def fCreateModel_FCN_simple(patchSize,
dr_rate=0.0,
iPReLU=0,
l1_reg=0.0,
l2_reg=1e-6):
# Total params: 1,223,831
# Replace the dense layer with a convolutional layer with filters=2 for the two classes
Strides = fgetStrides()
kernelnumber = fgetKernelNumber()
inp = Input(shape=(1, int(patchSize[0]), int(patchSize[1]),
int(patchSize[2])))
after_Conv_1 = fCreateVNet_Block(inp,
kernelnumber[0],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_1 = fCreateVNet_DownConv_Block(after_Conv_1,
after_Conv_1._keras_shape[1],
Strides[0],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Conv_2 = fCreateVNet_Block(after_DownConv_1,
kernelnumber[1],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_2 = fCreateVNet_DownConv_Block(after_Conv_2,
after_Conv_2._keras_shape[1],
Strides[1],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Conv_3 = fCreateVNet_Block(after_DownConv_2,
kernelnumber[2],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_3 = fCreateVNet_DownConv_Block(after_Conv_3,
after_Conv_3._keras_shape[1],
Strides[2],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
dropout_out = Dropout(dr_rate)(after_DownConv_3)
fclayer = Conv3D(
2,
kernel_size=(1, 1, 1),
kernel_initializer='he_normal',
weights=None,
padding='valid',
strides=(1, 1, 1),
kernel_regularizer=l1_l2(l1_reg, l2_reg),
)(dropout_out)
fclayer = GlobalAveragePooling3D()(fclayer)
outp = Activation('softmax')(fclayer)
cnn_spp = Model(inputs=inp, outputs=outp)
return cnn_spp
def fCreateModel_Inception_Archi1(patchSize,
dr_rate=0.0,
iPReLU=0,
l2_reg=1e-6):
# Total params: 5,598,511
# Each convolution layer before down sampling is replaced by the inception block,
# all other convolution layers are reserved.
Strides = fgetStrides()
kernelnumber = fgetKernelNumber()
inp = Input(shape=(1, int(patchSize[0]), int(patchSize[1]),
int(patchSize[2])))
after_Incep_1 = fConvIncep(inp,
KB=kernelnumber[0],
layernum=fgetLayerNumIncep(),
l2_reg=l2_reg)
after_DownConv_1 = fCreateVNet_DownConv_Block(
after_Incep_1,
after_Incep_1._keras_shape[1],
Strides[0],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Incep_2 = fConvIncep(after_DownConv_1,
KB=kernelnumber[1],
layernum=fgetLayerNumIncep(),
l2_reg=l2_reg)
after_DownConv_2 = fCreateVNet_DownConv_Block(
after_Incep_2,
after_Incep_2._keras_shape[1],
Strides[1],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Incep_3 = fConvIncep(after_DownConv_2,
KB=kernelnumber[2],
layernum=fgetLayerNumIncep(),
l2_reg=l2_reg)
after_DownConv_3 = fCreateVNet_DownConv_Block(
after_Incep_3,
after_Incep_3._keras_shape[1],
Strides[2],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
# fully connect layer as dense
flat_out = Flatten()(after_DownConv_3)
dropout_out = Dropout(dr_rate)(flat_out)
dense_out = Dense(units=2,
kernel_initializer='normal',
kernel_regularizer=l2(l2_reg))(dropout_out)
outp = Activation('softmax')(dense_out)
cnn_incep = Model(inputs=inp, outputs=outp)
return cnn_incep
def fpathway(input_t, dr_rate=0.0, iPReLU=0, l2_reg=1e-6):
Strides = fgetStrides()
KernelNumber = fgetKernelNumber()
after_res1_t = fCreateVNet_Block(input_t,
KernelNumber[0],
type=fgetLayerNumUnscaled(),
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_DownConv1_t = fCreateVNet_DownConv_Block(
after_res1_t,
after_res1_t._keras_shape[1],
Strides[0],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_res2_t = fCreateVNet_Block(after_DownConv1_t,
KernelNumber[1],
type=fgetLayerNumUnscaled(),
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_DownConv2_t = fCreateVNet_DownConv_Block(
after_res2_t,
after_res2_t._keras_shape[1],
Strides[1],
iPReLU=iPReLU,
l2_reg=l2_reg,
dr_rate=dr_rate)
after_res3_t = fCreateVNet_Block(after_DownConv2_t,
KernelNumber[2],
type=fgetLayerNumUnscaled(),
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_DownConv3_t = fCreateVNet_DownConv_Block(
after_res3_t,
after_res3_t._keras_shape[1],
Strides[2],
iPReLU=iPReLU,
l2_reg=l2_reg,
dr_rate=dr_rate)
return after_DownConv3_t
def fCreateModel_SPP(patchSize, dr_rate=0.0, iPReLU=0, l2_reg=1e-6):
# Total params: 1,036,856
# The third down sampling convolutional layer is replaced by the SPP module
Strides = fgetStrides()
kernelnumber = fgetKernelNumber()
inp = Input(shape=(1, int(patchSize[0]), int(patchSize[1]),
int(patchSize[2])))
after_Conv_1 = fCreateVNet_Block(inp,
kernelnumber[0],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_1 = fCreateVNet_DownConv_Block(after_Conv_1,
after_Conv_1._keras_shape[1],
Strides[0],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Conv_2 = fCreateVNet_Block(after_DownConv_1,
kernelnumber[1],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_2 = fCreateVNet_DownConv_Block(after_Conv_2,
after_Conv_2._keras_shape[1],
Strides[1],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Conv_3 = fCreateVNet_Block(after_DownConv_2,
kernelnumber[2],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_3 = fSPP(after_Conv_3, level=3)
dropout_out = Dropout(dr_rate)(after_DownConv_3)
dense_out = Dense(units=2,
kernel_initializer='normal',
kernel_regularizer=l2(l2_reg))(dropout_out)
outp = Activation('softmax')(dense_out)
cnn_spp = Model(inputs=inp, outputs=outp)
return cnn_spp
def fCreateModel_Inception_Archi2(patchSize,
dr_rate=0.0,
iPReLU=0,
l2_reg=1e-6):
# Total params: 2,883,791
# The three higher convolution layers except down sampling are replaced by the inception block,
# the three lower convolution layers are reserved.
# In work of GoogLeNet, it's beneficial for memory efficiency during training
# to start using inception modules at higher layers and to keep lower layers in traditional convolutional fashion.
Strides = fgetStrides()
kernelnumber = fgetKernelNumber()
inp = Input(shape=(1, int(patchSize[0]), int(patchSize[1]),
int(patchSize[2])))
after_Incep_1 = fCreateVNet_Block(inp,
kernelnumber[0],
type=fgetLayerNumIncep(),
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_DownConv_1 = fCreateVNet_DownConv_Block(
after_Incep_1,
after_Incep_1._keras_shape[1],
Strides[0],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Incep_2 = fConvIncep(after_DownConv_1,
KB=kernelnumber[1],
layernum=fgetLayerNumIncep(),
l2_reg=l2_reg)
after_DownConv_2 = fCreateVNet_DownConv_Block(
after_Incep_2,
after_Incep_2._keras_shape[1],
Strides[1],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Incep_3 = fIncepChain(after_DownConv_2,
layernum=fgetLayerNumIncep(),
l2_reg=l2_reg,
iPReLU=iPReLU)
after_DownConv_3 = fCreateVNet_DownConv_Block(
after_Incep_3,
after_Incep_3._keras_shape[1],
Strides[2],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
# fully connect layer as dense
flat_out = Flatten()(after_DownConv_3)
dropout_out = Dropout(dr_rate)(flat_out)
dense_out = Dense(units=2,
kernel_initializer='normal',
kernel_regularizer=l2(l2_reg))(dropout_out)
outp = Activation('softmax')(dense_out)
cnn_incep = Model(inputs=inp, outputs=outp)
return cnn_incep
def fCreateModel_FCN_MultiFM(patchSize,
dr_rate=0.0,
iPReLU=0,
l1_reg=0,
l2_reg=1e-6):
# Total params: 1,420,549
# The dense layer is repleced by a convolutional layer with filters=2 for the two classes
# The FM from the third down scaled convolutional layer is upsempled by deconvolution and
# added with the FM from the second down scaled convolutional layer.
# The combined FM goes through a convolutional layer with filters=2 for the two classes
# The two predictions are averages as the final result.
Strides = fgetStrides()
kernelnumber = fgetKernelNumber()
inp = Input(shape=(1, int(patchSize[0]), int(patchSize[1]),
int(patchSize[2])))
after_Conv_1 = fCreateVNet_Block(inp,
kernelnumber[0],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_1 = fCreateVNet_DownConv_Block(after_Conv_1,
after_Conv_1._keras_shape[1],
Strides[0],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Conv_2 = fCreateVNet_Block(after_DownConv_1,
kernelnumber[1],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_2 = fCreateVNet_DownConv_Block(after_Conv_2,
after_Conv_2._keras_shape[1],
Strides[1],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
after_Conv_3 = fCreateVNet_Block(after_DownConv_2,
kernelnumber[2],
type=fgetLayerNumConv(),
l2_reg=l2_reg)
after_DownConv_3 = fCreateVNet_DownConv_Block(after_Conv_3,
after_Conv_3._keras_shape[1],
Strides[2],
iPReLU=iPReLU,
dr_rate=dr_rate,
l2_reg=l2_reg)
# fully convolution over the FM from the deepest level
dropout_out1 = Dropout(dr_rate)(after_DownConv_3)
fclayer1 = Conv3D(
2,
kernel_size=(1, 1, 1),
kernel_initializer='he_normal',
weights=None,
padding='valid',
strides=(1, 1, 1),
kernel_regularizer=l1_l2(l1_reg, l2_reg),
)(dropout_out1)
fclayer1 = GlobalAveragePooling3D()(fclayer1)
# Upsample FM from the deepest level, add with FM from level 2,
UpedFM_Level3 = Conv3DTranspose(filters=97,
kernel_size=(3, 3, 1),
strides=(2, 2, 1),
padding='same')(after_DownConv_3)
conbined_FM_Level23 = add([UpedFM_Level3, after_DownConv_2])
fclayer2 = Conv3D(
2,
kernel_size=(1, 1, 1),
kernel_initializer='he_normal',
weights=None,
padding='valid',
strides=(1, 1, 1),
kernel_regularizer=l1_l2(l1_reg, l2_reg),
)(conbined_FM_Level23)
fclayer2 = GlobalAveragePooling3D()(fclayer2)
# combine the two predictions using average
fcl_aver = average([fclayer1, fclayer2])
predict = Activation('softmax')(fcl_aver)
cnn_fcl_msfm = Model(inputs=inp, outputs=predict)
return cnn_fcl_msfm