基于Bayesian粗糙集和布谷鸟算法的肺部肿瘤高维特征选择算法

High-dimensional feature selection algorithm for lung cancer based on Bayesian rough set and cuckoo search

在高维特征选择过程中最优特征子集生成和分类器参数优化方面,提出一种基于贝叶斯粗糙集(BRS)、遗传算法(GA)和布谷鸟算法(CS)的两阶段优化高维特征选择算法。该算法首先分析3000例肺部肿瘤CT图像的形状、灰度和纹理特征,提取104维特征分量共同量化ROI;然后进行两阶段优化:(1)从全局相对增益函数的角度分析了属性重要度,结合属性约简长度和基因编码权值函数的加权和构造适应度函数,通过选择、交叉和变异等遗传操作生成最优特征子集,在不降低分类精确度的前提下降低特征维度;(2)利用CS对支持向量机(SVM)参数进行全局寻优;最后通过实验验证本文算法的可行性和有效性。实验结果表明,该算法有效提升了肺部肿瘤良恶性识别能力,降低了算法的时间复杂度。

In the high-dimensional feature selection process,the best feature subset generation and classifier parameter optimization were concerned.A two stage optimization algorithm for high-dimensional feature selection based on bayesian rough set(BRS),genetic algorithm(GA)and cuckoo search(CS)was proposed.Firstly,analyzed the shape,gray and texture features of 3000 lung tumor CT images,and extracted 104 dimensional feature components to jointly quantify ROI.Then the two phase optimization was carried out.(1)the attribute importance was analyzed from the global relative gain function,and the fitness function was constructed by combining the weight of the attribute reduction length,gene encoding weight function and the attribute importance.The best feature subset was generated through genetic operation of selection,crossover and mutation,and the feature dimension was reduced without reducing the accuracy of classification.(2)CS was used to optimize the parameters of support vector machine(SVM).Finally,the feasibility and effectiveness of the algorithm were verified by experiments.Experimental results show that the algorithm effectively improves the ability of identifying benign and malignant lung tumors and reduces the time complexity of the algorithm.