|
1
|
|
|
# coding=utf-8
|
|
|
|
|
|
|
2
|
|
|
|
|
3
|
|
|
from typing import List
|
|
4
|
|
|
|
|
5
|
|
|
import numpy as np
|
|
6
|
|
|
import tensorflow as tf
|
|
7
|
|
|
import tensorflow.keras.layers as tfkl
|
|
8
|
|
|
|
|
9
|
|
|
from deepreg.model import layer, layer_util
|
|
10
|
|
|
from deepreg.model.backbone.interface import Backbone
|
|
11
|
|
|
from deepreg.registry import REGISTRY
|
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
@REGISTRY.register_backbone(name="global")
|
|
15
|
|
|
class GlobalNet(Backbone):
|
|
16
|
|
|
"""
|
|
17
|
|
|
Build GlobalNet for image registration.
|
|
18
|
|
|
|
|
19
|
|
|
Reference:
|
|
20
|
|
|
|
|
21
|
|
|
- Hu, Yipeng, et al.
|
|
22
|
|
|
"Label-driven weakly-supervised learning
|
|
23
|
|
|
for multimodal deformable image registration,"
|
|
24
|
|
|
https://arxiv.org/abs/1711.01666
|
|
25
|
|
|
"""
|
|
26
|
|
|
|
|
27
|
|
|
def __init__(
|
|
28
|
|
|
self,
|
|
29
|
|
|
image_size: tuple,
|
|
30
|
|
|
out_channels: int,
|
|
31
|
|
|
num_channel_initial: int,
|
|
32
|
|
|
extract_levels: List[int],
|
|
33
|
|
|
out_kernel_initializer: str,
|
|
34
|
|
|
out_activation: str,
|
|
35
|
|
|
name: str = "GlobalNet",
|
|
36
|
|
|
**kwargs,
|
|
37
|
|
|
):
|
|
38
|
|
|
"""
|
|
39
|
|
|
Image is encoded gradually, i from level 0 to E.
|
|
40
|
|
|
Then, a densely-connected layer outputs an affine
|
|
41
|
|
|
transformation.
|
|
42
|
|
|
|
|
43
|
|
|
:param image_size: tuple, such as (dim1, dim2, dim3)
|
|
44
|
|
|
:param out_channels: int, number of channels for the output
|
|
45
|
|
|
:param num_channel_initial: int, number of initial channels
|
|
46
|
|
|
:param extract_levels: list, which levels from net to extract
|
|
47
|
|
|
:param out_kernel_initializer: not used
|
|
48
|
|
|
:param out_activation: not used
|
|
49
|
|
|
:param name: name of the backbone.
|
|
50
|
|
|
:param kwargs: additional arguments.
|
|
51
|
|
|
"""
|
|
52
|
|
|
super().__init__(
|
|
53
|
|
|
image_size=image_size,
|
|
54
|
|
|
out_channels=out_channels,
|
|
55
|
|
|
num_channel_initial=num_channel_initial,
|
|
56
|
|
|
out_kernel_initializer=out_kernel_initializer,
|
|
57
|
|
|
out_activation=out_activation,
|
|
58
|
|
|
name=name,
|
|
59
|
|
|
**kwargs,
|
|
60
|
|
|
)
|
|
61
|
|
|
|
|
62
|
|
|
# save parameters
|
|
63
|
|
|
assert out_channels == 3
|
|
64
|
|
|
self._extract_levels = extract_levels
|
|
65
|
|
|
self._extract_max_level = max(self._extract_levels) # E
|
|
66
|
|
|
self.reference_grid = layer_util.get_reference_grid(image_size)
|
|
67
|
|
|
self.transform_initial = tf.constant_initializer(
|
|
68
|
|
|
value=list(np.eye(4, 3).reshape((-1)))
|
|
69
|
|
|
)
|
|
70
|
|
|
# init layer variables
|
|
71
|
|
|
num_channels = [
|
|
72
|
|
|
num_channel_initial * (2 ** level)
|
|
73
|
|
|
for level in range(self._extract_max_level + 1)
|
|
74
|
|
|
] # level 0 to E
|
|
75
|
|
|
self._downsample_convs = [
|
|
76
|
|
|
tf.keras.Sequential(
|
|
77
|
|
|
[
|
|
78
|
|
|
layer.Conv3dBlock(
|
|
79
|
|
|
filters=num_channels[i],
|
|
80
|
|
|
kernel_size=7 if i == 0 else 3,
|
|
81
|
|
|
padding="same",
|
|
82
|
|
|
),
|
|
83
|
|
|
layer.ResidualConv3dBlock(
|
|
84
|
|
|
filters=num_channels[i],
|
|
85
|
|
|
kernel_size=7 if i == 0 else 3,
|
|
86
|
|
|
padding="same",
|
|
87
|
|
|
),
|
|
88
|
|
|
]
|
|
89
|
|
|
)
|
|
90
|
|
|
for i in range(self._extract_max_level)
|
|
91
|
|
|
] # level 0 to E-1
|
|
92
|
|
|
self._downsample_pools = [
|
|
93
|
|
|
tfkl.MaxPool3D(pool_size=2, strides=2, padding="same")
|
|
94
|
|
|
for _ in range(self._extract_max_level)
|
|
95
|
|
|
] # level 0 to E-1
|
|
96
|
|
|
self._conv3d_block = layer.Conv3dBlock(
|
|
97
|
|
|
filters=num_channels[-1], kernel_size=3, padding="same"
|
|
98
|
|
|
) # level E
|
|
99
|
|
|
self._flatten = tfkl.Flatten()
|
|
100
|
|
|
self._dense_layer = tfkl.Dense(
|
|
101
|
|
|
units=12, bias_initializer=self.transform_initial
|
|
102
|
|
|
)
|
|
103
|
|
|
|
|
104
|
|
|
def call(
|
|
|
|
|
|
|
105
|
|
|
self, inputs: tf.Tensor, training=None, mask=None
|
|
106
|
|
|
) -> (tf.Tensor, tf.Tensor):
|
|
107
|
|
|
"""
|
|
108
|
|
|
Build GlobalNet graph based on built layers.
|
|
109
|
|
|
|
|
110
|
|
|
:param inputs: image batch, shape = (batch, f_dim1, f_dim2, f_dim3, ch)
|
|
111
|
|
|
:param training: None or bool.
|
|
112
|
|
|
:param mask: None or tf.Tensor.
|
|
113
|
|
|
:return:
|
|
114
|
|
|
ddf shape = (batch, dim1, dim2, dim3, 3)
|
|
115
|
|
|
theta shape = (batch, 4, 3)
|
|
116
|
|
|
"""
|
|
117
|
|
|
# down sample from level 0 to E
|
|
118
|
|
|
h_in = inputs
|
|
119
|
|
|
for level in range(self._extract_max_level): # level 0 to E - 1
|
|
120
|
|
|
skip = self._downsample_convs[level](inputs=h_in, training=training)
|
|
121
|
|
|
h_in = self._downsample_pools[level](inputs=skip)
|
|
122
|
|
|
h_out = self._conv3d_block(
|
|
123
|
|
|
inputs=h_in, training=training
|
|
124
|
|
|
) # level E of encoding
|
|
125
|
|
|
|
|
126
|
|
|
# predict affine parameters theta of shape = (batch, 4, 3)
|
|
127
|
|
|
theta = self._dense_layer(self._flatten(h_out))
|
|
128
|
|
|
theta = tf.reshape(theta, shape=(-1, 4, 3))
|
|
129
|
|
|
# warp the reference grid with affine parameters to output a ddf
|
|
130
|
|
|
grid_warped = layer_util.warp_grid(self.reference_grid, theta)
|
|
131
|
|
|
ddf = grid_warped - self.reference_grid
|
|
132
|
|
|
return ddf, theta
|
|
133
|
|
|
|
DeepRegNet /
DeepReg
Loading history...