基于深度对抗生成网络的彩超图像半监督分类研究

Semi-Supervised Classification for Ultrasound Images Based on GAN

深层神经网络模型逐渐被应用于乳腺超声等医学图像的分类。但乳腺彩超图像标签通常存在人为或自然的噪音。特别是深度神经网络的性能受数据量的影响,严重依赖于标注人员的专业能力。针对乳腺超声图像中固有的噪声和不确定性,提出一种基于生成对抗网络(GANs)的半监督学习分类模型。该方法充分利用噪声数据或所谓未标记数据的信息,有可能扩展学习到噪音数据中平滑分布的底层特征。该模型使用高像素的乳腺彩超图像,并使用指数移动平均(EMA)正则化来保证在线模型的稳定性,且能在以从非常小的训练数据集下获得良好的结果。

Deep neural network model is gradually applied to the classification of Breast Ultrasound Images.However,the label of data is influenced by noises coming from annotators or the data itself.In particular,the performance of deep neural network is affected by the amount of data,which seriously depends on the professional experience of annotator.In view of these inherent noise and uncertainty,a semi supervised learning model based on Generative Antagonism Networks(GANs)is proposed in this paper,to solve this problem.This method full the information of noise data or so-called unlabeled data,and it is possible to learn the features of the noise data.We can generate large-scale breast ultrasound images and use Exponential Moving Average(EMA)regularization to ensure the stability of the online model.Finally,this model can acquire good results from very datasets with a small volume of labeled data.