基于CT组学特征的肺部病变良恶性诊断模型构建研究

Construction of diagnostic model for benign and malignant lung lesions based on CT radiomics

目的:基于CT组学特征构建肺部病变良恶性诊断模型,以区分肺部病变的良、恶性。方法:回顾分析某院2014年1月至2016年1月行增强CT扫描的肺部病变患者的资料。首先对患者的CT图像进行标准化处理并重新采样,然后提取1029个组学特征,采用最小绝对收缩和选择方法对组学特征进行筛选;其次利用五折交叉验证将样本随机分为5个相等大小的子集,其中1个作为测试集,4个作为训练集;最后采用逻辑回归、随机森林和支持向量机方法构建肺部病变良恶性诊断模型。采用AUC值、准确率、敏感度和特异度指标评估模型的诊断性能,并采用Delong-test方法对3种诊断模型进行比较,得出最佳模型。结果:逻辑回归、随机森林和支持向量机方法构建的诊断模型平均AUC值分别为0.747、0.771和0.820,平均准确率分别为0.688、0.696和0.740,平均敏感度分别为0.690、0.691和0.740,平均特异度分别为0.685、0.704和0.740。在合并五折交叉验证后,逻辑回归、随机森林和支持向量机方法构建的诊断模型AUC值分别为0.740、0.762和0.790。结论:基于CT组学特征采用逻辑回归、随机森林和支持向量机3种方法构建的诊断模型在区分肺部病变良恶性方面均具有一定的准确性和预测能力,诊断能力相当,但支持向量机模型略优于其他2种模型。

Objective To construct a CT radiomics-based model for diagnosing benign and malignant lung lesions.Methods The lung lesion patients undergoing enhanced CT scans in some hospital from January 2014 to January 2016 were analyzed retrospectively.The CT images of the patients were standardized and resampled,and 1 029 radiomics features were extracted and then screened with the least absolute shrinkage and selection operator;the samples obtained were randomly divided into five equal-sized subsets using five-fold cross-validation method,of which one was used as the test set and four as the training sets;three diagnostic models for benign and malignant lung lesions were established with logistic regression,random forest and support vector machine methods respectively.The three models went through diagnostic performance evaluation with the indies of AUC value,accuracy,sensitivity and specificity,and then were compared to determine the optimal model by Delong-test method.Results The three models by logistic regression,random forest and support vector machine methods had the average AUC values being 0.747,0.771 and 0.820,the average accuracies being 0.688,0.696 and 0.740,the average sensitivities being 0.690,0.691 and 0.740 and the average specificities being 0.685,0.704 and 0.740,respectively.The AUC values of the models by logistic regression,random forest and support vector machine methods were 0.740,0.762 and 0.790respectively after the involvement of five-fold cross-validation method.Conclusion The CT radiomics-based diagnostic models built with logistic regression,random forest and support vector machine methods gain advantages in differentiating between benign and malignant lung lesions accurately and predictively.The support vector machine-based model outperforms slightly the other two models in spite of similar diagnosis abilities.