-
Notifications
You must be signed in to change notification settings - Fork 1
/
example-3d.py
executable file
·128 lines (104 loc) · 6.26 KB
/
example-3d.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
import numpy as np
import tensorflow as tf
from smoothTransformer import smoothTransformer3D
import matplotlib.pylab as plt
def getModel(moving, reference):
i = tf.keras.layers.Concatenate(axis=-1)([moving, reference])
# encoder
enc1 = tf.keras.layers.Conv3D(32, (3, 3, 3), dilation_rate=(1, 1, 1), kernel_initializer='he_normal', padding='same')(i)
enc1 = tf.keras.layers.LeakyReLU()(enc1)
enc2 = tf.keras.layers.Conv3D(32, (3, 3, 3), dilation_rate=(1, 1, 1), kernel_initializer='he_normal', padding='same')(enc1)
enc2 = tf.keras.layers.LeakyReLU()(enc2)
enc3 = tf.keras.layers.Conv3D(32, (3, 3, 3), dilation_rate=(2, 2, 2), kernel_initializer='he_normal', padding='same')(enc2)
enc3 = tf.keras.layers.LeakyReLU()(enc3)
enc = tf.keras.layers.Concatenate(axis=-1)([i,enc1,enc2,enc3])
enc = tf.keras.layers.Dropout(0.5)(enc)
# decoder
ddec = tf.keras.layers.Conv3D(32, (3, 3, 3), dilation_rate=(2, 2, 2), kernel_initializer='he_normal', padding='same')(enc)
ddec = tf.keras.layers.LeakyReLU()(ddec)
ddec = tf.keras.layers.Conv3D(32, (3, 3, 3), dilation_rate=(1, 1, 1), kernel_initializer='he_normal', padding='same')(ddec)
ddec = tf.keras.layers.LeakyReLU()(ddec)
ddec = tf.keras.layers.Conv3D(32, (3, 3, 3), dilation_rate=(1, 1, 1), kernel_initializer='he_normal', padding='same')(ddec)
ddec = tf.keras.layers.LeakyReLU()(ddec)
# Affine
affine = tf.keras.layers.GlobalAveragePooling3D()(enc)
affine = tf.keras.layers.Dense(12, bias_initializer='zeros', kernel_initializer='zeros')(affine)
affine = tf.keras.layers.Activation('linear')(affine) # this should be linear and initialized to 0 !
affine = tf.keras.layers.ActivityRegularization(l1=1e-7)(affine) # affine regularizer
# Deformable
deformable = tf.keras.layers.Conv3D(3, (3, 3, 3), padding='same', bias_initializer='zeros', kernel_initializer='zeros')(ddec)
deformable = tf.keras.layers.Activation('linear')(deformable) # this should be linear and initialized to 0 !
deformable = tf.keras.layers.ActivityRegularization(l1=1e-7)(deformable) # deformable regularizer
[forward, inverse, sampling_grid, sampling_grid_inverse] = smoothTransformer3D(maxgrad=10)([moving, reference, deformable, affine])
return forward, inverse, sampling_grid, sampling_grid_inverse
def getMnist3D(ntrain=2500, ntest=500, random_yawn=False):
def normalization(x):
return 255* ((x - np.min(x)) / (np.max(x) - np.min(x) + 1e-7))
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
x_train_3d = np.pad(np.repeat(x_train[:ntrain,...,None],6,axis=-1), ((0,0),(0,0),(0,0),(11,11)))
x_test_3d = np.pad(np.repeat(x_test[:ntest,...,None],6,axis=-1), ((0,0),(0,0),(0,0),(11,11)))
if random_yawn:
from scipy.ndimage import affine_transform
for i in range(x_train_3d.shape[0]):
theta = 4*np.random.random()-2
tra = [[1,0,0,-14],[0,1,0,-14],[0,0,1,-14],[0,0,0,1]]
rot = [[np.cos(theta), 0, np.sin(theta),0],[0,1,0,0],[-np.sin(theta), 0, np.cos(theta),0],[0,0,0,1]]
trainv = [[1,0,0,14],[0,1,0,14],[0,0,1,14],[0,0,0,1]]
affine = affine = np.matmul(trainv,np.matmul(rot, tra))
x_train_3d[i,...] = normalization(affine_transform(x_train_3d[i,...], affine))
for i in range(x_test_3d.shape[0]):
theta = 4*np.random.random()-2
tra = [[1,0,0,-14],[0,1,0,-14],[0,0,1,-14],[0,0,0,1]]
rot = [[np.cos(theta), 0, np.sin(theta),0],[0,1,0,0],[-np.sin(theta), 0, np.cos(theta),0],[0,0,0,1]]
trainv = [[1,0,0,14],[0,1,0,14],[0,0,1,14],[0,0,0,1]]
affine = affine = np.matmul(trainv,np.matmul(rot, tra))
x_test_3d[i,...] = normalization(affine_transform(x_test_3d[i,...], affine))
return (x_train_3d, y_train[:ntrain]), (x_test_3d, y_test[:ntest])
if __name__ == '__main__':
# input
moving = tf.keras.layers.Input((28,28,28,1))
reference = tf.keras.layers.Input((28,28,28,1))
forward, inverse, sampling_grid, sampling_grid_inverse = getModel(moving, reference)
model = tf.keras.models.Model(inputs=[moving, reference], outputs=[forward,inverse])
model.compile(optimizer='adam', loss='mse')
# generate a simple 3D mnist dataset
(x_train, _), (x_test, _) = getMnist3D(random_yawn=False)
y_train = x_train[np.random.permutation(x_train.shape[0]),...]
y_test = x_test[np.random.permutation(x_test.shape[0]),...]
# map input data to [0,1]
x_train = x_train[...,None]/255.
y_train = y_train[...,None]/255.
x_test = x_test[...,None]/255.
y_test = y_test[...,None]/255.
# train model
model.fit(x=[x_train, y_train], y=[y_train, x_train], batch_size=64, epochs=20, verbose=1, validation_split=0.5)
# get inference model
inference_model = tf.keras.models.Model(inputs=model.input, outputs=[forward, inverse, sampling_grid, sampling_grid_inverse])
i = np.random.randint(x_test.shape[0])
x = x_test[i:i+1,...]
y = y_test[i:i+1,...]
fwd, bwd, grid, grid_inv = inference_model.predict([x,y])
# plot example
# With no training the network should apply the identity transform to the moving image
xx, yy, zz = np.meshgrid(range(x.shape[1]), range(x.shape[2]), range(x.shape[3]), indexing='ij')
dx, dy, dz = grid[0,:,:,:,0] + xx, grid[0,:,:,:,1] + yy, grid[0,:,:,:,1] + zz
dxi, dyi, dzi = grid_inv[0,:,:,:,0] + xx, grid_inv[0,:,:,:,1] + yy, grid_inv[0,:,:,:,1] + zz
plt.figure(figsize=(10,4))
plt.subplot(1, 4, 1)
plt.imshow(np.squeeze(x)[:,:,14],cmap='gray')
plt.title('Moving Image')
plt.subplot(1, 4, 2)
plt.imshow(np.squeeze(fwd)[:,:,14],cmap='gray')
plt.contour(dx[...,14], 50, alpha=0.5, linewidths=0.5)
plt.contour(dy[...,14], 50, alpha=0.5, linewidths=0.5)
plt.title('Forward Deformation \n applied on Moving Image')
plt.subplot(1, 4, 3)
plt.imshow(np.squeeze(bwd)[:,:,14],cmap='gray')
plt.contour(dxi[...,14], 50, alpha=0.5, linewidths=0.5)
plt.contour(dyi[...,14], 50, alpha=0.5, linewidths=0.5)
plt.title('Inverse Deformation \n applied on Reference Image')
plt.subplot(1, 4, 4)
plt.imshow(np.squeeze(y)[:,:,14],cmap='gray')
plt.title('Reference Image')
plt.tight_layout()
plt.savefig('example-3d-output.png')