|
1
|
|
|
import warnings |
|
2
|
|
|
from typing import Tuple |
|
3
|
|
|
import torch |
|
4
|
|
|
import numpy as np |
|
5
|
|
|
from ...torchio import TypeData, CHANNELS_DIMENSION |
|
6
|
|
|
from .grid_sampler import GridSampler |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
class GridAggregator: |
|
10
|
|
|
r"""Aggregate patches for dense inference. |
|
11
|
|
|
|
|
12
|
|
|
This class is typically used to build a volume made of batches after |
|
13
|
|
|
inference of patches extracted by a :py:class:`~torchio.data.GridSampler`. |
|
14
|
|
|
|
|
15
|
|
|
Args: |
|
16
|
|
|
sampler: Instance of :py:class:`~torchio.data.GridSampler` used to |
|
17
|
|
|
extract the patches. |
|
18
|
|
|
|
|
19
|
|
|
.. note:: Adapted from NiftyNet. See `this NiftyNet tutorial |
|
20
|
|
|
<https://niftynet.readthedocs.io/en/dev/window_sizes.html>`_ for more |
|
21
|
|
|
information about patch based sampling. |
|
22
|
|
|
""" |
|
23
|
|
|
def __init__(self, sampler: GridSampler): |
|
24
|
|
|
sample = sampler.sample |
|
25
|
|
|
self.volume_padded = sampler.padding_mode is not None |
|
26
|
|
|
self.spatial_shape = sample.spatial_shape |
|
27
|
|
|
self._output_tensor = None |
|
28
|
|
|
self.patch_overlap = sampler.patch_overlap |
|
29
|
|
|
|
|
30
|
|
|
def crop_batch( |
|
31
|
|
|
self, |
|
32
|
|
|
batch: torch.Tensor, |
|
33
|
|
|
locations: np.ndarray, |
|
34
|
|
|
overlap: np.ndarray, |
|
35
|
|
|
) -> Tuple[TypeData, np.ndarray]: |
|
36
|
|
|
border = np.array(overlap) // 2 # overlap is even in grid sampler |
|
37
|
|
|
crop_locations = locations.astype(int).copy() |
|
38
|
|
|
indices_ini, indices_fin = crop_locations[:, :3], crop_locations[:, 3:] |
|
39
|
|
|
num_locations = len(crop_locations) |
|
40
|
|
|
|
|
41
|
|
|
# Do not crop patches at the border of the volume |
|
42
|
|
|
# Unless we're padding the volume in the grid sampler. In that case, |
|
43
|
|
|
# it doesn't matter if we don't crop patches at the border, because the |
|
44
|
|
|
# output volume will be cropped |
|
45
|
|
|
if not self.volume_padded: |
|
46
|
|
|
border_ini = np.tile(border, (num_locations, 1)) |
|
47
|
|
|
border_fin = border_ini.copy() |
|
48
|
|
|
mask_border_ini = indices_ini == 0 |
|
49
|
|
|
border_ini[mask_border_ini] = 0 |
|
50
|
|
|
for axis, size in enumerate(self.spatial_shape): |
|
51
|
|
|
mask_border_fin = indices_fin[:, axis] == size |
|
52
|
|
|
border_fin[mask_border_fin, axis] = 0 |
|
53
|
|
|
indices_ini += border_ini |
|
54
|
|
|
indices_fin -= border_fin |
|
55
|
|
|
|
|
56
|
|
|
crop_shapes = indices_fin - indices_ini |
|
57
|
|
|
patch_shape = batch.shape[2:] # ignore batch and channels dim |
|
58
|
|
|
cropped_patches = [] |
|
59
|
|
|
for patch, crop_shape in zip(batch, crop_shapes): |
|
60
|
|
|
diff = patch_shape - crop_shape |
|
61
|
|
|
left = (diff / 2).astype(int) |
|
62
|
|
|
i_ini, j_ini, k_ini = left |
|
63
|
|
|
i_fin, j_fin, k_fin = left + crop_shape |
|
64
|
|
|
cropped_patch = patch[:, i_ini:i_fin, j_ini:j_fin, k_ini:k_fin] |
|
65
|
|
|
cropped_patches.append(cropped_patch) |
|
66
|
|
|
return cropped_patches, crop_locations |
|
67
|
|
|
|
|
68
|
|
|
def initialize_output_tensor(self, batch: torch.Tensor) -> None: |
|
69
|
|
|
if self._output_tensor is not None: |
|
70
|
|
|
return |
|
71
|
|
|
num_channels = batch.shape[CHANNELS_DIMENSION] |
|
72
|
|
|
self._output_tensor = torch.zeros( |
|
73
|
|
|
num_channels, |
|
74
|
|
|
*self.spatial_shape, |
|
75
|
|
|
dtype=batch.dtype, |
|
76
|
|
|
) |
|
77
|
|
|
|
|
78
|
|
|
def add_batch( |
|
79
|
|
|
self, |
|
80
|
|
|
batch_tensor: torch.Tensor, |
|
81
|
|
|
locations: torch.Tensor, |
|
82
|
|
|
) -> None: |
|
83
|
|
|
"""Add batch processed by a CNN to the output prediction volume. |
|
84
|
|
|
|
|
85
|
|
|
Args: |
|
86
|
|
|
batch_tensor: 5D tensor, typically the output of a convolutional |
|
87
|
|
|
neural network, e.g. ``batch['image'][torchio.DATA]``. |
|
88
|
|
|
locations: 2D tensor with shape :math:`(B, 6)` representing the |
|
89
|
|
|
patch indices in the original image. They are typically |
|
90
|
|
|
extracted using ``batch[torchio.LOCATION]``. |
|
91
|
|
|
""" |
|
92
|
|
|
batch = batch_tensor.cpu() |
|
93
|
|
|
locations = locations.cpu().numpy() |
|
94
|
|
|
self.initialize_output_tensor(batch) |
|
95
|
|
|
cropped_patches, crop_locations = self.crop_batch( |
|
96
|
|
|
batch, |
|
97
|
|
|
locations, |
|
98
|
|
|
self.patch_overlap, |
|
99
|
|
|
) |
|
100
|
|
|
for patch, crop_location in zip(cropped_patches, crop_locations): |
|
101
|
|
|
i_ini, j_ini, k_ini, i_fin, j_fin, k_fin = crop_location |
|
102
|
|
|
self._output_tensor[ |
|
103
|
|
|
:, |
|
104
|
|
|
i_ini:i_fin, |
|
105
|
|
|
j_ini:j_fin, |
|
106
|
|
|
k_ini:k_fin] = patch |
|
107
|
|
|
|
|
108
|
|
|
def get_output_tensor(self) -> torch.Tensor: |
|
109
|
|
|
"""Get the aggregated volume after dense inference.""" |
|
110
|
|
|
if self._output_tensor.dtype == torch.int64: |
|
111
|
|
|
message = ( |
|
112
|
|
|
'Medical image frameworks such as ITK do not support int64.' |
|
113
|
|
|
' Casting to int32...' |
|
114
|
|
|
) |
|
115
|
|
|
warnings.warn(message) |
|
116
|
|
|
self._output_tensor = self._output_tensor.type(torch.int32) |
|
117
|
|
|
if self.volume_padded: |
|
118
|
|
|
from ...transforms import Crop |
|
119
|
|
|
border = self.patch_overlap // 2 |
|
120
|
|
|
cropping = border.repeat(2) |
|
121
|
|
|
crop = Crop(cropping) |
|
122
|
|
|
return crop(self._output_tensor) |
|
123
|
|
|
else: |
|
124
|
|
|
return self._output_tensor |
|
125
|
|
|
|
fepegar /
torchio
