基于半监督生成对抗网络X光图像分类算法  

X-Ray Image Classification Algorithm Based on Semi-Supervised Generative Adversarial Networks

作　　者：刘坤[1] 王典 荣梦学 Liu Kun;Wang Dian;Rong Mengxue(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院

出　　处：《光学学报》2019年第8期109-117,共9页Acta Optica Sinica

基　　金：国家自然科学基金(61271446);航空科学基金(2013ZC15005)

摘　　要：利用半监督学习体系结构中的生成对抗性网络,围绕标注数据稀缺性的问题进行研究,在传统无监督生成对抗网络的基础上用softmax替代最后的输出层,使其扩展为半监督生成对抗网络。对生成样本定义额外的类别标签,用于引导训练,采用半监督训练方式对网络参数进行优化,并将训练得到的判别网络运用于X光图像分类中。对于胸部X光图像,结合自动化分类诊断选取了6种肺部疾病的X光前视图进行实验,结果表明:所提算法提高了利用标注数据的监督学习性能,与其他半监督分类方法相比具有优越的性能。A generative adversarial network(GAN)in the semi-supervised learning architecture was used to address the problem of the scarcity of labeled data in X-ray image classification.Initially,we used a softmax layer to replace the output layer of an unsupervised GAN,extending it to a semi-supervised GAN.In addition,we defined additional labels for the GAN-synthesized samples to guide the training process and optimized the network parameters using a semi-supervised training strategy.Then,the discriminator network obtained by the training was used for X-ray image classification.From tested front-view chest X-ray images of six lung diseases,we find that the proposed method substantially enhances the supervised learning with limited labeled data.Further,the proposed method demonstrates superior classification performance compared with other semi-supervised methods.

关 键 词：图像处理 图像分类 X光图像 生成对抗网络 半监督学习 标注数据 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]