|
1
|
|
|
from typing import Optional, Tuple, Union, List |
|
2
|
|
|
import torch |
|
3
|
|
|
import numpy as np |
|
4
|
|
|
from ....data.subject import Subject |
|
5
|
|
|
from ....utils import to_tuple |
|
6
|
|
|
from ....torchio import DATA, TypeTuple, TypeData, TypeTripletInt |
|
7
|
|
|
from ... import IntensityTransform |
|
8
|
|
|
from .. import RandomTransform |
|
9
|
|
|
|
|
10
|
|
|
|
|
11
|
|
|
class RandomSwap(RandomTransform, IntensityTransform): |
|
12
|
|
|
r"""Randomly swap patches within an image. |
|
13
|
|
|
|
|
14
|
|
|
This is typically used in `context restoration for self-supervised learning |
|
15
|
|
|
<https://www.sciencedirect.com/science/article/pii/S1361841518304699>`_. |
|
16
|
|
|
|
|
17
|
|
|
Args: |
|
18
|
|
|
patch_size: Tuple of integers :math:`(w, h, d)` to swap patches |
|
19
|
|
|
of size :math:`h \times w \times d`. |
|
20
|
|
|
If a single number :math:`n` is provided, :math:`w = h = d = n`. |
|
21
|
|
|
num_iterations: Number of times that two patches will be swapped. |
|
22
|
|
|
p: Probability that this transform will be applied. |
|
23
|
|
|
seed: See :py:class:`~torchio.transforms.augmentation.RandomTransform`. |
|
24
|
|
|
keys: See :py:class:`~torchio.transforms.Transform`. |
|
25
|
|
|
""" |
|
26
|
|
|
def __init__( |
|
27
|
|
|
self, |
|
28
|
|
|
patch_size: TypeTuple = 15, |
|
29
|
|
|
num_iterations: int = 100, |
|
30
|
|
|
p: float = 1, |
|
31
|
|
|
seed: Optional[int] = None, |
|
32
|
|
|
keys: Optional[List[str]] = None, |
|
33
|
|
|
): |
|
34
|
|
|
super().__init__(p=p, seed=seed, keys=keys) |
|
35
|
|
|
self.patch_size = np.array(to_tuple(patch_size)) |
|
36
|
|
|
self.num_iterations = self.parse_num_iterations(num_iterations) |
|
37
|
|
|
|
|
38
|
|
|
@staticmethod |
|
39
|
|
|
def parse_num_iterations(num_iterations): |
|
40
|
|
|
if not isinstance(num_iterations, int): |
|
41
|
|
|
raise TypeError('num_iterations must be an int,' |
|
42
|
|
|
f'not {num_iterations}') |
|
43
|
|
|
if num_iterations < 0: |
|
44
|
|
|
raise ValueError('num_iterations must be positive,' |
|
45
|
|
|
f'not {num_iterations}') |
|
46
|
|
|
return num_iterations |
|
47
|
|
|
|
|
48
|
|
|
@staticmethod |
|
49
|
|
|
def get_params( |
|
50
|
|
|
tensor: torch.Tensor, |
|
51
|
|
|
patch_size: np.ndarray, |
|
52
|
|
|
num_iterations: int, |
|
53
|
|
|
) -> List[Tuple[np.ndarray, np.ndarray]]: |
|
54
|
|
|
spatial_shape = tensor.shape[-3:] |
|
55
|
|
|
locations = [] |
|
56
|
|
|
for _ in range(num_iterations): |
|
57
|
|
|
first_ini, first_fin = get_random_indices_from_shape( |
|
58
|
|
|
spatial_shape, |
|
59
|
|
|
patch_size, |
|
60
|
|
|
) |
|
61
|
|
|
while True: |
|
62
|
|
|
second_ini, second_fin = get_random_indices_from_shape( |
|
63
|
|
|
spatial_shape, |
|
64
|
|
|
patch_size, |
|
65
|
|
|
) |
|
66
|
|
|
larger_than_initial = np.all(second_ini >= first_ini) |
|
67
|
|
|
less_than_final = np.all(second_fin <= first_fin) |
|
68
|
|
|
if larger_than_initial and less_than_final: |
|
69
|
|
|
continue # patches overlap |
|
70
|
|
|
else: |
|
71
|
|
|
break # patches don't overlap |
|
72
|
|
|
locations.append((first_ini, second_ini)) |
|
73
|
|
|
return locations |
|
74
|
|
|
|
|
75
|
|
|
def apply_transform(self, sample: Subject) -> dict: |
|
76
|
|
|
for image in self.get_images(sample): |
|
77
|
|
|
tensor = image[DATA] |
|
|
|
|
|
|
78
|
|
|
locations = self.get_params( |
|
79
|
|
|
tensor, self.patch_size, self.num_iterations) |
|
80
|
|
|
image[DATA] = swap(tensor, self.patch_size, locations) |
|
81
|
|
|
return sample |
|
82
|
|
|
|
|
83
|
|
|
|
|
84
|
|
|
def swap( |
|
85
|
|
|
tensor: torch.Tensor, |
|
86
|
|
|
patch_size: TypeTuple, |
|
87
|
|
|
locations: List[Tuple[np.ndarray, np.ndarray]], |
|
88
|
|
|
) -> None: |
|
89
|
|
|
tensor = tensor.clone() |
|
90
|
|
|
patch_size = np.array(patch_size) |
|
91
|
|
|
for first_ini, second_ini in locations: |
|
92
|
|
|
first_fin = first_ini + patch_size |
|
93
|
|
|
second_fin = second_ini + patch_size |
|
94
|
|
|
first_patch = crop(tensor, first_ini, first_fin) |
|
95
|
|
|
second_patch = crop(tensor, second_ini, second_fin).clone() |
|
96
|
|
|
insert(tensor, first_patch, second_ini) |
|
97
|
|
|
insert(tensor, second_patch, first_ini) |
|
98
|
|
|
return tensor |
|
99
|
|
|
|
|
100
|
|
|
|
|
101
|
|
|
def insert(tensor: TypeData, patch: TypeData, index_ini: np.ndarray) -> None: |
|
102
|
|
|
index_fin = index_ini + np.array(patch.shape[-3:]) |
|
103
|
|
|
i_ini, j_ini, k_ini = index_ini |
|
104
|
|
|
i_fin, j_fin, k_fin = index_fin |
|
105
|
|
|
tensor[:, i_ini:i_fin, j_ini:j_fin, k_ini:k_fin] = patch |
|
106
|
|
|
|
|
107
|
|
|
|
|
108
|
|
|
def crop( |
|
109
|
|
|
image: Union[np.ndarray, torch.Tensor], |
|
110
|
|
|
index_ini: np.ndarray, |
|
111
|
|
|
index_fin: np.ndarray, |
|
112
|
|
|
) -> Union[np.ndarray, torch.Tensor]: |
|
113
|
|
|
i_ini, j_ini, k_ini = index_ini |
|
114
|
|
|
i_fin, j_fin, k_fin = index_fin |
|
115
|
|
|
return image[:, i_ini:i_fin, j_ini:j_fin, k_ini:k_fin] |
|
116
|
|
|
|
|
117
|
|
|
|
|
118
|
|
|
def get_random_indices_from_shape( |
|
119
|
|
|
spatial_shape: TypeTripletInt, |
|
120
|
|
|
patch_size: TypeTripletInt, |
|
121
|
|
|
) -> Tuple[np.ndarray, np.ndarray]: |
|
122
|
|
|
shape_array = np.array(spatial_shape) |
|
123
|
|
|
patch_size_array = np.array(patch_size) |
|
124
|
|
|
max_index_ini = shape_array - patch_size_array |
|
125
|
|
|
if (max_index_ini < 0).any(): |
|
126
|
|
|
message = ( |
|
127
|
|
|
f'Patch size {patch_size} cannot be' |
|
128
|
|
|
f' larger than image spatial shape {spatial_shape}' |
|
129
|
|
|
) |
|
130
|
|
|
raise ValueError(message) |
|
131
|
|
|
max_index_ini = max_index_ini.astype(np.uint16) |
|
132
|
|
|
coordinates = [] |
|
133
|
|
|
for max_coordinate in max_index_ini.tolist(): |
|
134
|
|
|
if max_coordinate == 0: |
|
135
|
|
|
coordinate = 0 |
|
136
|
|
|
else: |
|
137
|
|
|
coordinate = torch.randint(max_coordinate, size=(1,)).item() |
|
138
|
|
|
coordinates.append(coordinate) |
|
139
|
|
|
index_ini = np.array(coordinates, np.uint16) |
|
140
|
|
|
index_fin = index_ini + patch_size_array |
|
141
|
|
|
return index_ini, index_fin |
|
142
|
|
|
|
fepegar /
torchio
Loading history...