|
1
|
|
|
import numpy as np |
|
2
|
|
|
import torch |
|
3
|
|
|
|
|
4
|
|
|
from ....data.image import Image |
|
5
|
|
|
from ....data.subject import Subject |
|
6
|
|
|
from ...spatial_transform import SpatialTransform |
|
7
|
|
|
from .resample import Resample |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
class ToReferenceSpace(SpatialTransform): |
|
11
|
|
|
"""Modify the spatial metadata so it matches a reference space. |
|
12
|
|
|
|
|
13
|
|
|
This is useful, for example, to set meaningful spatial metadata of a neural |
|
14
|
|
|
network embedding, for visualization or further processing such as |
|
15
|
|
|
resampling a segmentation output. |
|
16
|
|
|
|
|
17
|
|
|
Example: |
|
18
|
|
|
|
|
19
|
|
|
>>> import torchio as tio |
|
20
|
|
|
>>> image = tio.datasets.FPG().t1 |
|
21
|
|
|
>>> embedding_tensor = my_network(image.tensor) # we lose metadata here |
|
22
|
|
|
>>> embedding_image = tio.ToReferenceSpace.from_tensor(embedding_tensor, image) |
|
23
|
|
|
""" |
|
24
|
|
|
|
|
25
|
|
|
def __init__(self, reference: Image, **kwargs): |
|
26
|
|
|
super().__init__(**kwargs) |
|
27
|
|
|
if not isinstance(reference, Image): |
|
28
|
|
|
raise TypeError('The reference must be a TorchIO image') |
|
29
|
|
|
self.reference = reference |
|
30
|
|
|
|
|
31
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
|
32
|
|
|
for image in self.get_images(subject): |
|
33
|
|
|
new_image = build_image_from_reference(image.data, self.reference) |
|
34
|
|
|
image.set_data(new_image.data) |
|
35
|
|
|
image.affine = new_image.affine |
|
36
|
|
|
return subject |
|
37
|
|
|
|
|
38
|
|
|
@staticmethod |
|
39
|
|
|
def from_tensor(tensor: torch.Tensor, reference: Image) -> Image: |
|
40
|
|
|
"""Build a TorchIO image from a tensor and a reference image.""" |
|
41
|
|
|
return build_image_from_reference(tensor, reference) |
|
42
|
|
|
|
|
43
|
|
|
|
|
44
|
|
|
def build_image_from_reference(tensor: torch.Tensor, reference: Image) -> Image: |
|
45
|
|
|
input_shape = np.array(reference.spatial_shape) |
|
46
|
|
|
output_shape = np.array(tensor.shape[-3:]) |
|
47
|
|
|
downsampling_factor = input_shape / output_shape |
|
48
|
|
|
input_spacing = np.array(reference.spacing) |
|
49
|
|
|
output_spacing = input_spacing * downsampling_factor |
|
50
|
|
|
downsample = Resample(output_spacing, image_interpolation='nearest') |
|
51
|
|
|
reference = downsample(reference) |
|
52
|
|
|
class_ = reference.__class__ |
|
53
|
|
|
result = class_(tensor=tensor, affine=reference.affine) |
|
54
|
|
|
return result |
|
55
|
|
|
|