import keras
from keras.layers import Conv3D, MaxPooling3D, Conv3DTranspose, Input
from keras.models import Model

def FCN3DNet():

    inputLayerA = Input(shape=(96,96,96,1))
    inputLayerB = Input(shape=(96,96,96,1))

    totalInput = keras.layers.concatenate([inputLayerA, inputLayerB])

    x = Conv3D(16, (2,2,2), padding='same', data_format='channels_last', activation='relu')(totalInput)

    x = MaxPooling3D(pool_size=(2,2,2))(x)

    x = Conv3D(32, (2,2,2), padding='same', activation='relu')(x)

    x = MaxPooling3D(pool_size=(2,2,2))(x)

    x = Conv3D(64, (2,2,2), padding='same', activation='relu')(x)

    x = MaxPooling3D(pool_size=(2,2,2))(x)

    x = Conv3DTranspose(32, (2,2,2), strides=(2,2,2), padding='same', activation='relu')(x)

    x = Conv3DTranspose(16, (2,2,2), strides=(2,2,2), padding='same', activation='relu')(x)

    x = Conv3DTranspose(3, (2,2,2), strides=(2,2,2), padding='same', activation='linear')(x)

    model = Model(inputs=[inputLayerA, inputLayerB], outputs=x)
    return model


model = FCN3DNet()


print(model.summary())
## The shapes reported in this summary have the number of channels truncated