基于肺部PET/CT图像不同纹理特征的K最近邻分类器

摘要

Objective To reduce the dimension of the high-dimensional texture parameters of PET/CT images and to improve the accuracy of classification by building the K-nearest neighbor( KNN) classifier based on different texture features. Methods The study retrospectively collected 52 cases with pulmonary nodules who underwent 18F -FDG PET/CT from department of Nuclear Medicine, Xuanwu Hospital Capital Medical University. Co-occurrence matrix texture features were extracted from the contourlet transformed PET/CT images. Univariate analysis was applied first to reduce dimensionality of texture features according to c value before principle components analysis. Principal components of texturefeatures from selected texture features were extracted by PCA. We built the KNN classifier for original textures, selected textures and principle components respectively to distinguish benign and malignant nodules,comparing the efficacy of models based on the evaluation indices such as accuracy, sensitivity, specificity and AUC. Results 1344 original texture features were extracted from the region of interest of PET/CT images,from which 89 texture features were selected. Eleven principal components were extracted through the PCA procedure. The accuracy of KNN classifiers based on principal components,selected textures and original textures are 0. 614, 0. 579 and 0. 263 with AUC of 0. 645,0. 610,0. 515 respectively. Conclusions The KNN classifier based on the texture of principal components is the best one among the classifiers based on original texture features,the selected texture features through univariate analysis and the texture of principal components.%目的 对PET/CT图像高维纹理参数进行降维,基于不同纹理参数建立肺结节良恶性的K最近邻(K-nearest neighbor,KNN)分类器,探究最佳建模方法 ,提高分类的准确率.方法采用回顾性研究的方式,收集52例首都医科大学宣武医院核医学科肺结节患者的PET/CT图像,对图像的感兴趣区域基于Contourlet变换提取灰度共生矩阵的纹理参数.对肺结节PET/CT图像的纹理参数首先采用单因素分析的方法,根据ROC曲线下面积筛选纹理参数,再对其进行主成分分析提取主要成分.基于主成分、根据ROC曲线筛选的纹理及原始纹理分别采用K最近邻分类算法建立肺结节良恶性的分类器,通过正确率、灵敏度、特异度、阳性预测值(positive predictive value,PPV)、阴性预测值(negative predictive value,NPV)、ROC曲线下面积(area under curve,AUC)这些指标评价分类效果.结果 PET/CT图像共提取1344个原始纹理参数,经单因素分析后筛选出89个纹理参数,对筛选后的纹理共提取11个主成分.基于主成分、筛选纹理、原始纹理的分类模型正确率分别为0.614、0.579、0.263;AUC分别为0.645、0.610、0.515.结论 在主成分纹理、单因素分析筛选的纹理、原始纹理中,基于主成分纹理建立K最近邻分类器的效果最好.