-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
367 lines (294 loc) · 11.9 KB
/
models.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
'''
VoxelMorph
Original code retrieved from:
https://github.com/voxelmorph/voxelmorph
Original paper:
Balakrishnan, G., Zhao, A., Sabuncu, M. R., Guttag, J., & Dalca, A. V. (2019).
VoxelMorph: a learning framework for deformable medical image registration.
IEEE transactions on medical imaging, 38(8), 1788-1800.
Modified and tested by:
Haiqiao Wang
2110246069@email.szu.edu.cn
Shenzhen University
'''
import torch
import torch.nn as nn
import torch.nn.functional as nnf
import numpy as np
from torch.distributions.normal import Normal
class SpatialTransformer(nn.Module):
"""
N-D Spatial Transformer
"""
def __init__(self, size, mode='bilinear'):
super().__init__()
self.mode = mode
# create sampling grid
vectors = [torch.arange(0, s) for s in size]
grids = torch.meshgrid(vectors)
grid = torch.stack(grids)
grid = torch.unsqueeze(grid, 0)
grid = grid.type(torch.FloatTensor)
# registering the grid as a buffer cleanly moves it to the GPU, but it also
# adds it to the state dict. this is annoying since everything in the state dict
# is included when saving weights to disk, so the model files are way bigger
# than they need to be. so far, there does not appear to be an elegant solution.
# see: https://discuss.pytorch.org/t/how-to-register-buffer-without-polluting-state-dict
self.register_buffer('grid', grid)
def forward(self, src, flow):
# new locations
new_locs = self.grid + flow
shape = flow.shape[2:]
# need to normalize grid values to [-1, 1] for resampler
for i in range(len(shape)):
new_locs[:, i, ...] = 2 * (new_locs[:, i, ...] / (shape[i] - 1) - 0.5)
# move channels dim to last position
# also not sure why, but the channels need to be reversed
if len(shape) == 2:
new_locs = new_locs.permute(0, 2, 3, 1)
new_locs = new_locs[..., [1, 0]]
elif len(shape) == 3:
new_locs = new_locs.permute(0, 2, 3, 4, 1)
new_locs = new_locs[..., [2, 1, 0]]
return nnf.grid_sample(src, new_locs, align_corners=True, mode=self.mode)
class VecInt(nn.Module):
"""
Integrates a vector field via scaling and squaring.
"""
def __init__(self, inshape, nsteps=7):
super().__init__()
assert nsteps >= 0, 'nsteps should be >= 0, found: %d' % nsteps
self.nsteps = nsteps
self.scale = 1.0 / (2 ** self.nsteps)
self.transformer = SpatialTransformer(inshape)
def forward(self, vec):
vec = vec * self.scale
for _ in range(self.nsteps):
vec = vec + self.transformer(vec, vec)
return vec
class ResizeTransform(nn.Module):
def __init__(self, vel_resize, ndims):
super().__init__()
self.factor = 1.0 / vel_resize
self.mode = 'linear'
if ndims == 2:
self.mode = 'bi' + self.mode
elif ndims == 3:
self.mode = 'tri' + self.mode
def forward(self, x):
if self.factor < 1:
# resize first to save memory
x = nnf.interpolate(x, align_corners=True, scale_factor=self.factor, mode=self.mode)
x = self.factor * x
elif self.factor > 1:
# multiply first to save memory
x = self.factor * x
x = nnf.interpolate(x, align_corners=True, scale_factor=self.factor, mode=self.mode)
# don't do anything if resize is 1
return x
class ConvBlock(nn.Module):
"""
Specific convolutional block followed by leakyrelu for unet.
"""
def __init__(self, ndims, in_channels, out_channels,kernal_size=3, stride=1, padding=1, alpha=0.1):
super().__init__()
Conv = getattr(nn, 'Conv%dd' % ndims)
self.main = Conv(in_channels, out_channels, kernal_size, stride, padding)
self.activation = nn.LeakyReLU(alpha)
def forward(self, x):
out = self.main(x)
out = self.activation(out)
return out
class ConvInsBlock(nn.Module):
"""
Specific convolutional block followed by leakyrelu for unet.
"""
def __init__(self, in_channels, out_channels,kernal_size=3, stride=1, padding=1, alpha=0.1):
super().__init__()
self.main = nn.Conv3d(in_channels, out_channels, kernal_size, stride, padding)
self.norm = nn.InstanceNorm3d(out_channels)
self.activation = nn.LeakyReLU(alpha)
def forward(self, x):
out = self.main(x)
out = self.norm(out)
out = self.activation(out)
return out
class UpConvBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=4, stride=2, alpha=0.1):
super(UpConvBlock, self).__init__()
self.upconv = nn.ConvTranspose3d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=1)
self.actout = nn.Sequential(
nn.InstanceNorm3d(out_channels),
nn.LeakyReLU(alpha)
)
def forward(self, x):
x = self.upconv(x)
return self.actout(x)
class ResBlock(nn.Module):
"""
VoxRes module
"""
def __init__(self, channel, alpha=0.1):
super(ResBlock, self).__init__()
self.block = nn.Sequential(
nn.InstanceNorm3d(channel),
nn.LeakyReLU(alpha),
nn.Conv3d(channel, channel, kernel_size=3, padding=1)
)
self.actout = nn.Sequential(
nn.InstanceNorm3d(channel),
nn.LeakyReLU(alpha),
)
def forward(self, x):
out = self.block(x) + x
return self.actout(out)
class Encoder(nn.Module):
"""
Main model
"""
def __init__(self, in_channel=1, first_out_channel=16):
super(Encoder, self).__init__()
c = first_out_channel
self.conv0 = ConvInsBlock(in_channel, c, 3, 1)
self.conv1 = nn.Sequential(
nn.Conv3d(c, 2*c, kernel_size=3, stride=2, padding=1),#80
ResBlock(2*c)
)
self.conv2 = nn.Sequential(
nn.Conv3d(2*c, 4*c, kernel_size=3, stride=2, padding=1),#40
ResBlock(4*c)
)
self.conv3 = nn.Sequential(
nn.Conv3d(4*c, 8*c, kernel_size=3, stride=2, padding=1),#20
ResBlock(8*c)
)
def forward(self, x):
out0 = self.conv0(x) # 1
out1 = self.conv1(out0) # 1/2
out2 = self.conv2(out1) # 1/4
out3 = self.conv3(out2) # 1/8
return [out0, out1, out2, out3]
class CConv(nn.Module):
def __init__(self, channel):
super(CConv, self).__init__()
c = channel
self.conv = nn.Sequential(
ConvInsBlock(c, c, 3, 1),
ConvInsBlock(c, c, 3, 1)
)
def forward(self, float_fm, fixed_fm, d_fm):
concat_fm = torch.cat([float_fm, fixed_fm, d_fm], dim=1)
x = self.conv(concat_fm)
return x
class RDP(nn.Module):
def __init__(self, inshape=(160,192,160), flow_multiplier=1.,in_channel=1, channels=16):
super(RDP, self).__init__()
self.flow_multiplier = flow_multiplier
self.channels = channels
self.step = 7
self.inshape = inshape
c = self.channels
self.encoder_moving = Encoder(in_channel=in_channel, first_out_channel=c)
self.encoder_fixed = Encoder(in_channel=in_channel, first_out_channel=c)
self.upsample = nn.Upsample(scale_factor=2, mode='nearest')
self.upsample_trilin = nn.Upsample(scale_factor=2, mode='trilinear', align_corners=True)#nn.Upsample(scale_factor=2, mode='trilinear', align_corners=True)
self.warp = nn.ModuleList()
self.diff = nn.ModuleList()
for i in range(4):
self.warp.append(SpatialTransformer([s // 2**i for s in inshape]))
self.diff.append(VecInt([s // 2**i for s in inshape]))
# bottleNeck
self.cconv_4 = nn.Sequential(
ConvInsBlock(16 * c, 8 * c, 3, 1),
ConvInsBlock(8 * c, 8 * c, 3, 1)
)
# warp scale 2
self.defconv4 = nn.Conv3d(8*c, 3, 3, 1, 1)
self.defconv4.weight = nn.Parameter(Normal(0, 1e-5).sample(self.defconv4.weight.shape))
self.defconv4.bias = nn.Parameter(torch.zeros(self.defconv4.bias.shape))
self.dconv4 = nn.Sequential(
ConvInsBlock(3*8*c, 8*c),
ConvInsBlock(8*c, 8*c)
)
self.upconv3 = UpConvBlock(8*c, 4*c, 4, 2)
self.cconv_3 = CConv(3*4*c)
# warp scale 1
self.defconv3 = nn.Conv3d(3*4*c, 3, 3, 1, 1)
self.defconv3.weight = nn.Parameter(Normal(0, 1e-5).sample(self.defconv3.weight.shape))
self.defconv3.bias = nn.Parameter(torch.zeros(self.defconv3.bias.shape))
self.dconv3 = ConvInsBlock(3 * 4 * c, 4 * c)
self.upconv2 = UpConvBlock(3*4*c, 2*c, 4, 2)
self.cconv_2 = CConv(3*2*c)
# warp scale 0
self.defconv2 = nn.Conv3d(3*2*c, 3, 3, 1, 1)
self.defconv2.weight = nn.Parameter(Normal(0, 1e-5).sample(self.defconv2.weight.shape))
self.defconv2.bias = nn.Parameter(torch.zeros(self.defconv2.bias.shape))
self.dconv2 = ConvInsBlock(3 * 2 * c, 2 * c)
self.upconv1 = UpConvBlock(3*2*c, c, 4, 2)
self.cconv_1 = CConv(3*c)
# decoder layers
self.defconv1 = nn.Conv3d(3*c, 3, 3, 1, 1)
self.defconv1.weight = nn.Parameter(Normal(0, 1e-5).sample(self.defconv1.weight.shape))
self.defconv1.bias = nn.Parameter(torch.zeros(self.defconv1.bias.shape))
#self.dconv1 = ConvInsBlock(3 * c, c)
def forward(self, moving, fixed):
# encode stage
M1, M2, M3, M4 = self.encoder_moving(moving)
F1, F2, F3, F4 = self.encoder_fixed(fixed)
# c=16, 2c, 4c, 8c # 160, 80, 40, 20
# first dec layer
C4 = torch.cat([F4, M4], dim=1)
C4 = self.cconv_4(C4) # (1,128,20,24,20)
flow = self.defconv4(C4) # (1,3,20,24,20)
flow = self.diff[3](flow)
warped = self.warp[3](M4, flow)
C4 = self.dconv4(torch.cat([F4, warped, C4], dim=1))
v = self.defconv4(C4) # (1,3,20,24,20)
w = self.diff[3](v)
D3 = self.upconv3(C4) # (1, 64, 40, 48, 40)
flow = self.upsample_trilin(2*(self.warp[3](flow, w)+w))
warped = self.warp[2](M3, flow) # (1, 64, 40, 48, 40)
C3 = self.cconv_3(F3, warped, D3) # (1, 3 * 64, 40, 48, 40)
v = self.defconv3(C3)
w = self.diff[2](v)
flow = self.warp[2](flow, w)+w
warped = self.warp[2](M3, flow) # (1, 64, 40, 48, 40)
D3 = self.dconv3(C3)
C3 = self.cconv_3(F3, warped, D3) # (1, 3 * 64, 40, 48, 40)
v = self.defconv3(C3)
w = self.diff[2](v)
D2 = self.upconv2(C3)
flow = self.upsample_trilin(2*(self.warp[2](flow, w)+w))
warped = self.warp[1](M2, flow)
C2 = self.cconv_2(F2, warped, D2)
v = self.defconv2(C2) # (1,3,80,96,80)
w = self.diff[1](v)
flow = self.warp[1](flow, w)+w
warped = self.warp[1](M2, flow)
D2 = self.dconv2(C2)
C2 = self.cconv_2(F2, warped, D2)
v = self.defconv2(C2) # (1,3,80,96,80)
w = self.diff[1](v)
flow = self.warp[1](flow, w)+w
warped = self.warp[1](M2, flow)
D2 = self.dconv2(C2)
C2 = self.cconv_2(F2, warped, D2)
v = self.defconv2(C2) # (1,3,80,96,80)
w = self.diff[1](v)
D1 = self.upconv1(C2) # (1,16,160,196,160)
flow = self.upsample_trilin(2*(self.warp[1](flow, w)+w)) # (1,3,160,196,160)
warped = self.warp[0](M1, flow) # (1,16,160,196,160)
C1 = self.cconv_1(F1, warped, D1) # (1,48,160,196,160)
v = self.defconv1(C1)
w = self.diff[0](v)
flow = self.warp[0](flow, w)+w # (1,3,160,196,160)
y_moved = self.warp[0](moving, flow)
return y_moved, flow
if __name__ == '__main__':
size = (1, 1, 80, 96, 80)
model = RDP(size[2:])
# print(str(model))
A = torch.ones(size)
B = torch.ones(size)
out, flow = model(A, B)
print(out.shape, flow.shape)