
|
79
|
|
|
if num_dimensions != 3: |
|
80
|
|
|
message = ( |
|
81
|
|
|
'The input tensor must have 3 dimensions (D, H, W),' |
|
82
|
|
|
f' but has {num_dimensions}: {tensor.shape}' |
|
83
|
|
|
) |
|
84
|
|
|
raise RuntimeError(message) |
|
85
|
|
|
tensor = tensor.unsqueeze(0) # add channels dimension |
|
86
|
|
|
return tensor |
|
87
|
|
|
|
|
88
|
|
|
@staticmethod |
|
89
|
|
|
def parse_affine(affine: np.ndarray) -> np.ndarray: |
|
90
|
|
|
if affine is None: |
|
91
|
|
|
return np.eye(4) |
|
92
|
|
|
if not isinstance(affine, np.ndarray): |
|
93
|
|
|
raise TypeError(f'Affine must be a NumPy array, not {type(affine)}') |
|
94
|
|
|
if affine.shape != (4, 4): |
|
95
|
|
|
raise ValueError(f'Affine shape must be (4, 4), not {affine.shape}') |
|
96
|
|
|
return affine |
|
97
|
|
|
|
|
98
|
|
|
def load(self, check_nans: bool = True) -> Tuple[torch.Tensor, np.ndarray]: |
|
99
|
|
|
r"""Load the image from disk. |
|
100
|
|
|
|
|
101
|
|
|
The file is expected to be monomodal/grayscale and 2D or 3D. |
|
102
|
|
|
A channels dimension is added to the tensor. |
|
103
|
|
|
|
|
104
|
|
|
Args: |
|
105
|
|
|
check_nans: If ``True``, issues a warning if NaNs are found |
|
106
|
|
|
in the image |
|
107
|
|
|
|
|
108
|
|
|
Returns: |
|
109
|
|
|
Tuple containing a 4D data tensor of size |
|
110
|
|
|
:math:`(1, D_{in}, H_{in}, W_{in})` |
|
111
|
|
|
and a 2D 4x4 affine matrix |
|
112
|
|
|
""" |
|
113
|
|
|
if self.path is None: |
|
114
|
|
|
return self.tensor, self.affine |
|
115
|
|
|
tensor, affine = read_image(self.path) |
|
116
|
|
|
# https://github.com/pytorch/pytorch/issues/9410#issuecomment-404968513 |
|
117
|
|
|
tensor = tensor[(None,) * (3 - tensor.ndim)] # force to be 3D |
|
118
|
|
|
tensor = tensor.unsqueeze(0) # add channels dimension |
|
119
|
|
|
if check_nans and torch.isnan(tensor).any(): |
|
120
|
|
|
warnings.warn(f'NaNs found in file "{self.path}"') |
|
121
|
|
|
return tensor, affine |
|
122
|
|
|
|
fepegar /
torchio
