基于深度神经网络的癌细胞识别系统研究.docxVIP

基于深度神经网络的癌细胞识别系统研究 杨晓玲　王振奇　李嘉 摘 要：癌症是发病率和死亡率极高的疾病，癌细胞正确识别与癌症等级正确判断具有极其重要的意义。深度神经网络（DNN）可用神经网络模拟大脑识别过程，底层提取初级特征，高层对底层特征进行组合与抽象。以乳腺癌细胞图像为例，采用BreaKHis官网数据集，在Linux操作系统安装Pycharm开发软件，以Tensorflow为框架，搭载Python2.7编译环境，增加现有神经网络的卷积层数和全连接层数，提出一种优化的深度神经网络癌细胞识别方法。实验结果表明，该方法能更加准确地识别癌细胞图像特征，有效降低现有神经网络分类错误，对癌细胞平均识别率达89.58%，对恶性癌细胞识别率最高可达96.75%。 关键词：癌细胞识别;数据集;神经网络;训练速度 DOI：10. 11907/rjdk. 191574 中图分类号：TP303 文献标识码：A 文章編号：1672-7800（2020）003-0065-04 Research Cancer Cell Recognition System Based on Deep Neural Network YANG Xiao-ling，WANG Zhen-qi，LI Jia （School of Electronic Information Engineering，Zhuhai College of Jilin University，Zhuhai 519041，China） Abstract： Cancer has become a major disease with high morbidity and mortality in China. Correct identification of cancer cells and correct judgement of cancer grade are of great significance to the development of Chinese medicine. This system employs official websites BreaKHis Data and takes breast cancer cell image as an example. Pycharm development software was installed on Linux operating system in Python2.7 compiler environment within the framework of Tensorflow to speed up network training and deepen convolutional layers and fully layers of existing neural networks. A cancer cell recognition method based on optimized deep neural network is proposed. The experimental results show that， this method can recognize the image features of cancer cells more accurately， effectiving reduce the existing neural networks classification errors， the average recognition rate of cancer cells was 89.58%， and the highest recognition rate of malignant cancer cells was 96.75%. Key Words： cancer cell recognition; data set; neural network; training speed 0 引言 深度学习算法是近几年兴起的特征学习及分类算法，具有强大的特征学习能力，可同时实现特征学习与分类。深度学习在癌细胞识别中的应用尚不普遍，目前最具代表性的研究是Cruz-Roa等[1]基于深度学习网络的基底细胞癌（Basal-Cell Carcinoma Cancer）自动检测系统。深度学习模型不仅能大幅提高图像识别精度，而且避免消耗大量时间进行人工特征提取，使运行效率大大提升。普通神经网络和深度神经网络区别如图1所示[2]。深度学习通过多个处理层组成复杂的计算模型，自动获取数据的表示与多个抽