计算机辅助乳腺癌诊断中的非平衡学习技术

针对计算机辅助诊断(CAD)中学习算法处理非平衡数据时,分类器预测具有大类样本的分类误差小,而稀有类样本的分类误差大的倾向性分类问题,提出基于反向k近邻的欠采样新方法.通过去除大类样本集中的噪声及冗余样本、保留具有类别代表性且可靠的样本作为有效样本以此平衡训练样本集,解决了欠采样引起的类别信息的丢失问题.基于UCI Breast-cancer数据集的仿真实验结果表明,该方法解决了非平衡学习问题的有效性,进一步的横向评测对比显示该算法性能显著优于其他同类算法.

Learning algorithm with non-balanced data for computer-aided diagnosis of breast cancer

When the learning algorithm handles non-balanced data in the computer-aided diagnosis, the prediction result of classifier is undesirably biased. The classification error of the big samples is small, while the classification error of the small samples is great. A reverse k nearest neighbor subsampling method was proposed in order to address the non-balanced learning issue. By removing the noisy and redundant samples from the big samples, and keeping the representative and reliable samples as the effective samples, the balanced training samples was realized, and the problem of the loss of the class information resulted from the subsampling was solved. The simulation results with the Breast-cancer dataset in UCI Machine Learning Repository show the validity of the algorithm to deal with the learning problems for non-balanced data. The experimental results show that the algorithm obviously outperforms existing methods．