摘要
针对乳腺超声图像分类中的标签噪音问题,该文设计了一种协作标签修正网络(COLC-Net)。该方法基于乳腺超声BI-RADS评级噪音分布特点,为乳腺超声图像定义了软标签,并设计了双网络协作训练,以蒸馏优秀知识修正软标签。随着软标签准确性的增加,可以降低噪音标签负作用,并增强准确标签知识的学习。与现有最新方法进行比较,结果证实了该方法具有更好的效果。
In order to solve the problem of label noise in breast ultrasound image classification,an efficient method called cooperative label correction network(COLC-Net)is proposed.In this method,based on the noise distribution characteristics of the breast ultrasound BI-RADS(breast imaging-reporting and data system)rating,soft labels are proposed for breast ultrasound images,and two networks are proposed for collaborative training.Excellent knowledge is distilled from the two networks to modify the soft labels.With the increase of the accuracy of soft labels,the negative effects of noise labels can be reduced and the learning of clean labels can be enhanced.In order to verify the effectiveness of the method,extensive comparisons are conducted with existing state-of-theart methods on the dataset.The results demonstrate the effectiveness of the proposed method.
作者
曹占涛
杨国武
陈琴
吴尽昭
李晓瑜
CAO Zhan-tao;YANG Guo-wu;CHEN Qin;WU Jin-zhao;LI Xiao-yu(School of Computer Science and Engineering,University of Electronic Science and Technology of China Chengdu 611731;School of Medicine,University of Electronic Science and Technology of China Chengdu 610072;College of Science,Guangxi University for Nationalities Nanning 530006;School of Information and Software Engineering,University of Electronic Science and Technology of China Chengdu 610054)
出处
《电子科技大学学报》
EI
CAS
CSCD
北大核心
2020年第4期597-602,共6页
Journal of University of Electronic Science and Technology of China
基金
国家自然科学基金(61572109,61772006)
广西“八桂学者”专项。
关键词
乳腺超声图像
深度学习
噪音标签
弱监督学习
breast ultrasound image
deep learning
noisy label
weakly supervised learning
