def model3D(img=(84, 256, 256)):
with tf.name_scope('WMH'):
seq = tg.Sequential()
convStride = (1, 1, 1)
poolStride = (2, 2, 2)
seq.add(
Conv3D(input_channels=1,
num_filters=10,
kernel_size=(3, 3, 3),
stride=convStride,
padding='SAME'))
seq.add(TFBatchNormalization(name='b1'))
seq.add(
MaxPool3D(poolsize=(2, 2, 2), stride=poolStride, padding='SAME'))
layerSize1 = updateConvLayerSize(img, poolStride)
#print("layer1: "+str(layerSize1))
seq.add(RELU())
seq.add(
Conv3D(input_channels=10,
num_filters=20,
kernel_size=(3, 3, 3),
stride=convStride,
padding='SAME'))
seq.add(TFBatchNormalization(name='b2'))
#layerSize2 = updateConvLayerSize(layerSize1,convStride)
#print("layer1: "+str(layerSize2))
seq.add(RELU())
seq.add(
Conv3D_Tranpose1(input_channels=20,
num_filters=10,
output_shape=layerSize1,
kernel_size=(3, 3, 3),
stride=convStride,
padding='SAME'))
seq.add(RELU())
seq.add(
Conv3D_Tranpose1(input_channels=10,
num_filters=3,
output_shape=img,
kernel_size=(3, 3, 3),
stride=(2, 2, 2),
padding='SAME'))
##
seq.add(RELU())
seq.add(
Conv3D(input_channels=3,
num_filters=2,
kernel_size=(3, 3, 3),
stride=convStride,
padding='SAME'))
##
seq.add(Softmax())
#seq.add(Sigmoid())
return seq
def model3D_2(img=(84, 256, 256)):
with tf.name_scope('WMH_2Chan_Input'):
seq = tg.Sequential()
convStride = (1, 1, 1)
poolStride = (2, 2, 2)
seq.add(
Conv3D(input_channels=1,
num_filters=8,
kernel_size=(5, 5, 5),
stride=convStride,
padding='SAME'))
seq.add(TFBatchNormalization(name='b1'))
seq.add(
MaxPool3D(poolsize=(2, 2, 2), stride=poolStride, padding='SAME'))
layerSize1 = updateConvLayerSize(img, poolStride)
seq.add(RELU())
seq.add(
Conv3D(input_channels=8,
num_filters=16,
kernel_size=(3, 3, 3),
stride=convStride,
padding='SAME'))
seq.add(TFBatchNormalization(name='b2'))
## Extra MaxPool
#seq.add(MaxPool3D(poolsize=(2,2,2), stride=poolStride, padding='SAME'))
#layerSize2 = updateConvLayerSize(layerSize1,convStride)
#seq.add(RELU())
## Extra Conv
#seq.add(Conv3D(input_channels=16, num_filters=16, kernel_size=(3,3,3), stride=convStride, padding='SAME'))
#seq.add(TFBatchNormalization(name='b3'))
seq.add(
Conv3D_Tranpose1(input_channels=16,
num_filters=8,
output_shape=layerSize1,
kernel_size=(3, 3, 3),
stride=convStride,
padding='SAME'))
seq.add(TFBatchNormalization(name='b4'))
seq.add(RELU())
seq.add(
Conv3D_Tranpose1(input_channels=8,
num_filters=2,
output_shape=img,
kernel_size=(3, 3, 3),
stride=poolStride,
padding='SAME'))
#seq.add(TFBatchNormalization(name='b5'))
#seq.add(RELU())
#seq.add(Conv3D(input_channels=2, num_filters=2, kernel_size=(1,1,1), stride=convStride, padding='SAME'))
##
##
#seq.add(RELU())
#seq.add(Conv3D(input_channels=2, num_filters=2, kernel_size=(3,3,3), stride=convStride, padding='SAME'))
# seq.add(RELU())
# seq.add(Conv3D(input_channels=3, num_filters=3, kernel_size=(3,3,3), stride=convStride, padding='SAME'))
# seq.add(RELU())
# seq.add(Conv3D(input_channels=3, num_filters=3, kernel_size=(3,3,3), stride=convStride, padding='SAME'))
#
##
seq.add(Softmax())
#seq.add(Sigmoid())
return seq
def Residual_UNET(input, img=(84, 256, 256)):
with tf.name_scope('WMH'):
convStride = (1, 1, 1)
poolStride = (2, 2, 2)
kSize3 = (3, 3, 3)
#kSize5 = (5,5,5)
#x_dim = 50
#component_dim = 100
#batchsize = 32
#learning_rate = 0.01
#x_ph = tf.placeholder('float32', [None, x_dim])
#start = StartNode(input_vars=[x_ph])
#h1 = HiddenNode(prev=[start], layers=[Linear(x_dim, component_dim), Softmax()])
#e1 = EndNode(prev=[h1], input_merge_mode=Sum())
#e3 = EndNode(prev=[h1, h2, h3], input_merge_mode=Sum())
start = StartNode(input_vars=[input])
Layer01 = [
Conv3D(input_channels=1,
num_filters=8,
kernel_size=kSize3,
stride=convStride,
padding='SAME')
]
#Layer01.append(RELU())
LayerPool = MaxPool3D(poolsize=(2, 2, 2),
stride=poolStride,
padding='SAME')
Layer02 = [LayerPool]
#Layer02.append(Testing(1))
Layer02.append(
Conv3D(input_channels=8,
num_filters=16,
kernel_size=kSize3,
stride=convStride,
padding='SAME'))
#Layer02.append(RELU())
#Layer02.append(Testing(2))
Layer02.append(ResidualBlock3D(16, 'L02'))
#Layer02.append(Testing(3))
layerSize1 = updateConvLayerSize(img, poolStride)
Layer03 = [LayerPool]
Layer03.append(
Conv3D(input_channels=16,
num_filters=32,
kernel_size=kSize3,
stride=convStride,
padding='SAME'))
#Layer03.append(RELU())
Layer03.append(ResidualBlock3D(32, 'L03'))
#layerSize2 = updateConvLayerSize(layerSize1,poolStride)
Layer03.append(
Conv3D_Tranpose1(input_channels=32,
num_filters=16,
output_shape=layerSize1,
kernel_size=kSize3,
stride=poolStride,
padding='SAME'))
#Layer03.append(RELU())
#Layer04 = [Conv3D(input_channels=32, num_filters=64, kernel_size=kSize5, stride=convStride, padding='SAME')]
#Layer04.append(ResidualBlock3D(64,'L03'))
conv8 = HiddenNode(prev=[start], layers=Layer01)
resBlock16 = HiddenNode(prev=[conv8], layers=Layer02)
resBlock32_16 = HiddenNode(prev=[resBlock16], layers=Layer03)
residualLong16 = HiddenNode(prev=[resBlock32_16, resBlock16],
input_merge_mode=Sum())
Layer04 = [ResidualBlock3D(16, 'L04')]
Layer04.append(
Conv3D_Tranpose1(input_channels=16,
num_filters=8,
output_shape=img,
kernel_size=kSize3,
stride=poolStride,
padding='SAME'))
Layer04.append(RELU())
resBlock16_8 = HiddenNode(prev=[residualLong16], layers=Layer04)
residualLong8 = HiddenNode(prev=[resBlock16_8, conv8],
input_merge_mode=Sum())
Layer05 = [ResidualBlock3D(8, 'L05')]
Layer05.append(
Conv3D(input_channels=8,
num_filters=2,
kernel_size=kSize3,
stride=convStride,
padding='SAME'))
Layer05.append(Softmax())
resBlock8_2 = HiddenNode(prev=[residualLong8], layers=Layer05)
endNode = EndNode(prev=[resBlock8_2], input_merge_mode=NoChange())
graph = Graph(start=[start], end=[endNode])
#o1, o2, o3 = graph.train_fprop()
#o1_mse = tf.reduce_mean((y1_ph - o1)**2)
#o2_mse = tf.reduce_mean((y2_ph - o2)**2)
return graph
def model_Inception_Resnet(img=(84, 256, 256)):
with tf.name_scope('WMH'):
seq = tg.Sequential()
convStride = (1, 1, 1)
poolStride = (2, 2, 2)
kernelSize = (3, 3, 3)
seq.add(
Conv3D(input_channels=1,
num_filters=8,
kernel_size=(5, 5, 5),
stride=convStride,
padding='SAME'))
#seq.add(TFBatchNormalization(name='b1'))
seq.add(
MaxPool3D(poolsize=(2, 2, 2), stride=poolStride, padding='SAME'))
layerSize1 = updateConvLayerSize(img, poolStride)
seq.add(RELU())
seq.add(InceptionResnet_3D(8, type='v2_out8'))
seq.add(
Conv3D(input_channels=8,
num_filters=16,
kernel_size=kernelSize,
stride=convStride,
padding='SAME'))
#seq.add(TFBatchNormalization(name='b2'))
seq.add(
MaxPool3D(poolsize=(2, 2, 2), stride=poolStride, padding='SAME'))
layerSize2 = updateConvLayerSize(layerSize1, poolStride)
seq.add(RELU())
seq.add(InceptionResnet_3D(16, type='v1_out16'))
seq.add(
Conv3D(input_channels=16,
num_filters=16,
kernel_size=kernelSize,
stride=convStride,
padding='SAME'))
seq.add(RELU())
seq.add(
Conv3D(input_channels=16,
num_filters=32,
kernel_size=kernelSize,
stride=convStride,
padding='SAME'))
#seq.add(TFBatchNormalization(name='b3'))
seq.add(
MaxPool3D(poolsize=(2, 2, 2), stride=poolStride, padding='SAME'))
#layerSize3 = updateConvLayerSize(layerSize2,poolStride)
seq.add(RELU())
seq.add(InceptionResnet_3D(32, type='v1_out16'))
seq.add(
Conv3D_Tranpose1(input_channels=16,
num_filters=16,
output_shape=layerSize2,
kernel_size=kernelSize,
stride=poolStride,
padding='SAME'))
seq.add(RELU())
seq.add(InceptionResnet_3D(16, type='v1_out16'))
seq.add(
Conv3D(input_channels=16,
num_filters=16,
kernel_size=kernelSize,
stride=convStride,
padding='SAME'))
seq.add(RELU())
seq.add(
Conv3D_Tranpose1(input_channels=16,
num_filters=8,
output_shape=layerSize1,
kernel_size=kernelSize,
stride=poolStride,
padding='SAME'))
seq.add(RELU())
seq.add(InceptionResnet_3D(8, type='v2_out8'))
seq.add(
Conv3D(input_channels=8,
num_filters=8,
kernel_size=kernelSize,
stride=convStride,
padding='SAME'))
seq.add(RELU())
# num_filter=3 --> Background, WhiteMatter, Others
seq.add(
Conv3D_Tranpose1(input_channels=8,
num_filters=3,
output_shape=img,
kernel_size=kernelSize,
stride=poolStride,
padding='SAME'))
##
seq.add(RELU())
seq.add(
Conv3D(input_channels=3,
num_filters=3,
kernel_size=(1, 1, 1),
stride=convStride,
padding='SAME'))
##
seq.add(Softmax())
return seq