很久以前就看到了unet3 医学图像分割的论文，本来想直接去github找keras/Tensorflow为了实现，似乎所有找到的东西都与源代码不同，所以我根据论文和源代码写了它们，但我不能保证它们与源代码完全一致，并发送它们来吸引玉石。很多讲unet3 博客写得很好。如果你想知道全文，你可以看看这个翻译【UNet3 (UNet )论文解读 九零猴，本文还简单介绍了自己的理解。

unet3 论文 源码（Pytorch）

简单来说，unet3 有三个特点： 1 跨尺度连接，防止语义在下采样/上采样之间损失 2 全尺度深监督，学习深层次特征 3 为了消除医学图像中噪声引起的假阳性分割，提出了分类指导模块 4 新的混合损失函数（TODO）

呃，前三点其实各有槽点，后面再说

注：孩子不懂事，代码写着玩，不一定正确，如有问题欢迎指出和讨论，转载请注明出处。 CGM输出这块的实现还是有待商榷的，我的代码里CGM分割掩膜分别输出，后面要手动乘以。

import tensorflow as tf import numpy as np from keras.models import Model from keras.layers import Conv2D, Input, concatenate, MaxPooling2D, UpSampling2D, Activation, BatchNormalization, LayerNormalization, Dropout, GlobalMaxPooling2D 

# helper function to build unet3  def normalization(input_tensor, normalization):      if normalization=='batch':         return(BatchNormalization()(input_tensor))     elif normalization=='layer':         return(LayerNormalization()(input_tensor))     elif normalization == None:         return input_tensor     else: raise ValueError('Invalid normalization') def conv2d_block(input_tensor, filters, kernel_size, norm_type, use_residual, act_type='relu', double_features = False, dilation=[1, 1]): x = Conv2D(filters, kernel_size, padding='same', dilation_rate=dilation[0], use_bias=False, kernel_initializer='he_normal')(input_tensor) x = normalization(x, norm_type) x = Activation(act_type)(x) if double_features: filters *= 2 x = Conv2D(filters, kernel_size, padding='same', dilation_rate=dilation[1], use_bias=False, kernel_initializer='he_normal')(x) x = normalization(x, norm_type) if use_residual: if K.int_shape(input_tensor)[-1] != K.int_shape(x)[-1]: shortcut = Conv2D(filters, kernel_size=1, padding='same', use_bias=False, kernel_initializer='he_normal')(input_tensor) shortcut = normalization(shortcut, norm_type) x = add([x, shortcut]) else: x = add([x, input_tensor]) x = Activation(act_type)(x) return x def down_layer_2d(input_tensor, down_pattern, filters, norm_type=None): if down_pattern == 'maxpooling': x = MaxPooling2D(pool_size=(2, 2))(input_tensor) elif down_pattern == 'avgpooling': x = AveragePooling2D(pool_size=(2, 2))(input_tensor) elif down_pattern == 'conv': x = Conv2D(filters, kernel_size=(2, 2), strides=(2, 2), padding='same', use_bias=False if norm_type is None else True, kernel_initializer='he_normal')(input_tensor) normalization(x, norm_type) elif down_pattern == 'normconv': x = normalization(input_tensor, norm_type) x = Conv2D(filters, kernel_size=(2, 2), strides=(2, 2), padding='same', kernel_initializer='he_normal')(x) else: raise ValueError('Invalid down_pattern') return x def conv_norm_act(input_tensor, filters, kernel_size , norm_type='batch', act_type='relu', dilation=1): output_tensor = Conv2D(filters, kernel_size, padding='same', dilation_rate=(dilation, dilation), use_bias=False if norm_type is not None else True, kernel_initializer='he_normal')(input_tensor) output_tensor = normalization(output_tensor, normalization=norm_type) output_tensor = Activation(act_type)(output_tensor) return output_tensor def aggregate(l1, l2, l3, l4, l5, filters, kernel_size, norm_type='batch', act_type='relu'): out = concatenate([l1, l2, l3, l4, l5], axis = -1) out = Conv2D(filters * 5, kernel_size, padding = 'same', use_bias=False if norm_type is not None else True, kernel_initializer = 'he_normal')(out) out = normalization(out, norm_type) out = Activation(act_type)(out) return out def cgm_block(input_tensor, class_num, dropout_rate = 0.): x = Dropout(rate = dropout_rate)(input_tensor) x = Conv2D(class_num, 1, padding='same', kernel_initializer='he_normal')(x) # x = BatchNormalization()(x) x = GlobalMaxPooling2D()(x) # 用全局最大池化代替原文中的自适应最大池化，这里的效果应该是一样的 x = Activation('sigmoid', name='cgm_output')(x) # x = Lambda(lambda x: K.expand_dims(x, axis=1))(x) # x = Lambda(lambda x: K.expand_dims(x, axis=1), name = 'cgm_output')(x) # x = Reshape((batch_size, 1, 1, class_num))(x) return x