如何搭建Unet-3D网络

UNET模型是由卷积神经网络演变过来，一般的卷积神经网络将其任务集中在图像分类上，其中输入多是图像输出是标签，而在医学和其他学科中，需要输出是图像，输出也是图像。
UNET模型主要解决的就是这类问题，它能够定位和区分边界，并且输出的大小和输出的大小相同。

文章目录

- 模型概况
- - 用2D网络逐行分析
  - - - 1. 下采样，卷积
      - 2. 上采样，反卷积

模型概况

# 下采样部分
def build_model(input_layer, start_neurons):conv1 = Conv3D(start_neurons * 1, (3, 3), activation='relu', padding="same")(input_layer)conv1 = Conv3D(start_neurons * 1, (3, 3), activation='relu', padding="same")(conv1)pool1 = MaxPool3D((2, 2))(conv1)pool1 = Dropout(0.25)(pool1)conv2 = Conv3D(start_neurous * 2, (3, 3), activation='relu', padding="same")(pool1)conv2 = Conv3D(start_neurous * 2, (3, 3), activation='relu', padding="same")(conv2)pool2 = MaxPool3D((2, 2))(conv2)pool2 = Dropout(0.5)(pool2)conv3 = Conv3D(start_neurous * 4, (3, 3), activation='relu', padding="same")(pool2)conv3 = Conv3D(start_neurous * 4, (3, 3), activation='relu', padding="same")(conv3)pool3 = MaxPool3D((2, 2))(conv3)pool3 = Dropout(0.5)(pool3)conv4 = Conv3D(start_neurous * 8, (3, 3), activation='relu', padding="same")(pool3)conv4 = Conv3D(start_neurous * 8, (3, 3), activation='relu', padding="same")(conv4)pool4 = MaxPool3D((2, 2))(conv4)pool4 = Dropout(0.5)(pool4)# 中间部分convm = Conv3D(start_neurous * 16, (3, 3), activation='relu', padding="same")(pool4)convm = Conv3D(start_neurous * 16, (3, 3), activation='relu', padding="same")(convm)# 上采样部分	deconv4 = Conv3DTranspose(start_neurons * 8, (3, 3), strides=(2, 2), padding="same")(convm)uconv4 = concatenate([deconv4, conv4])uconv4 = Dropout(0.5)(uconv4)uconv4 = Conv3D(start_neurons * 8, (3, 3), activation='relu', padding="same")(uconv4)uconv4 = Conv3D(start_neurons * 8, (3, 3), activation='relu', padding="same")(uconv4)deconv3 = Conv3DTranspose(start_neurons * 4, (3, 3), strides=(2, 2), padding="same")(uconv4)uconv3 = concatenate([deconv3, conv3])uconv3 = Dropout(0.5)(uconv3)uconv3 = Conv3D(start_neurons * 4, (3, 3), activation='relu', padding="same")(uconv3)uconv3 = Conv3D(start_neurons * 4, (3, 3), activation='relu', padding="same")(uconv3)deconv2 = Conv3DTranspose(start_neurons * 2, (3, 3), strides=(2, 2), padding="same")(uconv3)uconv2 = concatenate([deconv2, conv2])uconv2 = Dropout(0.5)(uconv2)uconv2 = Conv3D(start_neurons * 2, (3, 3), activation='relu', padding="same")(uconv2)uconv2 = Conv3D(start_neurons * 2, (3, 3), activation='relu', padding="same")(uconv2)deconv1 = Conv3DTranspose(start_neurons * 1, (3, 3), strides(2, 2), padding="same")(uconv2)uconv1 = concatenate([deconv1, conv1])uconv1 = Dropout(0.5)(uconv1)uconv1 = Conv3D(start_neurons * 1, (3, 3), acvivation='relu', padding="same")(uconv1)uconv1 = Conv3D(start_neurons * 1, (3, 3), activation='relu', padding="same")(uconv1)output_layer = Conv3D(start_neurons * 1, (1, 1), activation='relu', padding="same")(uconv1)input_layer = Input(image_size_target, image_size_target, 1)
output_layer = build_model(input_layer, 16)

用2D网络逐行分析

用2D网络分析，只是为了方便使用原文章里的图片，3D和2D网络搭建上基本没有区别，这样也可以做到对比学习。

1. 下采样，卷积

卷积路径：

conv_layer1 -> conv_layer2 -> max_pooling -> dropout(optional)

这部分结构对应的代码是：

conv1 = Conv2D(start_neurons * 1, (3, 3), activation="relu", padding="same")(input_layer)
conv1 = Conv2D(start_neurons * 1, (3, 3), activation="relu", padding="same")(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
pool1 = Dropout(0.25)(pool1)

图中对应的结构是：
在这里插入图片描述
注意，每个进程构成两个卷积层，通道数从1变成64，因为卷积过程会增加图像的深度。红色向下的箭头是最大池化处理，它将图像大小减半（从572x572到568x568，变小是由于填充问题）padding采用的是“same”。
该过程重复三次，代表卷积深度三次：
在这里插入图片描述
具体代码：

conv2 = Conv2D(start_neurons * 2, (3, 3), activation="relu", padding="same")(pool1)
conv2 = Conv2D(start_neurons * 2, (3, 3), activation="relu", padding="same")(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
pool2 = Dropout(0.5)(pool2)conv3 = Conv2D(start_neurons * 4, (3, 3), activation="relu", padding="same")(pool2)
conv3 = Conv2D(start_neurons * 4, (3, 3), activation="relu", padding="same")(conv3)
pool3 = MaxPooling2D((2, 2))(conv3)
pool3 = Dropout(0.5)(pool3)conv4 = Conv2D(start_neurons * 8, (3, 3), activation="relu", padding="same")(pool3)
conv4 = Conv2D(start_neurons * 8, (3, 3), activation="relu", padding="same")(conv4)
pool4 = MaxPooling2D((2, 2))(conv4)
pool4 = Dropout(0.5)(pool4)

现在到达了UNET的最底层:

convm = Conv2D(start_neurons * 16, (3, 3), activation="relu", padding="same")(pool4)
convm = Conv2D(start_neurons * 16, (3, 3), activation="relu", padding="same")(convm)

此时图像大小已经是28x28x1024了，下采样已经结束，开始上采样。

2. 上采样，反卷积

反卷积路径：

conv_2d_transpose -> concatenate -> conv_layer1 -> conv_layer2

图中具体结构对应：
在这里插入图片描述
对应代码：

deconv4 = Conv2DTranspose(start_neurons * 8, (3, 3), strides=(2, 2), padding="same")(convm)
uconv4 = concatenate([deconv4, conv4])
uconv4 = Dropout(0.5)(uconv4)
uconv4 = Conv2D(start_neurons * 8, (3, 3), activation="relu", padding="same")(uconv4)
uconv4 = Conv2D(start_neurons * 8, (3, 3), activation="relu", padding="same")(uconv4)

这里反卷积是一种扩大图像大小的技术，也就是上采样技术。基本上是对图像进行填充，然后进行卷积运算。
反卷积后，图像从28x28x1024放到到56x56x512，然后该图像与来自卷积操作中的相应位置的图像进行张量拼接，生成一个大小为56x56x1024的图像，主要是为了更精确的预测，与下采样是的图像进行结合。第四行和第五行是另外加了两个卷积层。
与下采样相同，这个过程重复3遍：

deconv3 = Conv2DTranspose(start_neurons * 4, (3, 3), strides=(2, 2), padding="same")(uconv4)
uconv3 = concatenate([deconv3, conv3])
uconv3 = Dropout(0.5)(uconv3)
uconv3 = Conv2D(start_neurons * 4, (3, 3), activation="relu", padding="same")(uconv3)
uconv3 = Conv2D(start_neurons * 4, (3, 3), activation="relu", padding="same")(uconv3)deconv2 = Conv2DTranspose(start_neurons * 2, (3, 3), strides=(2, 2), padding="same")(uconv3)
uconv2 = concatenate([deconv2, conv2])
uconv2 = Dropout(0.5)(uconv2)
uconv2 = Conv2D(start_neurons * 2, (3, 3), activation="relu", padding="same")(uconv2)
uconv2 = Conv2D(start_neurons * 2, (3, 3), activation="relu", padding="same")(uconv2)deconv1 = Conv2DTranspose(start_neurons * 1, (3, 3), strides=(2, 2), padding="same")(uconv2)
uconv1 = concatenate([deconv1, conv1])
uconv1 = Dropout(0.5)(uconv1)
uconv1 = Conv2D(start_neurons * 1, (3, 3), activation="relu", padding="same")(uconv1)
uconv1 = Conv2D(start_neurons * 1, (3, 3), activation="relu", padding="same")(uconv1)

现在已经走到了最高层
在这里插入图片描述
输出:

output_layer = Conv2D(1, (1,1), padding="same", activation="sigmoid")(uconv1)