def build_model(input_shape):
xin = Input(input_shape)
#shift the below down by one
x1 = conv_block(xin,8,activation='relu')
x1_ident = AveragePooling3D()(xin)
x1_merged = merge([x1, x1_ident],mode='concat', concat_axis=1)
x2_1 = conv_block(x1_merged,24,activation='relu') #outputs 37 ch
x2_ident = AveragePooling3D()(x1_ident)
x2_merged = merge([x2_1,x2_ident],mode='concat', concat_axis=1)
#by branching we reduce the #params
x3_ident = AveragePooling3D()(x2_ident)
x3_malig = conv_block(x2_merged,48,activation='relu') #outputs 25 + 16 ch = 41
x3_malig_merged = merge([x3_malig,x3_ident],mode='concat', concat_axis=1)
x4_ident = AveragePooling3D()(x3_ident)
x4_malig = conv_block(x3_malig_merged,64,activation='relu') #outputs 25 + 16 ch = 41
x4_merged = merge([x4_malig,x4_ident],mode='concat', concat_axis=1)
x5_malig = conv_block(x4_merged,64) #outputs 25 + 16 ch = 41
xpool_malig = BatchNormalization(momentum=0.995)(GlobalMaxPooling3D()(x5_malig))
xout_malig = Dense(1, name='o_mal', activation='relu')(xpool_malig) #relu output
x5_diam = conv_block(x4_merged,64) #outputs 25 + 16 ch = 41
xpool_diam = BatchNormalization(momentum=0.995)(GlobalMaxPooling3D()(x5_diam))
xout_diam = Dense(1, name='o_diam', activation='relu')(xpool_diam) #relu output
x5_lob = conv_block(x4_merged,64) #outputs 25 + 16 ch = 41
xpool_lob = BatchNormalization(momentum=0.995)(GlobalMaxPooling3D()(x5_lob))
xout_lob = Dense(1, name='o_lob', activation='relu')(xpool_lob) #relu output
x5_spic = conv_block(x4_merged,64) #outputs 25 + 16 ch = 41
xpool_spic = BatchNormalization(momentum=0.995)(GlobalMaxPooling3D()(x5_spic))
xout_spic = Dense(1, name='o_spic', activation='relu')(xpool_spic) #relu output
model = Model(input=xin,output=[xout_diam, xout_lob, xout_spic, xout_malig])
if input_shape[1] == 32:
lr_start = .01
elif input_shape[1] == 64:
lr_start = .003
elif input_shape[1] == 128:
lr_start = .002
# elif input_shape[1] == 96:
# lr_start = 5e-4
opt = Nadam(lr_start,clipvalue=1.0)
print 'compiling model'
model.compile(optimizer=opt,loss='mse',loss_weights={'o_diam':0.06, 'o_lob':0.5, 'o_spic':0.5, 'o_mal':1.0})
return model
def up_conv_block_seunet(x, x2, f, dropout=False):
x = UpSampling3D(size=(2, 2, 2))(x)
channels_nb = K.int_shape(x2)[-1]
if channels_nb==16:
channels_nb_bottleneck = channels_nb // 16
else:
channels_nb_bottleneck = channels_nb // 32
x3=GlobalMaxPooling3D()(x2)
x3 = Dense(channels_nb_bottleneck, activation='relu')(x3)
x3 = Dense(channels_nb, activation='sigmoid')(x3)
y = Lambda(lambda x: attetion(x))([x2, x3])
x = Concatenate(axis=-1)([x, y])
f_new = f + channels_nb
x = Conv3D(f_new, (3, 3, 3), padding="same")(x)
x = Conv3D(f_new, (3, 3, 3), padding="same")(x)
x = BatchNormalization(axis=-1)(x)
if dropout:
x = Dropout(0.5)(x)
x = Activation("relu")(x)
return x
def build_model(input_shape):
xin = Input(input_shape)
#shift the below down by one
x1 = conv_block(xin, 8, activation='relu')
x1_ident = AveragePooling3D()(xin)
x1_merged = merge([x1, x1_ident], mode='concat', concat_axis=1)
x2_1 = conv_block(x1_merged, 24, activation='relu') #outputs 37 ch
x2_ident = AveragePooling3D()(x1_ident)
x2_merged = merge([x2_1, x2_ident], mode='concat', concat_axis=1)
#by branching we reduce the #params
x3_ident = AveragePooling3D()(x2_ident)
x3_malig = conv_block(x2_merged, 36,
activation='relu') #outputs 25 + 16 ch = 41
x3_malig_merged = merge([x3_malig, x3_ident], mode='concat', concat_axis=1)
x4_ident = AveragePooling3D()(x3_ident)
x4_malig = conv_block(x3_malig_merged, 48,
activation='relu') #outputs 25 + 16 ch = 41
x4_malig_merged = merge([x4_malig, x4_ident], mode='concat', concat_axis=1)
x5_malig = conv_block(x4_malig_merged, 64) #outputs 25 + 16 ch = 41
xpool_malig = BatchNormalization(momentum=0.995)(
GlobalMaxPooling3D()(x5_malig))
xout_malig = Dense(1, name='o_mal',
activation='sigmoid')(xpool_malig) #sigmoid output
model = Model(input=xin, output=xout_malig)
if input_shape[1] == 32:
lr_start = .01
elif input_shape[1] == 64:
lr_start = .003
elif input_shape[1] == 128:
lr_start = .002
# elif input_shape[1] == 96:
# lr_start = 5e-4
opt = Nadam(lr_start, clipvalue=1.0)
print 'compiling model'
model.compile(optimizer=opt, loss='mae')
return model
def build_model(input_shape):
xin = Input(input_shape)
x1 = conv_block(xin,8,activation='crelu')
x1_ident = AveragePooling3D()(xin)
x1_merged = concatenate([x1, x1_ident], axis=1)
x2_1 = conv_block(x1_merged,24,activation='crelu',init='orthogonal')
x2_ident = AveragePooling3D()(x1_ident)
x2_merged = concatenate([x2_1,x2_ident], axis=1)
#by branching we reduce the #params
x3_1 = conv_block(x2_merged,36,activation='crelu',init='orthogonal')
x3_ident = AveragePooling3D()(x2_ident)
x3_merged = concatenate([x3_1,x3_ident], axis=1)
x4_1 = conv_block(x3_merged,36,activation='crelu',init='orthogonal')
x4_ident = AveragePooling3D()(x3_ident)
x4_merged = concatenate([x4_1,x4_ident], axis=1)
x5_1 = conv_block(x4_merged,64,pool=False,init='orthogonal')
xpool = BatchNormalization()(GlobalMaxPooling3D()(x5_1))
xout = dense_branch(xpool,outsize=1,activation='sigmoid')
model = Model(input=xin,output=xout)
if input_shape[1] == 32 :
lr_start = 1e-5
elif input_shape[1] == 64:
lr_start = 1e-5
elif input_shape[1] == 128:
lr_start = 1e-4
elif input_shape[1] == 96:
lr_start = 5e-4
elif input_shape[1] == 16:
lr_start = 1e-6
opt = Nadam(lr_start,clipvalue=1.0)
print('compiling model')
model.compile(optimizer=opt, loss='binary_crossentropy', metrics=['accuracy'])
return model
def build_model(input_shape):
xin = Input(input_shape)
#shift the below down by one
x1 = conv_block(xin, 8, norm=True, drop_rate=0) #outputs 9 ch
x1_ident = AveragePooling3D()(xin)
x1_merged = merge([x1, x1_ident], mode='concat', concat_axis=1)
x2_1 = conv_block(x1_merged, 24, norm=True,
drop_rate=0) #outputs 16+9 ch = 25
x2_ident = AveragePooling3D()(x1_ident)
x2_merged = merge([x2_1, x2_ident], mode='concat', concat_axis=1)
#by branching we reduce the #params
x3_1 = conv_block(x2_merged, 64, norm=True,
drop_rate=0) #outputs 25 + 16 ch = 41
x3_ident = AveragePooling3D()(x2_ident)
x3_merged = merge([x3_1, x3_ident], mode='concat', concat_axis=1)
x4_1 = conv_block(x3_merged, 72, norm=True,
drop_rate=0) #outputs 25 + 16 ch = 41
x4_ident = AveragePooling3D()(x3_ident)
x4_merged = merge([x4_1, x4_ident], mode='concat', concat_axis=1)
x5_1 = conv_block(x4_merged, 72, norm=True, pool=False,
drop_rate=0) #outputs 25 + 16 ch = 41
xpool = GlobalMaxPooling3D()(x5_1)
xpool_norm = BatchNormalization()(xpool)
#xpool_norm = GaussianDropout(.1)(xpool_norm)
#from here let's branch and predict different things
xout_diam = dense_branch(xpool_norm,
name='o_d',
outsize=1,
activation='relu')
#sphericity
# xout_spher= dense_branch(xpool_norm,name='o_spher',outsize=4,activation='softmax')
# xout_text = dense_branch(xpool_norm,name='o_t',outsize=4,activation='softmax')
#calcification
# xout_calc = dense_branch(xpool_norm,name='o_c',outsize=7,activation='softmax')
xout_cad_falsepositive = dense_branch(xpool_norm,
name='o_fp',
outsize=3,
activation='softmax')
# xout_cat = merge([xout_text,xout_spher,xout_calc],name='o_cat',mode='concat', concat_axis=1)
xout_margin = dense_branch(xpool_norm,
name='o_marg',
outsize=1,
activation='sigmoid')
xout_lob = dense_branch(xpool_norm,
name='o_lob',
outsize=1,
activation='sigmoid')
xout_spic = dense_branch(xpool_norm,
name='o_spic',
outsize=1,
activation='sigmoid')
xout_malig = dense_branch(xpool_norm,
name='o_mal',
outsize=1,
activation='sigmoid')
# xout_numeric = merge([xout_margin, xout_lob, xout_spic, xout_malig],name='o_num',mode='concat',concat_axis=1)
model = Model(input=xin,
output=[
xout_diam, xout_lob, xout_spic, xout_malig,
xout_cad_falsepositive
])
if input_shape[1] == 32:
lr_start = .005
elif input_shape[1] == 64:
lr_start = .001
elif input_shape[1] == 128:
lr_start = 1e-4
elif input_shape[1] == 96:
lr_start = 5e-4
opt = Nadam(lr_start, clipvalue=1.0)
print 'compiling model'
model.compile(optimizer=opt,
loss={
'o_d': 'mse',
'o_lob': 'binary_crossentropy',
'o_spic': 'binary_crossentropy',
'o_mal': 'binary_crossentropy',
'o_fp': 'categorical_crossentropy'
},
loss_weights={
'o_d': 1.0,
'o_lob': 5.0,
'o_spic': 5.0,
'o_mal': 5.0,
'o_fp': 5.0
})
return model
def build_model(input_shape):
xin = Input(input_shape)
#shift the below down by one
x1 = conv_block(xin, 8, activation='crelu') #outputs 13 ch
x1_ident = AveragePooling3D()(xin)
x1_merged = merge([x1, x1_ident], mode='concat', concat_axis=1)
x2_1 = conv_block(x1_merged,
24,
activation='crelu',
init=looks_linear_init) #outputs 37 ch
x2_ident = AveragePooling3D()(x1_ident)
x2_merged = merge([x2_1, x2_ident], mode='concat', concat_axis=1)
#by branching we reduce the #params
x3_ident = AveragePooling3D()(x2_ident)
x3_diam = conv_block(x2_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x3_lob = conv_block(x2_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x3_spic = conv_block(x2_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x3_malig = conv_block(x2_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x3_diam_merged = merge([x3_diam, x3_ident], mode='concat', concat_axis=1)
x3_lob_merged = merge([x3_lob, x3_ident], mode='concat', concat_axis=1)
x3_spic_merged = merge([x3_spic, x3_ident], mode='concat', concat_axis=1)
x3_malig_merged = merge([x3_malig, x3_ident], mode='concat', concat_axis=1)
x4_ident = AveragePooling3D()(x3_ident)
x4_diam = conv_block(x3_diam_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x4_lob = conv_block(x3_lob_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x4_spic = conv_block(x3_spic_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x4_malig = conv_block(x3_malig_merged,
36,
activation='crelu',
init=looks_linear_init) #outputs 25 + 16 ch = 41
x4_diam_merged = merge([x4_diam, x4_ident], mode='concat', concat_axis=1)
x4_lob_merged = merge([x4_lob, x4_ident], mode='concat', concat_axis=1)
x4_spic_merged = merge([x4_spic, x4_ident], mode='concat', concat_axis=1)
x4_malig_merged = merge([x4_malig, x4_ident], mode='concat', concat_axis=1)
x5_diam = conv_block(x4_diam_merged,
64,
pool=False,
init=looks_linear_init) #outputs 25 + 16 ch = 41
x5_lob = conv_block(x4_lob_merged, 64, pool=False,
init=looks_linear_init) #outputs 25 + 16 ch = 41
x5_spic = conv_block(x4_spic_merged,
64,
pool=False,
init=looks_linear_init) #outputs 25 + 16 ch = 41
x5_malig = conv_block(x4_malig_merged,
64,
pool=False,
init=looks_linear_init) #outputs 25 + 16 ch = 41
xpool_diam = BatchNormalization()(GlobalMaxPooling3D()(x5_diam))
xpool_lob = BatchNormalization()(GlobalMaxPooling3D()(x5_lob))
xpool_spic = BatchNormalization()(GlobalMaxPooling3D()(x5_spic))
xpool_malig = BatchNormalization()(GlobalMaxPooling3D()(x5_malig))
#from here let's branch and predict different things
xout_diam = dense_branch(xpool_diam,
name='o_d',
outsize=1,
activation='relu')
xout_lob = dense_branch(xpool_lob,
name='o_lob',
outsize=1,
activation='sigmoid')
xout_spic = dense_branch(xpool_spic,
name='o_spic',
outsize=1,
activation='sigmoid')
xout_malig = dense_branch(xpool_malig,
name='o_mal',
outsize=1,
activation='sigmoid')
#sphericity
# xout_spher= dense_branch(xpool_norm,name='o_spher',outsize=4,activation='softmax')
# xout_text = dense_branch(xpool_norm,name='o_t',outsize=4,activation='softmax')
#calcification
# xout_calc = dense_branch(xpool_norm,name='o_c',outsize=7,activation='softmax')
model = Model(input=xin,
output=[xout_diam, xout_lob, xout_spic, xout_malig])
if input_shape[1] == 32:
lr_start = .003
elif input_shape[1] == 64:
lr_start = .001
elif input_shape[1] == 128:
lr_start = 1e-4
elif input_shape[1] == 96:
lr_start = 5e-4
opt = Nadam(lr_start, clipvalue=1.0)
print 'compiling model'
model.compile(optimizer=opt,
loss={
'o_d': 'mse',
'o_lob': 'binary_crossentropy',
'o_spic': 'binary_crossentropy',
'o_mal': 'binary_crossentropy'
},
loss_weights={
'o_d': 1.0,
'o_lob': 5.0,
'o_spic': 5.0,
'o_mal': 5.0
})
return model