基于迁移学习的乳腺结构紊乱异常识别  

Architectural distortion recognition in mammograms based on transfer learning

作　　者：刘小明[1,2] 翟蕾蕾[1,2] 朱婷[1,2] 

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]武汉科技大学智能信息处理与实时工业系统湖北省重点实验室,湖北武汉430065

出　　处：《计算机工程与设计》2017年第9期2530-2535,2579,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61403287;61472293;31201121);中国博士后科学基金项目(2014M552039);湖北省自然科学基金项目(2014CFB288)

摘　　要：针对乳腺X图像中结构紊乱识别困难、样本数量较少的问题,提出基于迁移学习的结构紊乱识别方法,把基于Gabor的毛刺模式特征、GLCM特征以及熵特征等新特征运用其中。基于恶性肿块与结构紊乱的相似性,把恶性肿块作为源域中正样本,负样本由结构紊乱检测算法中的伪正样本构成,对正负样本区域提取多种特征,把结构紊乱作为目标域的训练和测试集分别进行特征提取,使用自适应支持向量机(A-SVM)进行分类。实验在乳腺钼靶摄影数字化数据库(DDSM)上进行,实验结果表明,该方法克服了结构紊乱样本数量少的问题,提高了结构紊乱的识别率。Architectural distortion detection and recognition in mammograms are difficult problems, as it is hard to separate archi-tectural distortion from normal mammary gland, and the sample size is too small to learn a classifier with good performance. To overcome these difficulties, a transfer learning method was proposed for architectural distortion recognition, which combined the Gabor based spiculate patterns features, GLCM features and the Entropy features. In the source domain, the malignant masses were used as the positive samples and false positive samples of architectural distortion detection were used as the negative sam-ples ,and features were extracted from them. In the target domain, the true and false architectural distortion was used as training and testing set to extract features. An adaptive support vector machines （A-SVM） classifier was trained to perform the classifi-cation of a ROI. Images in DDSM （digital database for screening mammography） were used for experiments. Results show that the proposed method can overcome the problem of insufficient samples of architectural distortion and improve the recognition ac-curacy.

关 键 词：结构紊乱 迁移学习 自适应支持向量机 计算机辅助检测 乳腺癌 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]