多示例学习在多ROI乳腺肿瘤分类中的应用.pdfVIP

第1 卷 第1 期 智能计算机与应用 Vol． 1 No． 1 2011 年6 月 INTELLIGENT COMPUTER AND APPLICATIONS Jun． 2011 多示例学习在多ROI 乳腺肿瘤分类中的应用 胡 聪，黄剑华，张英涛，唐降龙 （哈尔滨工业大学 计算机科学与技术学院，哈尔滨 150001） 摘 要：在传统超声乳腺肿瘤分类系统中， 提取算法的精确性很大程度上决定了最终分类效果的好坏，同时传统分类系统中每个样本 ROI 都要求只提取肿瘤ROI ，在样本被提取出多个ROI 的情况下传统分类算法适用性差。 本文提出了两种适合于本类图像的包结构构造方法，首 先将多示例学习算法引入到超声乳腺肿瘤分类问题中，从而解决了上述多ROI 分类问题。 通过实验验证，文中方法在此类问题中较传统超声 乳腺肿瘤分类方法分类效果有较大提高。 关键词：乳腺超声； 多示例学习；纹理特征； 分类 中图分类号：TP391．4 文献标识码： A 文章编号：2095－2163 （2011）01－0066－04 The Application of Multi－instance Learning in Multi－ROI Breast Tumor Classification HU Cong ，HUANG Jianhua ，ZHANG Yingtao ，TANG Xianglong （School of Computer Science ＆ Technology, Harbin Institute of Technology, Harbin 150001, China ） Abstract: In traditional ultrasound breast tumor classification system, ROI extraction algorithms largely determine the accuracy of the final classifi- cation result. And each example is required to extract only the tumor ROI in the traditional classification system. The applicability of the traditional classification algorithm is not good in the case of each example having several ROIs. In this paper, two bag constructors suitable for such kind of US images have been proposed. Then multi-instance learning algorithms are first introduced to the classification of breast tumors in US image, which ad- dress the above-mentioned classification of multi-ROI. Experiments show that classification results of the proposed method in such problems have greater improvement than that of the traditional ultrasonic classification of breast tumor. Key words: Breast