|
1
|
|
|
import numpy as np |
|
2
|
|
|
from torch.utils.data import Dataset |
|
3
|
|
|
|
|
4
|
|
|
from ..sampler.sampler import PatchSampler |
|
5
|
|
|
from ...utils import to_tuple |
|
6
|
|
|
from ...torchio import LOCATION, TypeTuple, TypeTripletInt |
|
7
|
|
|
from ..subject import Subject |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
class GridSampler(PatchSampler, Dataset): |
|
11
|
|
|
r"""Extract patches across a whole volume. |
|
12
|
|
|
|
|
13
|
|
|
Grid samplers are useful to perform inference using all patches from a |
|
14
|
|
|
volume. It is often used with a :py:class:`~torchio.data.GridAggregator`. |
|
15
|
|
|
|
|
16
|
|
|
Args: |
|
17
|
|
|
sample: Instance of :py:class:`~torchio.data.subject.Subject` |
|
18
|
|
|
from which patches will be extracted. |
|
19
|
|
|
patch_size: Tuple of integers :math:`(d, h, w)` to generate patches |
|
20
|
|
|
of size :math:`d \times h \times w`. |
|
21
|
|
|
If a single number :math:`n` is provided, |
|
22
|
|
|
:math:`d = h = w = n`. |
|
23
|
|
|
patch_overlap: Tuple of integers :math:`(d_o, h_o, w_o)` specifying the |
|
24
|
|
|
overlap between patches for dense inference. If a single number |
|
25
|
|
|
:math:`n` is provided, :math:`d_o = h_o = w_o = n`. |
|
26
|
|
|
|
|
27
|
|
|
.. note:: Adapted from NiftyNet. See `this NiftyNet tutorial |
|
28
|
|
|
<https://niftynet.readthedocs.io/en/dev/window_sizes.html>`_ for more |
|
29
|
|
|
information. |
|
30
|
|
|
""" |
|
31
|
|
|
def __init__( |
|
32
|
|
|
self, |
|
33
|
|
|
sample: Subject, |
|
34
|
|
|
patch_size: TypeTuple, |
|
35
|
|
|
patch_overlap: TypeTuple, |
|
36
|
|
|
): |
|
37
|
|
|
self.sample = sample |
|
38
|
|
|
PatchSampler.__init__(self, patch_size) |
|
39
|
|
|
patch_size = to_tuple(patch_size, length=3) |
|
40
|
|
|
patch_overlap = to_tuple(patch_overlap, length=3) |
|
41
|
|
|
sizes = self.sample.spatial_shape, patch_size, patch_overlap |
|
42
|
|
|
self.parse_sizes(*sizes) |
|
43
|
|
|
self.locations = self.get_patches_locations(*sizes) |
|
44
|
|
|
|
|
45
|
|
|
def __len__(self): |
|
46
|
|
|
return len(self.locations) |
|
47
|
|
|
|
|
48
|
|
|
def __getitem__(self, index): |
|
49
|
|
|
# Assume 3D |
|
50
|
|
|
location = self.locations[index] |
|
51
|
|
|
index_ini = location[:3] |
|
52
|
|
|
index_fin = location[3:] |
|
53
|
|
|
cropped_sample = self.extract_patch(self.sample, index_ini, index_fin) |
|
54
|
|
|
cropped_sample[LOCATION] = location |
|
55
|
|
|
return cropped_sample |
|
56
|
|
|
|
|
57
|
|
|
@staticmethod |
|
58
|
|
|
def parse_sizes( |
|
59
|
|
|
image_size: TypeTripletInt, |
|
60
|
|
|
patch_size: TypeTripletInt, |
|
61
|
|
|
patch_overlap: TypeTripletInt, |
|
62
|
|
|
) -> None: |
|
63
|
|
|
image_size = np.array(image_size) |
|
64
|
|
|
patch_size = np.array(patch_size) |
|
65
|
|
|
patch_overlap = np.array(patch_overlap) |
|
66
|
|
|
if np.any(patch_size > image_size): |
|
67
|
|
|
message = ( |
|
68
|
|
|
f'Patch size {tuple(patch_size)} cannot be' |
|
69
|
|
|
f' larger than image size {tuple(image_size)}' |
|
70
|
|
|
) |
|
71
|
|
|
raise ValueError(message) |
|
72
|
|
|
if np.any(patch_overlap >= patch_size): |
|
73
|
|
|
message = ( |
|
74
|
|
|
f'Patch overlap {tuple(patch_overlap)} must be smaller' |
|
75
|
|
|
f' than patch size {tuple(image_size)}' |
|
76
|
|
|
) |
|
77
|
|
|
raise ValueError(message) |
|
78
|
|
|
|
|
79
|
|
|
def extract_patch( |
|
80
|
|
|
self, |
|
81
|
|
|
sample: Subject, |
|
82
|
|
|
index_ini: TypeTripletInt, |
|
83
|
|
|
index_fin: TypeTripletInt, |
|
84
|
|
|
) -> Subject: |
|
85
|
|
|
crop = self.get_crop_transform( |
|
86
|
|
|
sample.spatial_shape, |
|
87
|
|
|
index_ini, |
|
88
|
|
|
index_fin - index_ini, |
|
89
|
|
|
) |
|
90
|
|
|
cropped_sample = crop(sample) |
|
91
|
|
|
return cropped_sample |
|
92
|
|
|
|
|
93
|
|
|
@staticmethod |
|
94
|
|
|
def get_patches_locations( |
|
95
|
|
|
image_size: TypeTripletInt, |
|
96
|
|
|
patch_size: TypeTripletInt, |
|
97
|
|
|
patch_overlap: TypeTripletInt, |
|
98
|
|
|
) -> np.ndarray: |
|
99
|
|
|
indices = [] |
|
100
|
|
|
zipped = zip(image_size, patch_size, patch_overlap) |
|
101
|
|
|
for im_size_dim, patch_size_dim, patch_overlap_dim in zipped: |
|
102
|
|
|
end = im_size_dim + 1 - patch_size_dim |
|
103
|
|
|
step = patch_size_dim - patch_overlap_dim |
|
104
|
|
|
indices_dim = list(range(0, end, step)) |
|
105
|
|
|
if im_size_dim % step: |
|
106
|
|
|
indices_dim.append(im_size_dim - patch_size_dim) |
|
107
|
|
|
indices.append(indices_dim) |
|
108
|
|
|
indices_ini = np.array(np.meshgrid(*indices)).reshape(3, -1).T |
|
109
|
|
|
indices_ini = np.unique(indices_ini, axis=0) |
|
110
|
|
|
indices_fin = indices_ini + np.array(patch_size) |
|
111
|
|
|
locations = np.hstack((indices_ini, indices_fin)) |
|
112
|
|
|
return np.array(sorted(locations.tolist())) |
|
113
|
|
|
|
fepegar /
torchio
