论文3D Deeply Supervised Network f

论文地址：https://arxiv.org/abs/1607.00582
这是一篇MICCAI 2016关于肝脏分割的论文，使用了3D卷积神经网络，难点是虚线里面的部分，如何体现出三个输出的监督作用，最初感觉是在损失函数里面体现出来，尝试把损失函数的一个预测变量改成三个，但运行网络总是报错，后来反复理解这三个输出如何起作用，以及下图中三个输出和label之间的虚线，搜索了与多输出有关的技术博客，最后完全理解了这个网络结构的运行过程，具体到代码的差异性如下红框所示：

模型结构图.png
模型结构在下采样的过程中有三个输出分支，使用Deconvolution来进行上采样，模型结构比较容易理解，主要是代码实现过程中和平时有一点差异，下图红框是需要注意的地方：
Keras代码.png
整个网络架构的代码如下：

    def DSN(self):
        
        inputs = Input((32,self.img_rows, self.img_cols,1))
         

        conv1 = Conv3D(8, (7, 9, 9), padding='same', activation= 'selu', kernel_initializer = 'he_normal')(inputs)
        conv1 = Conv3D(8, (7, 9, 9), padding='same',  activation= 'selu',kernel_initializer = 'he_normal')(conv1)
        print("conv1 shape:", conv1.shape)
        pool1 = MaxPooling3D(pool_size=(2, 2, 2))(conv1)
        print("pool1 shape:", pool1.shape)
        convT1 = Conv3DTranspose(2, (2, 2, 2), padding='valid', strides = (2, 2, 2), activation= 'selu', kernel_initializer = 'he_normal')(pool1)
        print("convT1 shape:", convT1.shape)
        out1 = Conv3D(1, 1, activation = 'softmax')(convT1)

        conv2 = Conv3D(16, (5, 7, 7), padding='same', activation= 'selu', kernel_initializer = 'he_normal')(pool1)
        conv2 = Conv3D(32, (5, 7, 7), padding='same',  activation= 'selu',kernel_initializer = 'he_normal')(conv2)
        print("conv2 shape:", conv2.shape)
        pool2 = MaxPooling3D(pool_size=(2, 2, 2))(conv2)
        print("pool2 shape:", pool2.shape)
        convT2 = Conv3DTranspose(2, (2, 2, 2), padding='valid', strides = (2, 2, 2), activation= 'selu', kernel_initializer = 'he_normal')(pool2)
        convT2 = Conv3DTranspose(2, (2, 2, 2), padding='valid', strides = (2, 2, 2), activation= 'selu', kernel_initializer = 'he_normal')(convT2)        
        print("convT2 shape:", convT2.shape)        
        out2 = Conv3D(1, 1, activation = 'softmax')(convT2)
        
        conv3 = Conv3D(32, (3, 5, 5), padding='same', activation= 'selu', kernel_initializer = 'he_normal')(pool2)
        conv3 = Conv3D(32, (1, 1, 1), padding='same',  activation= 'selu',kernel_initializer = 'he_normal')(conv3)
        print("conv3 shape:", conv3.shape)
        convT3 = Conv3DTranspose(2, (2, 2, 2), padding='valid', strides = (2, 2, 2), activation= 'selu', kernel_initializer = 'he_normal')(conv3)
        convT3 = Conv3DTranspose(2, (2, 2, 2), padding='valid', strides = (2, 2, 2), activation= 'selu', kernel_initializer = 'he_normal')(convT3)
        print("convT3 shape:", convT3.shape)        
        out3 = Conv3D(1, 1, activation = 'sigmoid')(convT3)        
        
        model = Model(input=inputs, output=[out3, out2, out1])
        adam = Adam(lr=0.0001)
        model.summary()
         
        model.compile(optimizer=adam, loss=self.dice_coef_loss, loss_weights=[0.6, 0.3, 0.1])
        with open('seg_liver3D.json', 'w') as files:
            files.write(model.to_json())
        print('model compile')
        return model

    def train(self):
        print("loading data")
        imgs_train, label_train = self.load_train_data()
        print("loading data done")
        model = self.get_unet()
        print("got unet")

        # 保存的是模型和权重,
        model_checkpoint = ModelCheckpoint('seg_liver3D.h5', monitor='loss', verbose=0, save_best_only=True, 
                                           save_weights_only=True, mode='min')
        print('Fitting model...')
        model.fit(imgs_train, [label_train, label_train, label_train] ,batch_size=2, epochs=15, verbose=1, 
                  callbacks=[model_checkpoint], validation_split=0.2, shuffle=True)

网络架构中使用转置卷积来进行上采样过程。