Dilated U-Net Keras版本实现

最新推荐文章于 2025-02-07 22:52:40 发布

脉望虫

最新推荐文章于 2025-02-07 22:52:40 发布

阅读量1.5k

点赞数

CC 4.0 BY-SA版权

分类专栏： Python 文章标签：深度学习 tensorflow 神经网络卷积

本文链接：https://blog.youkuaiyun.com/qq_36763031/article/details/108475711

Python 专栏收录该内容

16 篇文章

订阅专栏

本文介绍一种改进的U-Net模型，用于MRI头颈部肿瘤的分割。该模型引入了空洞卷积和普通卷积的双通路结构，提高了特征提取能力。通过Keras实现并在肿瘤分割任务上取得良好效果。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

最近读了一篇会议文章，提出在编码路径每个阶段汇中，使用除了一般卷积两个卷积之外还有一个空洞卷积路径的双通路结构，文章中称为Modified U-Net分割MRI中头颈部位的肿瘤，文章名称为：Segmentation of Head and Neck Tumours Using Modified U-net。网络结构为：

网络结构

下面给出该网络Keras版本的复现：

#dilated UNet
import tensorflow.python.keras.backend as K
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.layers import Input, ZeroPadding2D, concatenate, add
from tensorflow.python.keras.layers.core import Dropout, Activation
from tensorflow.python.keras.layers.convolutional import UpSampling2D, Conv2D, Conv2DTranspose
from tensorflow.python.keras.layers.pooling import AveragePooling2D, MaxPooling2D
from tensorflow.python.keras.layers.normalization import BatchNormalization
from tensorflow.python.keras.optimizers import nadam, adam, SGD
import losses


def doubleConv2D(input, outdim, is_dilated = False, is_batchnorm=False, name = ''):
    #Double  convolution layer
    if is_dilated:
        dilated_rated = (3, 3)
    else:
        dilated_rated = (1, 1)
    x = Conv2D(outdim, (3, 3), strides=(1, 1), dilation_rate = dilated_rated, kernel_initializer='glorot_normal', padding="same", name=name + '_1')(input)
    if is_batchnorm:
        x =BatchNormalization(name=name + '_1_bn')(x)
    x = Activation('relu',name=name + '_1_act')(x)

    x = Conv2D(outdim, (3, 3), strides=(1, 1), dilation_rate = dilated_rated, kernel_initializer='glorot_normal', padding="same", name=name + '_2')(x)
    if is_batchnorm:
        x = BatchNormalization(name=name + '_2_bn')(x)
    x = Activation('relu', name=name + '_2_act')(x)
    return x

def dualConv2D(input, outdim, is_batchnorm=False, name = ''):
    '''
    parallel convolution block in which the upper is dilated convolution and the lower is
    common 3*3 convolution
    '''
    conv_x = doubleConv2D(input, outdim, is_batchnorm=False, name = name + '_conv')
    dilated_x = doubleConv2D(input, outdim, is_dilated = True, is_batchnorm=False, name = name + '_dilated')
    concat_x = concatenate([dilated_x , conv_x], name = name + '_concat')
    return concat_x

def dilatedUNet(opt, input_size, classes, lossfxn):

    img_input = Input(shape=input_size, name='input')
    dualblock1 = dualConv2D(img_input, 32, is_batchnorm=False, name = 'encoder_stage_1')
    pool1 = MaxPooling2D(pool_size=(2, 2))(dualblock1)

    dualblock2 = dualConv2D(pool1, 64, is_batchnorm=False, name='encoder_stage_2')
    pool2 = MaxPooling2D(pool_size=(2, 2))(dualblock2)

    dualblock3 = dualConv2D(pool2, 128, is_batchnorm=False, name='encoder_stage_3')
    pool3 = MaxPooling2D(pool_size=(2, 2))(dualblock3)

    dualblock4 = dualConv2D(pool3, 256, is_batchnorm=False, name='encoder_stage_4')
    pool4 = MaxPooling2D(pool_size=(2, 2))(dualblock4)

    conv5 = doubleConv2D(pool4, 1024, is_batchnorm=False, name = 'encoder_stage_5')

    up1 = concatenate(
        [Conv2DTranspose(512, (3, 3), strides=(2, 2), padding='same', activation='relu', kernel_initializer='glorot_normal')(conv5),
         dualblock4], name='up1')
    upconv1 = doubleConv2D(up1, 512, is_batchnorm=False, name = 'decoder_stage_1')

    up2 = concatenate(
        [Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same', activation='relu',
                         kernel_initializer='glorot_normal')(upconv1),
         dualblock3], name='up2')

    upconv2 = doubleConv2D(up2, 256, is_batchnorm=False, name='decoder_stage_2')

    up3 = concatenate(
        [Conv2DTranspose(128, (3, 3), strides=(2, 2), padding='same', activation='relu',
                         kernel_initializer='glorot_normal')(upconv2),
         dualblock2], name='up3')

    upconv3 = doubleConv2D(up3, 128, is_batchnorm=False, name='decoder_stage_3')

    up4 = concatenate(
        [Conv2DTranspose(64, (3, 3), strides=(2, 2), padding='same', activation='relu',
                         kernel_initializer='glorot_normal')(upconv3),
         dualblock1], name='up4')

    upconv4 = doubleConv2D(up4, 128, is_batchnorm=False, name='decoder_stage_4')

    out = Conv2D(classes, (1, 1), activation='softmax', name='final')(upconv4)

    model = Model(inputs=img_input, outputs= out )

    model.compile(optimizer=opt, loss=lossfxn, metrics=[losses.dsc])

    return model

实际表现尚待检验！