|
1
|
|
|
# coding=utf-8
|
|
2
|
|
|
|
|
3
|
|
|
from typing import List, Optional, Tuple, Union
|
|
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_util
|
|
10
|
|
|
from deepreg.model.backbone.u_net import UNet
|
|
11
|
|
|
from deepreg.registry import REGISTRY
|
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class AffineHead(tfkl.Layer):
|
|
15
|
|
|
def __init__(
|
|
16
|
|
|
self,
|
|
17
|
|
|
image_size: tuple,
|
|
18
|
|
|
name: str = "AffineHead",
|
|
19
|
|
|
):
|
|
20
|
|
|
"""
|
|
21
|
|
|
Init.
|
|
22
|
|
|
|
|
23
|
|
|
:param image_size: such as (dim1, dim2, dim3)
|
|
24
|
|
|
:param name: name of the layer
|
|
25
|
|
|
"""
|
|
26
|
|
|
super().__init__(name=name)
|
|
27
|
|
|
self.reference_grid = layer_util.get_reference_grid(image_size)
|
|
28
|
|
|
self.transform_initial = tf.constant_initializer(
|
|
29
|
|
|
value=list(np.eye(4, 3).reshape((-1)))
|
|
30
|
|
|
)
|
|
31
|
|
|
self._flatten = tfkl.Flatten()
|
|
32
|
|
|
self._dense = tfkl.Dense(units=12, bias_initializer=self.transform_initial)
|
|
33
|
|
|
|
|
34
|
|
|
def call(
|
|
35
|
|
|
self, inputs: Union[tf.Tensor, List], **kwargs
|
|
36
|
|
|
) -> Tuple[tf.Tensor, tf.Tensor]:
|
|
37
|
|
|
"""
|
|
38
|
|
|
|
|
39
|
|
|
:param inputs: a tensor or a list of tensor with length 1
|
|
40
|
|
|
:param kwargs: additional args
|
|
41
|
|
|
:return: ddf and theta
|
|
42
|
|
|
|
|
43
|
|
|
- ddf has shape (batch, dim1, dim2, dim3, 3)
|
|
44
|
|
|
- theta has shape (batch, 4, 3)
|
|
45
|
|
|
"""
|
|
46
|
|
|
if isinstance(inputs, list):
|
|
47
|
|
|
inputs = inputs[0]
|
|
48
|
|
|
theta = self._dense(self._flatten(inputs))
|
|
49
|
|
|
theta = tf.reshape(theta, shape=(-1, 4, 3))
|
|
50
|
|
|
# warp the reference grid with affine parameters to output a ddf
|
|
51
|
|
|
grid_warped = layer_util.warp_grid(self.reference_grid, theta)
|
|
52
|
|
|
ddf = grid_warped - self.reference_grid
|
|
53
|
|
|
return ddf, theta
|
|
54
|
|
|
|
|
55
|
|
|
def get_config(self):
|
|
56
|
|
|
"""Return the config dictionary for recreating this class."""
|
|
57
|
|
|
config = super().get_config()
|
|
58
|
|
|
config.update(image_size=self.reference_grid.shape[:3])
|
|
59
|
|
|
return config
|
|
60
|
|
|
|
|
61
|
|
|
|
|
62
|
|
|
@REGISTRY.register_backbone(name="global")
|
|
63
|
|
|
class GlobalNet(UNet):
|
|
64
|
|
|
"""
|
|
65
|
|
|
Build GlobalNet for image registration.
|
|
66
|
|
|
|
|
67
|
|
|
GlobalNet is a special UNet where the decoder for up-sampling is skipped.
|
|
68
|
|
|
The network's outputs come from the bottom layer from the encoder directly.
|
|
69
|
|
|
|
|
70
|
|
|
Reference:
|
|
71
|
|
|
|
|
72
|
|
|
- Hu, Yipeng, et al.
|
|
73
|
|
|
"Label-driven weakly-supervised learning
|
|
74
|
|
|
for multimodal deformable image registration,"
|
|
75
|
|
|
https://arxiv.org/abs/1711.01666
|
|
76
|
|
|
"""
|
|
77
|
|
|
|
|
78
|
|
|
def __init__(
|
|
79
|
|
|
self,
|
|
80
|
|
|
image_size: tuple,
|
|
81
|
|
|
num_channel_initial: int,
|
|
82
|
|
|
extract_levels: Optional[Tuple[int, ...]] = None,
|
|
83
|
|
|
depth: Optional[int] = None,
|
|
84
|
|
|
name: str = "GlobalNet",
|
|
85
|
|
|
**kwargs,
|
|
86
|
|
|
):
|
|
87
|
|
|
"""
|
|
88
|
|
|
Image is encoded gradually, i from level 0 to E.
|
|
89
|
|
|
Then, a densely-connected layer outputs an affine
|
|
90
|
|
|
transformation.
|
|
91
|
|
|
|
|
92
|
|
|
:param image_size: tuple, such as (dim1, dim2, dim3)
|
|
93
|
|
|
:param num_channel_initial: int, number of initial channels
|
|
94
|
|
|
:param extract_levels: list, which levels from net to extract, deprecated.
|
|
95
|
|
|
If depth is not given, depth = max(extract_levels) will be used.
|
|
96
|
|
|
:param depth: depth of the encoder. If given, extract_levels is not used.
|
|
97
|
|
|
:param name: name of the backbone.
|
|
98
|
|
|
:param kwargs: additional arguments.
|
|
99
|
|
|
"""
|
|
100
|
|
|
if depth is None:
|
|
101
|
|
|
if extract_levels is None:
|
|
102
|
|
|
raise ValueError(
|
|
103
|
|
|
"GlobalNet requires `depth` or `extract_levels` "
|
|
104
|
|
|
"to define the depth of encoder. "
|
|
105
|
|
|
"If `depth` is not given, "
|
|
106
|
|
|
"the maximum value of `extract_levels` will be used."
|
|
107
|
|
|
"However the argument `extract_levels` is deprecated "
|
|
108
|
|
|
"and will be removed in future release."
|
|
109
|
|
|
)
|
|
110
|
|
|
depth = max(extract_levels)
|
|
111
|
|
|
super().__init__(
|
|
112
|
|
|
image_size=image_size,
|
|
113
|
|
|
num_channel_initial=num_channel_initial,
|
|
114
|
|
|
depth=depth,
|
|
115
|
|
|
extract_levels=(depth,),
|
|
116
|
|
|
name=name,
|
|
117
|
|
|
**kwargs,
|
|
118
|
|
|
)
|
|
119
|
|
|
|
|
120
|
|
|
def build_output_block(
|
|
121
|
|
|
self,
|
|
122
|
|
|
image_size: Tuple[int, ...],
|
|
123
|
|
|
extract_levels: Tuple[int, ...],
|
|
124
|
|
|
out_channels: int,
|
|
125
|
|
|
out_kernel_initializer: str,
|
|
126
|
|
|
out_activation: str,
|
|
127
|
|
|
) -> Union[tf.keras.Model, tfkl.Layer]:
|
|
128
|
|
|
"""
|
|
129
|
|
|
Build a block for output.
|
|
130
|
|
|
|
|
131
|
|
|
The input to this block is a list of length 1.
|
|
132
|
|
|
The output has two tensors.
|
|
133
|
|
|
|
|
134
|
|
|
:param image_size: such as (dim1, dim2, dim3)
|
|
135
|
|
|
:param extract_levels: not used
|
|
136
|
|
|
:param out_channels: not used
|
|
137
|
|
|
:param out_kernel_initializer: not used
|
|
138
|
|
|
:param out_activation: not used
|
|
139
|
|
|
:return: a block consists of one or multiple layers
|
|
140
|
|
|
"""
|
|
141
|
|
|
return AffineHead(image_size=image_size)
|
|
142
|
|
|
|