|
1
|
|
|
from collections import defaultdict |
|
2
|
|
|
from typing import Tuple, Union, Dict |
|
3
|
|
|
|
|
4
|
|
|
import torch |
|
5
|
|
|
import numpy as np |
|
6
|
|
|
|
|
7
|
|
|
from ....data.subject import Subject |
|
8
|
|
|
from ... import IntensityTransform, FourierTransform |
|
9
|
|
|
from .. import RandomTransform |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class RandomSpike(RandomTransform, IntensityTransform, FourierTransform): |
|
13
|
|
|
r"""Add random MRI spike artifacts. |
|
14
|
|
|
|
|
15
|
|
|
Also known as `Herringbone artifact |
|
16
|
|
|
<https://radiopaedia.org/articles/herringbone-artifact?lang=gb>`_, |
|
17
|
|
|
crisscross artifact or corduroy artifact, it creates stripes in different |
|
18
|
|
|
directions in image space due to spikes in k-space. |
|
19
|
|
|
|
|
20
|
|
|
Args: |
|
21
|
|
|
num_spikes: Number of spikes :math:`n` present in k-space. |
|
22
|
|
|
If a tuple :math:`(a, b)` is provided, then |
|
23
|
|
|
:math:`n \sim \mathcal{U}(a, b) \cap \mathbb{N}`. |
|
24
|
|
|
If only one value :math:`d` is provided, |
|
25
|
|
|
:math:`n \sim \mathcal{U}(0, d) \cap \mathbb{N}`. |
|
26
|
|
|
Larger values generate more distorted images. |
|
27
|
|
|
intensity: Ratio :math:`r` between the spike intensity and the maximum |
|
28
|
|
|
of the spectrum. |
|
29
|
|
|
If a tuple :math:`(a, b)` is provided, then |
|
30
|
|
|
:math:`r \sim \mathcal{U}(a, b)`. |
|
31
|
|
|
If only one value :math:`d` is provided, |
|
32
|
|
|
:math:`r \sim \mathcal{U}(-d, d)`. |
|
33
|
|
|
Larger values generate more distorted images. |
|
34
|
|
|
**kwargs: See :class:`~torchio.transforms.Transform` for additional keyword arguments. |
|
35
|
|
|
|
|
36
|
|
|
.. note:: The execution time of this transform does not depend on the |
|
37
|
|
|
number of spikes. |
|
38
|
|
|
""" |
|
39
|
|
|
def __init__( |
|
40
|
|
|
self, |
|
41
|
|
|
num_spikes: Union[int, Tuple[int, int]] = 1, |
|
42
|
|
|
intensity: Union[float, Tuple[float, float]] = (1, 3), |
|
43
|
|
|
**kwargs |
|
44
|
|
|
): |
|
45
|
|
|
super().__init__(**kwargs) |
|
46
|
|
|
self.intensity_range = self._parse_range( |
|
47
|
|
|
intensity, 'intensity_range') |
|
48
|
|
|
self.num_spikes_range = self._parse_range( |
|
49
|
|
|
num_spikes, 'num_spikes', min_constraint=0, type_constraint=int) |
|
50
|
|
|
|
|
51
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
52
|
|
|
arguments = defaultdict(dict) |
|
53
|
|
|
for image_name in self.get_images_dict(subject): |
|
54
|
|
|
spikes_positions_param, intensity_param = self.get_params( |
|
55
|
|
|
self.num_spikes_range, |
|
56
|
|
|
self.intensity_range, |
|
57
|
|
|
) |
|
58
|
|
|
arguments['spikes_positions'][image_name] = spikes_positions_param |
|
59
|
|
|
arguments['intensity'][image_name] = intensity_param |
|
60
|
|
|
transform = Spike(**self.add_include_exclude(arguments)) |
|
61
|
|
|
transformed = transform(subject) |
|
62
|
|
|
return transformed |
|
63
|
|
|
|
|
64
|
|
|
def get_params( |
|
65
|
|
|
self, |
|
66
|
|
|
num_spikes_range: Tuple[int, int], |
|
67
|
|
|
intensity_range: Tuple[float, float], |
|
68
|
|
|
) -> Tuple[np.ndarray, float]: |
|
69
|
|
|
ns_min, ns_max = num_spikes_range |
|
70
|
|
|
num_spikes_param = torch.randint(ns_min, ns_max + 1, (1,)).item() |
|
71
|
|
|
intensity_param = self.sample_uniform(*intensity_range) |
|
72
|
|
|
spikes_positions = torch.rand(num_spikes_param, 3).numpy() |
|
73
|
|
|
return spikes_positions, intensity_param.item() |
|
74
|
|
|
|
|
75
|
|
|
|
|
76
|
|
|
class Spike(IntensityTransform, FourierTransform): |
|
77
|
|
|
r"""Add MRI spike artifacts. |
|
78
|
|
|
|
|
79
|
|
|
Also known as `Herringbone artifact |
|
80
|
|
|
<https://radiopaedia.org/articles/herringbone-artifact?lang=gb>`_, |
|
81
|
|
|
crisscross artifact or corduroy artifact, it creates stripes in different |
|
82
|
|
|
directions in image space due to spikes in k-space. |
|
83
|
|
|
|
|
84
|
|
|
Args: |
|
85
|
|
|
spikes_positions: |
|
86
|
|
|
intensity: Ratio :math:`r` between the spike intensity and the maximum |
|
87
|
|
|
of the spectrum. |
|
88
|
|
|
**kwargs: See :class:`~torchio.transforms.Transform` for additional keyword arguments. |
|
89
|
|
|
|
|
90
|
|
|
.. note:: The execution time of this transform does not depend on the |
|
91
|
|
|
number of spikes. |
|
92
|
|
|
""" |
|
93
|
|
|
def __init__( |
|
94
|
|
|
self, |
|
95
|
|
|
spikes_positions: Union[np.ndarray, Dict[str, np.ndarray]], |
|
96
|
|
|
intensity: Union[float, Dict[str, float]], |
|
97
|
|
|
**kwargs |
|
98
|
|
|
): |
|
99
|
|
|
super().__init__(**kwargs) |
|
100
|
|
|
self.spikes_positions = spikes_positions |
|
101
|
|
|
self.intensity = intensity |
|
102
|
|
|
self.args_names = 'spikes_positions', 'intensity' |
|
103
|
|
|
self.invert_transform = False |
|
104
|
|
|
|
|
105
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
106
|
|
|
spikes_positions = self.spikes_positions |
|
107
|
|
|
intensity = self.intensity |
|
108
|
|
|
for image_name, image in self.get_images_dict(subject).items(): |
|
109
|
|
|
if self.arguments_are_dict(): |
|
110
|
|
|
spikes_positions = self.spikes_positions[image_name] |
|
111
|
|
|
intensity = self.intensity[image_name] |
|
112
|
|
|
transformed_tensors = [] |
|
113
|
|
|
for channel in image.data: |
|
114
|
|
|
transformed_tensor = self.add_artifact( |
|
115
|
|
|
channel, |
|
116
|
|
|
spikes_positions, |
|
117
|
|
|
intensity, |
|
118
|
|
|
) |
|
119
|
|
|
transformed_tensors.append(transformed_tensor) |
|
120
|
|
|
image.data = torch.stack(transformed_tensors) |
|
121
|
|
|
return subject |
|
122
|
|
|
|
|
123
|
|
|
def add_artifact( |
|
124
|
|
|
self, |
|
125
|
|
|
tensor: torch.Tensor, |
|
126
|
|
|
spikes_positions: np.ndarray, |
|
127
|
|
|
intensity_factor: float, |
|
128
|
|
|
): |
|
129
|
|
|
if intensity_factor == 0 or len(spikes_positions) == 0: |
|
130
|
|
|
return tensor |
|
131
|
|
|
spectrum = self.fourier_transform(tensor) |
|
132
|
|
|
shape = np.array(spectrum.shape) |
|
133
|
|
|
mid_shape = shape // 2 |
|
134
|
|
|
indices = np.floor(spikes_positions * shape).astype(int) |
|
135
|
|
|
for index in indices: |
|
136
|
|
|
diff = index - mid_shape |
|
137
|
|
|
i, j, k = mid_shape + diff |
|
138
|
|
|
# As of torch 1.7, "max is not yet implemented for complex tensors" |
|
139
|
|
|
artifact = spectrum.cpu().numpy().max() * intensity_factor |
|
140
|
|
|
if self.invert_transform: |
|
141
|
|
|
spectrum[i, j, k] -= artifact |
|
142
|
|
|
else: |
|
143
|
|
|
spectrum[i, j, k] += artifact |
|
144
|
|
|
# If we wanted to add a pure cosine, we should add spikes to both |
|
145
|
|
|
# sides of k-space. However, having only one is a better |
|
146
|
|
|
# representation og the actual cause of the artifact in real |
|
147
|
|
|
# scans. Therefore the next two lines have been removed. |
|
148
|
|
|
# #i, j, k = mid_shape - diff |
|
149
|
|
|
# #spectrum[i, j, k] = spectrum.max() * intensity_factor |
|
150
|
|
|
result = self.inv_fourier_transform(spectrum).real.float() |
|
151
|
|
|
return result |
|
152
|
|
|
|
fepegar /
torchio
