基于难样本挖掘和深度学习的乳腺癌检测方法  

Breast cancer detection method based on difficult sample mining and deep learning

作　　者：张永梅[1] 陈彤 马健喆 胡蕾[3] ZHANG Yong-mei;CHEN Tong;MA Jian-zhe;HU Lei(School of Information Science and Technology,North China University of Technology,Beijing 100144,China;Department of Electronic and Information Engineering,Hong Kong Polytechnic University,Hong Kong 999077,China;School of Computer Information Engineering,Jiangxi Normal University,Nanchang 330022,China)

机构地区：[1]北方工业大学信息学院,北京100144 [2]香港理工大学电子与信息工程系,中国香港999077 [3]江西师范大学计算机信息工程学院,江西南昌330022

出　　处：《计算机工程与设计》2021年第6期1727-1734,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61371143、61662033);教育部高等教育司产学合作协同育人基金项目(201801121002);全国高等学校计算机教育研究会2019年度课题基金项目(CERACU2019R05);“天诚汇智”创新促教基金项目(2018A03029);2019年度北京市教委基本科研业务费基金项目(110052971921/002)。

摘　　要：为解决当前乳腺癌影像检测任务中数据样本量少、医学专业性强等问题,提出一种基于难样本挖掘和深度学习的乳腺癌检测方法。以乳腺癌医学图像数据库中含有正负样本的X光影像训练模型,通过改进Mask R-CNN特征金字塔结构对目标区域的高低层特征充分学习,利用难样本挖掘方法对正样本及难负样本进一步筛选,降低患病区域误检率,避免深度学习模型依赖样本量所造成的过拟合问题。实验结果表明,该方法在公开数据集上的检测精度达到85.92%-86.75%,精度提高了3%左右,提高了原检测方法的准确性和鲁棒性。To solve the problems of low amount and strong medical expertise of data sets in current breast cancer image detection tasks,a breast cancer detection method based on difficult sample mining and deep learning was proposed.The X-ray image containing positive and negative samples in the breast cancer database were used.The feature pyramid structure of Mask R-CNN was improved to fully learn the high-level and low-level features of the target areas and the difficult sample mining method was combined to further select positive and negative samples.The false detection rate in the diseased area was decreased and the over-fitting problem caused by the deep learning model relying on the sample amount was avoided.Experimental results show the detection accuracy of the method on public data sets reaches 85.92%-86.75%,the accuracy is improved by 3%,and the accuracy and robustness of the original detection method are improved.

关 键 词：深度学习 乳腺癌检测 难样本挖掘 特征金字塔 多尺度特征 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]