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

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.