基于CT影像组学模型鉴别诊断小细胞肺癌与非小细胞肺癌

Radiomics model based on CT for differential diagnosis of small cell lung cancer and non-small cell lung cancer

目的观察基于CT影像组学模型鉴别诊断小细胞肺癌(SCLC)与非小细胞肺癌(NSCLC)的效能。方法回顾性分析1524例经手术病理确诊的肺癌患者,其中526例SCLC(SCLC组),998例NSCLC(NSCLC组)。采用特征提取软件MaZda(Version 4.6)提取CT图像中病灶最大层面的纹理特征参数,以Correlation相关性分析和最小绝对收缩和选择算子(LASSO)算法对数据进行降维,筛选组间差异明显的影像组学特征,构建影像组学模型。以7∶3比例将全部患者分为训练集和验证集,采用7种机器学习模型,包括Logistic回归、随机森林(RF)、贝叶斯算法(NB)、决策树(DT)、卷积神经网络(CNN)、邻近算法(KNN)和支持向量机(SVM)模型对数据集进行处理,根据其在验证集的准确率选择最佳分类器模型,采用受试者工作特征(ROC)曲线分析该分类器模型对SCLC与NSCLC的鉴别诊断效能。结果针对每个病灶提取306个纹理特征参数,最终筛选出20个组间差异明显的影像组学特征,并以之构建预测模型。模型训练结果显示,KNN模型鉴别诊断SCLC与NSCLC的准确率最高,其在训练集的AUC为0.88、准确率81.34%、特异度97.00%、敏感度51.63%,在验证集的AUC为0.82、准确率78.82%、特异度95.00%、敏感度48.10%。结论基于CT影像组学结合机器学习算法建立的诊断模型可用于鉴别SCLC与NSCLC,以KNN模型的效能更优。

Objective To observe the efficiency of radiomics model based on CT for differential diagnosis of small cell lung cancer(SCLC)and non-small cell lung cancer(NSCLC).Methods Totally 1524 patients with postoperative pathological confirmed lung cancer were retrospectively analyzed,including 526 SCLC patients(SCLC group)and 998 NSCLC patients(NSCLC group).MaZda(Version 4.6)was used to extract texture feature parameters of the largest level of each lesion on CT images,and the radiomics features were downscaled with Correlation and least absolute shrinkage and selection operator(LASSO)algorithm,features being obvious different between groups were retained,then the optimal radiomics features were selected to build predictive model.The data were divided into training set and validation set at the ratio of 7∶3.Seven machine learning algorithms,including Logistic regression,random forest(RF),Bayesian algorithm(NB),decision tree(DT),convolutional neural network(CNN),K-neighborhood algorithm(KNN)and support vector machine(SVM)models were used to classify the set.The best classifier model was selected according to the accuracy of the validation set,and the receiver operating characteristic(ROC)curve was drawn to analyze the differential diagnostic efficacy of the classifier for SCLC and NSCLC.Results A total of 306 texture feature parameters were extracted.Among radiomics features being significantly different between groups,20 optimal radiomics features were retained to construct the prediction model.KNN was the best classifier in training set.The area under the curve(AUC)of the established predictive model for predicting SCLC and NSCLC in training set was 0.88,with accuracy of 81.34%,specificity and sensitivity of 97.00%and 51.63%,respectively,while the AUC in validation set was 0.82,with accuracy of 78.82%,specificity and sensitivity of 95.00%and 48.10%,respectively.Conclusion Radiomics model based on CT could effectively predict SCLC and NSCLC,and KNN model had better performances than the others.