|
1
|
|
|
import copy |
|
2
|
|
|
from typing import Union, Sequence, Generator, Tuple |
|
3
|
|
|
|
|
4
|
|
|
import numpy as np |
|
5
|
|
|
import torch |
|
6
|
|
|
from torch.utils.data import IterableDataset |
|
7
|
|
|
|
|
8
|
|
|
from ...torchio import DATA |
|
9
|
|
|
from ...utils import to_tuple |
|
10
|
|
|
from ..subject import Subject |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
class ImageSampler(IterableDataset): |
|
14
|
|
|
r"""Extract random patches from a volume. |
|
15
|
|
|
|
|
16
|
|
|
Args: |
|
17
|
|
|
sample: Sample generated by a |
|
18
|
|
|
:py:class:`~torchio.data.dataset.ImagesDataset`, from which image |
|
19
|
|
|
patches will be extracted. |
|
20
|
|
|
patch_size: Tuple of integers :math:`(d, h, w)` to generate patches |
|
21
|
|
|
of size :math:`d \times h \times w`. |
|
22
|
|
|
If a single number :math:`n` is provided, :math:`d = h = w = n`. |
|
23
|
|
|
""" |
|
24
|
|
|
def __init__(self, sample: Subject, patch_size: Union[int, Sequence[int]]): |
|
25
|
|
|
self.sample = sample |
|
26
|
|
|
patch_size = to_tuple(patch_size, length=3) |
|
27
|
|
|
self.patch_size = np.array(patch_size, dtype=np.uint16) |
|
28
|
|
|
|
|
29
|
|
|
def __iter__(self) -> Generator[Subject, None, None]: |
|
30
|
|
|
while True: |
|
31
|
|
|
yield self.extract_patch() |
|
32
|
|
|
|
|
33
|
|
|
def extract_patch(self) -> Subject: |
|
34
|
|
|
index_ini, index_fin = self.get_random_indices( |
|
35
|
|
|
self.sample, self.patch_size) |
|
36
|
|
|
cropped_sample = self.copy_and_crop( |
|
37
|
|
|
self.sample, |
|
38
|
|
|
index_ini, |
|
39
|
|
|
index_fin, |
|
40
|
|
|
) |
|
41
|
|
|
return cropped_sample |
|
42
|
|
|
|
|
43
|
|
|
@staticmethod |
|
44
|
|
|
def get_random_indices(sample: Subject, patch_size: Tuple[int, int, int]): |
|
45
|
|
|
# Assume all images in sample have the same shape |
|
46
|
|
|
sample.check_consistent_shape() |
|
47
|
|
|
first_image_name = list(sample.keys())[0] |
|
48
|
|
|
first_image_array = sample[first_image_name][DATA] |
|
49
|
|
|
# first_image_array should have shape (1, H, W, D) |
|
50
|
|
|
shape = np.array(first_image_array.shape[1:], dtype=np.uint16) |
|
51
|
|
|
return get_random_indices_from_shape(shape, patch_size) |
|
52
|
|
|
|
|
53
|
|
View Code Duplication |
@staticmethod |
|
|
|
|
|
|
54
|
|
|
def copy_and_crop( |
|
55
|
|
|
sample: Subject, |
|
56
|
|
|
index_ini: np.ndarray, |
|
57
|
|
|
index_fin: np.ndarray, |
|
58
|
|
|
) -> dict: |
|
59
|
|
|
cropped_sample = copy.deepcopy(sample) |
|
60
|
|
|
iterable = sample.get_images_dict(intensity_only=False).items() |
|
61
|
|
|
for image_name, image in iterable: |
|
62
|
|
|
cropped_sample[image_name] = copy.deepcopy(image) |
|
63
|
|
|
sample_image_dict = image |
|
64
|
|
|
cropped_image_dict = cropped_sample[image_name] |
|
65
|
|
|
cropped_image_dict[DATA] = crop( |
|
66
|
|
|
sample_image_dict[DATA], index_ini, index_fin) |
|
|
|
|
|
|
67
|
|
|
# torch doesn't like uint16 |
|
68
|
|
|
cropped_sample['index_ini'] = index_ini.astype(int) |
|
69
|
|
|
return cropped_sample |
|
70
|
|
|
|
|
71
|
|
|
|
|
72
|
|
|
def crop( |
|
73
|
|
|
image: Union[np.ndarray, torch.Tensor], |
|
74
|
|
|
index_ini: np.ndarray, |
|
75
|
|
|
index_fin: np.ndarray, |
|
76
|
|
|
) -> Union[np.ndarray, torch.Tensor]: |
|
77
|
|
|
i_ini, j_ini, k_ini = index_ini |
|
78
|
|
|
i_fin, j_fin, k_fin = index_fin |
|
79
|
|
|
return image[..., i_ini:i_fin, j_ini:j_fin, k_ini:k_fin] |
|
80
|
|
|
|
|
81
|
|
|
|
|
82
|
|
|
def get_random_indices_from_shape( |
|
83
|
|
|
shape: Tuple[int, int, int], |
|
84
|
|
|
patch_size: Tuple[int, int, int], |
|
85
|
|
|
) -> Tuple[np.ndarray, np.ndarray]: |
|
86
|
|
|
shape_array = np.array(shape) |
|
87
|
|
|
patch_size_array = np.array(patch_size) |
|
88
|
|
|
max_index_ini = shape_array - patch_size_array |
|
89
|
|
|
if (max_index_ini < 0).any(): |
|
90
|
|
|
message = ( |
|
91
|
|
|
f'Patch size {patch_size} must not be' |
|
92
|
|
|
f' larger than image size {shape}' |
|
93
|
|
|
) |
|
94
|
|
|
raise ValueError(message) |
|
95
|
|
|
max_index_ini = max_index_ini.astype(np.uint16) |
|
96
|
|
|
coordinates = [] |
|
97
|
|
|
for max_coordinate in max_index_ini.tolist(): |
|
98
|
|
|
if max_coordinate == 0: |
|
99
|
|
|
coordinate = 0 |
|
100
|
|
|
else: |
|
101
|
|
|
coordinate = torch.randint(max_coordinate, size=(1,)).item() |
|
102
|
|
|
coordinates.append(coordinate) |
|
103
|
|
|
index_ini = np.array(coordinates, np.uint16) |
|
104
|
|
|
index_fin = index_ini + patch_size_array |
|
105
|
|
|
return index_ini, index_fin |
|
106
|
|
|
|
fepegar /
torchio
Loading history...