|
1
|
|
|
import copy |
|
2
|
|
|
import warnings |
|
3
|
|
|
|
|
4
|
|
|
import torch |
|
5
|
|
|
import torchio as tio |
|
6
|
|
|
from torchio.transforms.intensity_transform import IntensityTransform |
|
7
|
|
|
from ..utils import TorchioTestCase |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
class TestInvertibility(TorchioTestCase): |
|
11
|
|
|
|
|
12
|
|
|
def test_all_random_transforms(self): |
|
13
|
|
|
transform = self.get_large_composed_transform() |
|
14
|
|
|
# Remove RandomLabelsToImage as it will add a new image to the subject |
|
15
|
|
|
for t in transform.transforms: |
|
16
|
|
|
if t.name == 'RandomLabelsToImage': |
|
17
|
|
|
transform.transforms.remove(t) |
|
18
|
|
|
break |
|
19
|
|
|
# Ignore elastic deformation and gamma warnings during execution |
|
20
|
|
|
# Ignore some transforms not invertible |
|
21
|
|
|
with warnings.catch_warnings(): |
|
22
|
|
|
warnings.simplefilter('ignore', RuntimeWarning) |
|
23
|
|
|
transformed = transform(self.sample_subject) |
|
24
|
|
|
inverting_transform = transformed.get_inverse_transform() |
|
25
|
|
|
transformed_back = inverting_transform(transformed) |
|
26
|
|
|
self.assertEqual( |
|
27
|
|
|
transformed.t1.shape, |
|
28
|
|
|
transformed_back.t1.shape, |
|
29
|
|
|
) |
|
30
|
|
|
self.assertTensorEqual( |
|
31
|
|
|
transformed.label.affine, |
|
32
|
|
|
transformed_back.label.affine, |
|
33
|
|
|
) |
|
34
|
|
|
|
|
35
|
|
|
def test_ignore_intensity(self): |
|
36
|
|
|
composed = self.get_large_composed_transform() |
|
37
|
|
|
with warnings.catch_warnings(): |
|
38
|
|
|
warnings.simplefilter('ignore', RuntimeWarning) |
|
39
|
|
|
transformed = composed(self.sample_subject) |
|
40
|
|
|
inverse_transform = transformed.get_inverse_transform(warn=False) |
|
41
|
|
|
for transform in inverse_transform: |
|
42
|
|
|
assert not isinstance(transform, IntensityTransform) |
|
43
|
|
|
|
|
44
|
|
|
def test_different_interpolation(self): |
|
45
|
|
|
def model_probs(subject): |
|
46
|
|
|
subject = copy.deepcopy(subject) |
|
47
|
|
|
subject.im.set_data(torch.rand_like(subject.im.data)) |
|
48
|
|
|
return subject |
|
49
|
|
|
|
|
50
|
|
|
def model_label(subject): |
|
51
|
|
|
subject = model_probs(subject) |
|
52
|
|
|
subject.im.set_data(torch.bernoulli(subject.im.data)) |
|
53
|
|
|
return subject |
|
54
|
|
|
|
|
55
|
|
|
transform = tio.RandomAffine(image_interpolation='bspline') |
|
56
|
|
|
subject = copy.deepcopy(self.sample_subject) |
|
57
|
|
|
tensor = (torch.rand(1, 20, 20, 20) > 0.5).float() # 0s and 1s |
|
58
|
|
|
subject = tio.Subject(im=tio.ScalarImage(tensor=tensor)) |
|
59
|
|
|
transformed = transform(subject) |
|
60
|
|
|
assert transformed.im.data.min() < 0 |
|
61
|
|
|
assert transformed.im.data.max() > 1 |
|
62
|
|
|
|
|
63
|
|
|
subject_probs = model_probs(transformed) |
|
64
|
|
|
transformed_back = subject_probs.apply_inverse_transform() |
|
65
|
|
|
assert transformed_back.im.data.min() < 0 |
|
66
|
|
|
assert transformed_back.im.data.max() > 1 |
|
67
|
|
|
transformed_back_linear = subject_probs.apply_inverse_transform( |
|
68
|
|
|
image_interpolation='linear', |
|
69
|
|
|
) |
|
70
|
|
|
assert transformed_back_linear.im.data.min() >= 0 |
|
71
|
|
|
assert transformed_back_linear.im.data.max() <= 1 |
|
72
|
|
|
|
|
73
|
|
|
subject_label = model_label(transformed) |
|
74
|
|
|
transformed_back = subject_label.apply_inverse_transform() |
|
75
|
|
|
assert transformed_back.im.data.min() < 0 |
|
76
|
|
|
assert transformed_back.im.data.max() > 1 |
|
77
|
|
|
transformed_back_linear = subject_label.apply_inverse_transform( |
|
78
|
|
|
image_interpolation='nearest', |
|
79
|
|
|
) |
|
80
|
|
|
assert transformed_back_linear.im.data.unique().tolist() == [0, 1] |
|
81
|
|
|
|
fepegar /
torchio
