|
1
|
|
|
from typing import Union, Sequence, Tuple |
|
2
|
|
|
|
|
3
|
|
|
import torch |
|
4
|
|
|
import numpy as np |
|
5
|
|
|
|
|
6
|
|
|
from ...utils import to_tuple |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
class PatchSampler: |
|
10
|
|
|
r"""Randomly extract patches from a volume. |
|
11
|
|
|
|
|
12
|
|
|
Args: |
|
13
|
|
|
sample: Sample generated by a |
|
14
|
|
|
:py:class:`~torchio.data.dataset.ImagesDataset`, from which image |
|
15
|
|
|
patches will be extracted. |
|
16
|
|
|
patch_size: Tuple of integers :math:`(d, h, w)` to generate patches |
|
17
|
|
|
of size :math:`d \times h \times w`. |
|
18
|
|
|
If a single number :math:`n` is provided, :math:`d = h = w = n`. |
|
19
|
|
|
""" |
|
20
|
|
|
def __init__(self, patch_size: Union[int, Sequence[int]]): |
|
21
|
|
|
patch_size = np.array(to_tuple(patch_size, length=3)) |
|
22
|
|
|
if np.any(patch_size < 1): |
|
23
|
|
|
message = ( |
|
24
|
|
|
'Patch dimensions must be positive integers,' |
|
25
|
|
|
f' not {patch_size}' |
|
26
|
|
|
) |
|
27
|
|
|
raise ValueError(message) |
|
28
|
|
|
self.patch_size = patch_size.astype(np.uint16) |
|
29
|
|
|
|
|
30
|
|
|
def __call__(self): |
|
31
|
|
|
raise NotImplementedError |
|
32
|
|
|
|
|
33
|
|
|
def get_probability_map(self): |
|
34
|
|
|
raise NotImplementedError |
|
35
|
|
|
|
|
36
|
|
|
def extract_patch(self): |
|
37
|
|
|
raise NotImplementedError |
|
38
|
|
|
|
|
39
|
|
|
|
|
40
|
|
|
|
|
41
|
|
|
def crop( |
|
42
|
|
|
image: Union[np.ndarray, torch.Tensor], |
|
43
|
|
|
index_ini: np.ndarray, |
|
44
|
|
|
index_fin: np.ndarray, |
|
45
|
|
|
) -> Union[np.ndarray, torch.Tensor]: |
|
46
|
|
|
i_ini, j_ini, k_ini = index_ini |
|
47
|
|
|
i_fin, j_fin, k_fin = index_fin |
|
48
|
|
|
return image[..., i_ini:i_fin, j_ini:j_fin, k_ini:k_fin] |
|
49
|
|
|
|
|
50
|
|
|
|
|
51
|
|
|
def get_random_indices_from_shape( |
|
52
|
|
|
shape: Tuple[int, int, int], |
|
53
|
|
|
patch_size: Tuple[int, int, int], |
|
54
|
|
|
) -> Tuple[np.ndarray, np.ndarray]: |
|
55
|
|
|
shape_array = np.array(shape) |
|
56
|
|
|
patch_size_array = np.array(patch_size) |
|
57
|
|
|
max_index_ini = shape_array - patch_size_array |
|
58
|
|
|
if (max_index_ini < 0).any(): |
|
59
|
|
|
message = ( |
|
60
|
|
|
f'Patch size {patch_size} must not be' |
|
61
|
|
|
f' larger than image size {shape}' |
|
62
|
|
|
) |
|
63
|
|
|
raise ValueError(message) |
|
64
|
|
|
max_index_ini = max_index_ini.astype(np.uint16) |
|
65
|
|
|
coordinates = [] |
|
66
|
|
|
for max_coordinate in max_index_ini.tolist(): |
|
67
|
|
|
if max_coordinate == 0: |
|
68
|
|
|
coordinate = 0 |
|
69
|
|
|
else: |
|
70
|
|
|
coordinate = torch.randint(max_coordinate, size=(1,)).item() |
|
71
|
|
|
coordinates.append(coordinate) |
|
72
|
|
|
index_ini = np.array(coordinates, np.uint16) |
|
73
|
|
|
index_fin = index_ini + patch_size_array |
|
74
|
|
|
return index_ini, index_fin |
|
75
|
|
|
|
fepegar /
torchio
