-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathMPUM.py
executable file
·235 lines (211 loc) · 10.7 KB
/
MPUM.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
from torch import nn
import numpy as np
import torch
import einops
from categories import prediction as prediction_label_tissue
class Controller(nn.Module):
def __init__(self, config,in_channel,out_channel,stride=1):
super().__init__()
self.config = config
if config["modelsize"] == "tiny":
self.latent = 32
elif config["modelsize"] == "base":
self.latent = 64
elif config["modelsize"] == "huge":
self.latent = 128
self.conv1 = nn.Conv3d(in_channel,self.latent,bias=False,kernel_size=1)
if config["modality"] == "ct":
self.fc1 = nn.Linear(config["modalitydimension"], self.latent * 3 * 3 * 3)
elif config["modality"] == "mr":
self.fc1 = nn.Linear(config["modalitydimension"], self.latent * 3 * 3 * 3)
elif config["modality"] == "pet":
self.fc1 = nn.Linear(config["modalitydimension"], self.latent * 3 * 3 * 3)
self.conv2 = nn.Conv3d(config["tissuenumber"],out_channel,bias=False,kernel_size=3,padding=1,stride=stride)
self.in_channel = in_channel
def forward(self,orix,controllerspace):
x = self.conv1(orix)
kernel = self.fc1(controllerspace)
kernel = einops.rearrange(kernel, 'a (b c d e) -> a b c d e', b=self.latent, c=3,d=3,e=3)
saliency = nn.functional.conv3d(x,kernel,None,1,1)
x = self.conv2(saliency)
return x,saliency
class Block(nn.Module):
def __init__(self, config,in_channel, out_channel, kernel_size=3, padding=1, stride=1):
super().__init__()
self.config = config
self.in_channel = in_channel
self.out_channel = out_channel
# branch1
self.conv1 = nn.Conv3d(in_channel, out_channel, kernel_size, stride=stride, padding=padding)
self.norm1 = nn.InstanceNorm3d(out_channel, affine=True,track_running_stats=False)
self.act1 = nn.LeakyReLU(inplace=True)
# branch2
self.tissueTUC = Controller(config,in_channel,config["tissuenumber"],stride=stride)
# branch3
# self.lesionTUC = Controller(config,in_channel,config["tissuenumber"],stride=stride)
self.tail = nn.Conv3d(out_channel+ config["tissuenumber"],out_channel,kernel_size=1,padding=0,stride=1)
if stride == 2 or (stride==1 and in_channel != out_channel):
self.res = nn.Conv3d(in_channel, out_channel, kernel_size, stride=stride, padding=padding)
else:
self.res = None
self.norm2 = nn.InstanceNorm3d(out_channel, affine=True,track_running_stats=False)
self.act2 = nn.LeakyReLU(inplace=True)
def forward(self, orix,controllerspace,lastblock=False):
# branch1
x1 = self.conv1(orix)
x1 = self.act1(self.norm1(x1))
# branch2
x2,saliency = self.tissueTUC(orix,controllerspace)
# branch3
# x3 = self.lesionTUC(orix)
if lastblock:
return None,x2
x12 = torch.concat([x1,x2],axis=1)
x = self.tail(x12)
if self.res:
x += self.res(orix)
else:
x += x
x = self.act2(self.norm2(x))
return x,saliency
class MPUM(nn.Module):
def __init__(self,config,tissueclip, in_channel):
super().__init__()
self.input_channels = in_channel
self._deep_supervision = False
self.config = config
self.register_buffer('tissueclip', tissueclip)
# self.register_buffer('lesionclip', lesionclip)
self.backprojection_matrix_tissue = nn.Sequential(
nn.Linear(self.tissueclip.shape[1],1024),
)
# self.backprojection_matrix_lesion = nn.Sequential(
# nn.Linear(self.tissueclip.shape[1]+self.lesionclip.shape[1],1024),
# )
if config["modelsize"] == "tiny":
self.channels = [32,64,128,256]
elif config["modelsize"] == "base":
self.channels = [64,128,256,512]
elif config["modelsize"] == "huge":
self.channels = [128,256,512,1024]
self.head = nn.Conv3d(1,self.channels[0],kernel_size=3,padding=1,stride=2)
self.Block1_2 = Block(config,self.channels[0],self.channels[0])
self.Block2_1 = Block(config,self.channels[0],self.channels[1],stride=2)
self.Block2_2 = Block(config,self.channels[1],self.channels[1] )
self.Block3_1 = Block(config,self.channels[1],self.channels[2] ,stride=2)
self.Block3_2 = Block(config,self.channels[2],self.channels[2] )
self.Block4_1 = Block(config,self.channels[2],self.channels[3] ,stride=2)
self.Block4_2 = Block(config,self.channels[3],self.channels[3])
# self.Block5_1 = Block(self.channels[3],self.channels[4] ,stride=2)
# self.Block5_2 = Block(self.channels[4],self.channels[4] )
self.upsample = nn.Upsample(scale_factor=2)
# self.Block4_3 = Block(self.channels[4] + self.channels[3],self.channels[3] )
# self.Block4_4 = Block(self.channels[3],self.channels[3] )
self.Block3_3 = Block(config,self.channels[3] + self.channels[2],self.channels[2] )
self.Block3_4 = Block(config,self.channels[2],self.channels[2])
self.Block2_3 = Block(config,self.channels[2] + self.channels[1],self.channels[1] )
self.Block2_4 = Block(config,self.channels[1],self.channels[1])
self.Block1_3 = Block(config,self.channels[1] + self.channels[0],self.channels[0] )
self.Block1_4 = Block(config,self.channels[0],self.channels[0])
self.ct_projection_matrix = nn.Linear(1024,config["modalitydimension"],bias=False)
self.pet_projection_matrix = nn.Linear(1024,config["modalitydimension"],bias=False)
self.mr_projection_matrix = nn.Linear(1024,config["modalitydimension"],bias=False)
self.saliency=False
def forward(self, x):
controllerspace = self.backprojection_matrix_tissue(self.tissueclip)
if self.config["modality"] == "ct":
controllerspace = self.ct_projection_matrix(controllerspace)
elif self.config["modality"] == "pet":
controllerspace = self.pet_projection_matrix(controllerspace)
elif self.config["modality"] == "mr":
controllerspace = self.mr_projection_matrix(controllerspace)
# deep_supervision = []
x1_1 = self.head(x)
x1_2,o7 = self.Block1_2(x1_1,controllerspace)
x2_1,_ = self.Block2_1(x1_2,controllerspace)
x2_2,o6 = self.Block2_2(x2_1,controllerspace)
x3_1,_ = self.Block3_1(x2_2,controllerspace)
x3_2,o5 = self.Block3_2(x3_1,controllerspace)
x4_1,_ = self.Block4_1(x3_2,controllerspace)
x4_2,o4 = self.Block4_2(x4_1,controllerspace)
# x5_1,_ = self.Block5_1(x4_2,ctembedding)
# x5_2,o5 = self.Block5_2(x5_1,ctembedding)
# x4_3,_ = self.Block4_3(torch.concat([self.upsample(x5_2),x4_2],dim=1),ctembedding)
# x4_4,o4 = self.Block4_4(x4_3,ctembedding)
x3_3,_ = self.Block3_3(torch.concat([self.upsample(x4_2),x3_2],dim=1),controllerspace)
x3_4,o3 = self.Block3_4(x3_3,controllerspace)
x2_3,_ = self.Block2_3(torch.concat([self.upsample(x3_4),x2_2],dim=1),controllerspace)
x2_4,o2 = self.Block2_4(x2_3,controllerspace)
x1_3,_ = self.Block1_3(torch.concat([self.upsample(x2_4),x1_2],dim=1),controllerspace)
_,o1 = self.Block1_4(self.upsample(x1_3),controllerspace,lastblock=True)
# print(self.saliency)
# if self.saliency:
# return o1,[o1,o2,o3,o4]
return o1
# else:
# return o1
# if self._deep_supervision:
# return o1,[o2,o3,o4,o5,o6,o7]
# else:
# return o1
# def load_model(self, checkpoint):
# # 获取当前模型的状态字典
# current_state = self.state_dict()
# # 为每个不匹配的权重进行切片操作
# for key, param in checkpoint.items():
# if key in current_state:
# if param.size() != current_state[key].size():
# # 进行切片以匹配模型中的相应参数形状
# slices = tuple(slice(0, min(param.size(i), current_state[key].size(i))) for i in range(len(param.size())))
# param = param[slices]
# print(f"Adjusted {key} from {checkpoint[key].size()} to {param.size()}")
# current_state[key] = param
# else:
# print(f"Skipping {key} because it's not in the current model.")
# self.load_state_dict(current_state,strict=False)
def get_unispace(self):
return self.backprojection_matrix_tissue(self.tissueclip)
def get_ctspace(self):
controllerspace = self.backprojection_matrix_tissue(self.tissueclip)
controllerspace = self.ct_projection_matrix(controllerspace)
return controllerspace
def get_mrspace(self):
controllerspace = self.backprojection_matrix_tissue(self.tissueclip)
controllerspace = self.mr_projection_matrix(controllerspace)
return controllerspace
def get_petspace(self):
controllerspace = self.backprojection_matrix_tissue(self.tissueclip)
controllerspace = self.pet_projection_matrix(controllerspace)
return controllerspace
def get_label(self):
return prediction_label_tissue
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
# 输入维度：(batch_size, 133, 128, 128, 128)
self.features = nn.Sequential(
# 第一层减少空间维度
nn.Conv3d(133, 64, kernel_size=3, stride=2, padding=1), # 输出：(batch_size, 64, 64, 64, 64)
nn.LeakyReLU(0.2, inplace=True),
# 进一步降低维度和增加通道数
nn.Conv3d(64, 128, kernel_size=3, stride=2, padding=1), # 输出：(batch_size, 128, 32, 32, 32)
nn.InstanceNorm3d(128),
nn.LeakyReLU(0.2, inplace=True),
# 继续减小维度
nn.Conv3d(128, 256, kernel_size=3, stride=2, padding=1), # 输出：(batch_size, 256, 16, 16, 16)
nn.InstanceNorm3d(256),
nn.LeakyReLU(0.2, inplace=True),
# 最后一层，转换为更小的空间尺寸
nn.Conv3d(256, 512, kernel_size=3, stride=2, padding=1), # 输出：(batch_size, 512, 8, 8, 8)
nn.InstanceNorm3d(512),
nn.LeakyReLU(0.2, inplace=True),
)
# 分类层
self.classifier = nn.Sequential(
nn.Conv3d(512, 1, kernel_size=3, stride=1, padding=0), # 输出：(batch_size, 1, 1, 1, 1)
nn.Sigmoid()
)
def forward(self, x):
x = self.features(x)
x = self.classifier(x)
return x