def __init__(self,
num_channels,
num_classes,
num_conv_layers=2,
kernel_size=(3, 3, 3),
nonlinearity='relu',
use_batchnorm=True,
dropout_rate=0.25,
use_spatial_dropout=True,
data_format='channels_last',
name="vnet"):
super(VNet, self).__init__(name=name)
self.conv_1 = Conv3D_Block(num_channels,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.conv_2 = Conv3D_Block(num_channels*2,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.conv_3 = Conv3D_Block(num_channels*4,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.conv_4 = Conv3D_Block(num_channels*8,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,use_dropout=True,dropout_rate=dropout_rate,use_spatial_dropout=use_spatial_dropout,data_format=data_format)
self.conv_5 = Conv3D_Block(num_channels*16,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,use_dropout=True,dropout_rate=dropout_rate,use_spatial_dropout=use_spatial_dropout,data_format=data_format)
self.up_5 = Up_Conv3D(num_channels*8,(2,2,2),nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_6 = Up_Conv3D(num_channels*4,(2,2,2),nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_7 = Up_Conv3D(num_channels*2,(2,2,2),nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_8 = Up_Conv3D(num_channels,(2,2,2),nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_conv4 = Conv3D_Block(num_channels*8,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_conv3 = Conv3D_Block(num_channels*4,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_conv2 = Conv3D_Block(num_channels*2,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
self.up_conv1 = Conv3D_Block(num_channels,num_conv_layers,kernel_size,nonlinearity,use_batchnorm=use_batchnorm,data_format=data_format)
# convolution num_channels at the output
self.conv_output = tf.keras.layers.Conv3D(2, kernel_size, activation=nonlinearity, padding='same', data_format=data_format)
self.conv_1x1 = tf.keras.layers.Conv3D(num_classes, kernel_size, padding='same', data_format=data_format)
def __init__(self,
num_channels,
num_classes,
num_conv_layers=2,
kernel_size=(3, 3, 3),
nonlinearity='relu',
use_batchnorm=True,
dropout_rate=0.25,
use_spatial_dropout=True,
data_format='channels_last',
merge_connections=False,
output_activation=None,
name="vnet_tiny"):
super(VNet_Tiny, self).__init__(name=name)
self.merge_connections = merge_connections
self.num_classes = num_classes
self.conv_1 = Conv3D_Block(num_channels, num_conv_layers, kernel_size,
nonlinearity, use_batchnorm=use_batchnorm,
data_format=data_format)
self.conv_2 = Conv3D_Block(num_channels * 2, num_conv_layers, kernel_size,
nonlinearity, use_batchnorm=use_batchnorm,
data_format=data_format)
self.up_2 = Up_Conv3D(num_channels, (2, 2, 2), nonlinearity,
use_batchnorm=use_batchnorm, data_format=data_format)
self.up_conv1 = Conv3D_Block(num_channels, num_conv_layers, kernel_size, nonlinearity,
use_batchnorm=use_batchnorm, data_format=data_format)
# convolution num_channels at the output
self.conv_output = tf.keras.layers.Conv3D(num_classes, kernel_size, activation=nonlinearity, padding='same',
data_format=data_format)
if output_activation is None:
output_activation = 'sigmoid'
if num_classes > 1:
output_activation = 'softmax'
if num_classes == 1:
self.conv_1x1_binary = tf.keras.layers.Conv3D(num_classes, (1, 1, 1), activation='sigmoid',
padding='same', data_format=data_format)
else:
self.conv_1x1 = tf.keras.layers.Conv3D(num_classes, kernel_size=(1, 1, 1), activation='softmax',
padding='same', data_format=data_format)
def __init__(self,
num_channels,
num_classes,
num_conv_layers=2,
kernel_size=(3, 3, 3),
nonlinearity='relu',
use_batchnorm=True,
dropout_rate=0.25,
use_spatial_dropout=True,
data_format='channels_last',
merge_connections=False,
num_compressors=3,
compressor_filters=3,
action='add',
output_activation=None,
name="vnet_small_relative"):
super(VNet_Small_Relative, self).__init__(name=name)
self.merge_connections = merge_connections
self.num_classes = num_classes
self.action = action
assert self.action in ['add',
'multiply'], f"{self.action} not in action list"
self.conv_1 = Conv3D_Block(num_channels,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c1")
self.conv_2 = Conv3D_Block(num_channels * 2,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c2")
self.conv_3 = Conv3D_Block(num_channels * 4,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c3")
self.up_3 = Up_Conv3D(num_channels * 2, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu3")
self.up_2 = Up_Conv3D(num_channels, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu2")
self.up_conv2 = Conv3D_Block(num_channels * 2,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc2")
self.up_conv1 = Conv3D_Block(num_channels,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc1")
# compression section
self.compressors = []
for idx, i in enumerate(range(num_compressors)):
cfilters = compressor_filters
if num_compressors == (idx + 1):
cfilters = 1
self.compressors.append(
tf.keras.layers.Conv3D(cfilters,
kernel_size=3,
strides=2,
padding="valid",
activation=nonlinearity,
data_format=data_format,
name=f"compessor_{idx}"))
self.comp_dense_1 = tf.keras.layers.Dense(16, activation="relu")
self.comp_dense_2 = tf.keras.layers.Dense(4, activation="relu")
self.comp_dense_3 = tf.keras.layers.Dense(1, activation="tanh")
# convolution num_channels at the output
self.conv_output = tf.keras.layers.Conv3D(num_classes,
kernel_size,
activation=nonlinearity,
padding='same',
data_format=data_format)
if output_activation is None:
output_activation = 'sigmoid'
if num_classes > 1:
output_activation = 'softmax'
if num_classes == 1:
self.conv_1x1_binary = tf.keras.layers.Conv3D(
num_classes, (1, 1, 1),
activation='sigmoid',
padding='same',
data_format=data_format)
else:
self.conv_1x1 = tf.keras.layers.Conv3D(num_classes,
kernel_size=(1, 1, 1),
activation='softmax',
padding='same',
data_format=data_format)
def __init__(self,
num_channels,
num_classes,
num_conv_layers=2,
kernel_size=(3, 3, 3),
nonlinearity='relu',
use_batchnorm=True,
dropout_rate=0.25,
use_spatial_dropout=True,
data_format='channels_last',
merge_connections=True,
output_activation=None,
noise=0.0001,
use_res_connect=False,
use_stride_2=False,
name="vnet_small"):
self.params = str(inspect.currentframe().f_locals)
super(VNet_Small, self).__init__(name=name)
self.merge_connections = merge_connections
self.num_classes = num_classes
self.noise = noise
self.use_res_connect = use_res_connect
self.use_stride_2 = use_stride_2
self.conv_1 = Conv3D_Block(num_channels,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c1")
if self.use_stride_2:
self.conv_1_stride = tf.keras.layers.Conv3D(
num_channels,
kernel_size=kernel_size,
strides=2,
activation="selu",
padding="same")
self.conv_2 = Conv3D_Block(num_channels * 2,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c2")
if self.use_stride_2:
self.conv_2_stride = tf.keras.layers.Conv3D(
num_channels,
kernel_size=kernel_size,
strides=2,
activation="selu",
padding="same")
self.conv_3 = Conv3D_Block(num_channels * 4,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c3")
self.up_3 = Up_Conv3D(num_channels * 2, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu3")
self.up_2 = Up_Conv3D(num_channels, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu2")
self.up_conv2 = Conv3D_Block(num_channels * 2,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc2")
self.up_conv1 = Conv3D_Block(num_channels,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc1")
# convolution num_channels at the output
self.conv_output = tf.keras.layers.Conv3D(num_classes,
kernel_size,
activation=nonlinearity,
padding='same',
data_format=data_format)
if output_activation is None:
output_activation = 'sigmoid'
if num_classes > 1:
output_activation = 'softmax'
if num_classes == 1:
self.conv_1x1_binary = tf.keras.layers.Conv3D(
num_classes, (1, 1, 1),
activation='sigmoid',
padding='same',
data_format=data_format)
else:
self.conv_1x1 = tf.keras.layers.Conv3D(num_classes,
kernel_size=(1, 1, 1),
activation='softmax',
padding='same',
data_format=data_format)
def __init__(self,
num_channels,
num_classes,
num_conv_layers=3,
start_kernel_size=(5, 5, 5),
kernel_size=(3, 3, 3),
nonlinearity='relu',
use_batchnorm=True,
dropout_rate=0.25,
use_spatial_dropout=True,
data_format='channels_last',
merge_connections=True,
action='add',
output_activation=None,
use_res_connect=False,
use_stride_2=False,
name="vnet_large_relative"):
self.params = str(inspect.currentframe().f_locals)
super(VNet_Large_Relative, self).__init__(name=name)
self.merge_connections = merge_connections
self.num_classes = num_classes
self.action = action
assert self.action in ['add',
'multiply'], f"{self.action} not in action list"
self.use_res_connect = use_res_connect
self.use_stride_2 = use_stride_2
self.conv_1 = Conv3D_Block(num_channels,
num_conv_layers,
start_kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c1")
if self.use_stride_2:
self.conv_1_stride = tf.keras.layers.Conv3D(
num_channels,
kernel_size=kernel_size,
strides=2,
activation="selu",
padding="same")
self.conv_2 = Conv3D_Block(num_channels * 2,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c2")
if self.use_stride_2:
self.conv_2_stride = tf.keras.layers.Conv3D(
num_channels,
kernel_size=kernel_size,
strides=2,
activation="selu",
padding="same")
self.conv_3 = Conv3D_Block(num_channels * 4,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c3")
if self.use_stride_2:
self.conv_3_stride = tf.keras.layers.Conv3D(
num_channels,
kernel_size=kernel_size,
strides=2,
activation="selu",
padding="same")
self.conv_4 = Conv3D_Block(num_channels * 8,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c4")
if self.use_stride_2:
self.conv_4_stride = tf.keras.layers.Conv3D(
num_channels,
kernel_size=kernel_size,
strides=2,
activation="selu",
padding="same")
self.conv_5 = Conv3D_Block(num_channels * 16,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="c5")
self.up_5 = Up_Conv3D(num_channels * 8, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu5")
self.up_4 = Up_Conv3D(num_channels * 4, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu4")
self.up_3 = Up_Conv3D(num_channels * 2, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu3")
self.up_2 = Up_Conv3D(num_channels, (2, 2, 2),
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="cu2")
self.up_conv4 = Conv3D_Block(num_channels * 8,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc4")
self.up_conv3 = Conv3D_Block(num_channels * 4,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc3")
self.up_conv2 = Conv3D_Block(num_channels * 2,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc2")
self.up_conv1 = Conv3D_Block(num_channels,
num_conv_layers,
kernel_size,
nonlinearity,
use_batchnorm=use_batchnorm,
data_format=data_format,
name="upc1")
self.pos_dense_1 = tf.keras.layers.Dense(16, activation="relu")
self.pos_dense_2 = tf.keras.layers.Dense(4, activation="relu")
self.pos_dense_3 = tf.keras.layers.Dense(1, activation="tanh")
# convolution num_channels at the output
self.conv_output = tf.keras.layers.Conv3D(num_classes,
kernel_size,
activation=nonlinearity,
padding='same',
data_format=data_format)
if output_activation is None:
output_activation = 'sigmoid'
if num_classes > 1:
output_activation = 'softmax'
if num_classes == 1:
self.conv_1x1_binary = tf.keras.layers.Conv3D(
num_classes, (1, 1, 1),
activation='sigmoid',
padding='same',
data_format=data_format)
else:
self.conv_1x1 = tf.keras.layers.Conv3D(num_classes,
kernel_size=(1, 1, 1),
activation='softmax',
padding='same',
data_format=data_format)