基于特征矢量化的肺结节特征选择算法*

针对肺结节良恶性分类模型中特征选择过程无法避免特征多样性不受破坏的问题，提出一种将肺结节特征矢量化处理的特征选择方法。首先，假设每个肺结节特征都是由数据、类型构成的一个矢量。然后，按照特征类型添加特征到相应的特征子集，并分别利用Relief算法评价特征、特征子集的分类重要性。最后，通过动态阈值的方式筛选得到优化后的特征子集。在150个肺结节样本的分类实验中，采用本文算法所取得的敏感性为94.7%、特异性为93.7%、虚警率为5.2%、受试者工作特性曲线下面积为97.3%。分析表明，本文算法几乎不破坏肺结节特征的多样性，能够显著提高肺结节良恶性分类的准确性。

Feature selection based on feature vectorization on computer tomography scan of pulmonary nodules

To solve the problem of feature selection techniques by which the diversity of features was damaged in the process of distinguishing malignant pulmonary nodules from benign pulmonary nodules, this paper developed a new feature selection based on feature vectorization (FSBFV). Firstly, it assumed that each feature was a vector composing of its data and type. Secondly, it divided the feature space into several feature subsets according to their types, and then it applied Relief to evaluate the quality of features and feature subsets, as well as applied dynamic threshold to select those features with high quality and feature subsets with high quality. Finally, it combined selected feature subsets into the one. The classification experiment of 150 nodules was carried out. The sensibility, specificity, false alarm rate, and the area under the receiver operating characteristic curve based on the proposed method were 94.7%, 93.7%, 5.2%, and 97.3%, respectively. Further analysis indicates that the proposed method makes sure the diversity of features in the optimized set, and is helpful to improve the classification accuracy of benign and malignant pulmonary nodules.