代价敏感特征选择和半监督学习相结合的乳腺癌辅助诊断

Breast Cancer Assistant Diagnosis by Combining Cost Sensitive Feature Selection with Semi-supervised Learning

乳腺癌在X光图像下的主要表现是肿块和微钙化点.传统的诊断方法一般假设从图像肿块和微钙化点中提取的特征是正确有效的,且采用监督分类器进行诊断.但在实际中,一方面不能完全保证所有被提取特征的正确性;另一方面,由于高昂的标记代价,导致大量样本无标记.针对上述两个问题,本文提出了一种新颖的诊断方法.一方面,为了消除特征冗余和选择出对分类有用的判别特征,提出改进的代价敏感选择性集成法用于选择特征;另一方面,为了利用未标记样本信息,设计了一致性协同学习半监督分类器.在公共乳腺癌数据库DDSM上的实验表明,所设计的乳腺癌辅助诊断方法与其他方法相比具有更好的诊断性能.

Masses and microcalcification clusters are the main characteristics in the digital mammography of breast cancer. It is traditionally thought that the features extracted from the masses and microcalcification clusters are always correct and effective, and therefore used for a supervised design of classifier to diagnose. In practice, however, one cannot necessarily promise effectiveness of the features. Furthermore, not all labels of the samples can be obtained due to the expensive labeling cost. In this paper, we design a novel diagnosis method for microcalcification clusters. The proposed method first uses an algorithm of modified cost sensitive selective ensemble （CSSE） to select the features that are most useful for classification and without redundant information. Then we design a semi-supervised consistent co-training （CoCo-Training） algorithm as a diagnosis classifier by taking sufficient advantage of the unlabeled samples. Experiments on the benchmark DDSM show that the proposed diagnose method outperforms others.