The pseudo-label scheme in breast tumor classification based on BI-RADS features

Fan Zhang, Qinghua Huang, Xuelong Li

Abstract: The proposed method employs the Breast Imaging Reporting and Data System (BI-RADS) feature to classify the breast tumor. Compared with the ultrasound breast tumor classification methods based on the image, the “semantic gap” between the clinical feature and image feature is eliminated. In order to address the shortage of the labeled data, the pseudo-labeled scheme based on SVM is designed. The SVM classifier is trained by few labeled samples, and the hybrid dataset which contains the pseudo-labeled sample marked by SVM and few labeled samples is adopted to train the decision tree. 500 ultrasound breast tumor cases are collected to evaluate the proposed method. According to the result of the experiment, compared with the decision tree trained by the labeled dataset only, the accuracy of decision tree train by hybrid dataset improves 2.65%, the NPV improves 7.00%, and the Sensitivity increases 3.30%.