卷积神经网络的宫颈细胞图像分类  

Classification of Cervical Cells Based on Convolution Neural Network

作　　者：赵越 曾立波[1] 吴琼水[1] Zhao Yue;Zeng Libo;Wu Qiongshui(Electronic Information School,Wuhan University,Wuhan 430079)

机构地区：[1]武汉大学电子信息学院,武汉430079

出　　处：《计算机辅助设计与图形学学报》2018年第11期2049-2054,共6页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家科技支撑计划(2011BAF02B00)

摘　　要：为实现计算机辅助系统精准、快速地检测宫颈异常细胞,提出一种基于卷积神经网络的宫颈细胞自动分类方法.首先复制预训练网络结构及参数来初始化分类网络,将宫颈细胞图像分批次传入网络;然后采用Softmax函数将网络输出数据归一化为各标签对应的概率值,并使用交叉熵作为损失函数;最后改进网络结构加入对数据的批归一化处理,通过反向传播算法优化参数使损失函数最小化,最终选择训练所得最优网络.使用5折交叉验证法在Herlev数据集上的实验结果表明,对比Herlev常用基准方法,该方法的特异性、调和平均数和准确率分别提高了19.46%, 10.71%和5.09%.To achieve accurate and rapid detection for abnormal cervical cells in computer-assisted cytology test,an automatic classification method based on convolution neural network is proposed.First,the classification network was initialized with the pre-trained network structure and parameters,and cervical cell images were imported into it in batches.Then the output data is normalized to the probability of each label by Softmax,and cross-entropy is set as the loss function.The network structure was improved by batch normalization,and parameters were optimized by back propagation to minimize the loss function.Eventually,the optimal network was selected.The five-fold cross validation shows specificity,H-mean and F-measure are improved by 19.46%,10.71%and 5.09%respectively in contrast with benchmark on Herlev dataset.

关 键 词：卷积神经网络 宫颈细胞 迁移学习 批归一化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]