特征提取联合机器学习在肺腺癌与肺鳞癌病理分型诊断中的应用研究

Application of Feature Extraction Combined with Machine Learning in Pathological Typing Diagnosis of Lung Adenocarcinoma and Lung Squamous Cell Carcinoma

目的探讨基于CT影像特征提取结合机器学习在肺腺癌(Adenocarcinoma,ADC)与肺鳞状细胞癌(Squamous Cell Carcinoma,SCC)病理分型鉴别诊断中的应用价值。方法回顾性分析1013例经手术病理证实为肺ADC或SCC的患者资料,根据病理结果将肺ADC分为第1组(n=515),肺SCC分为第2组(n=498),比较两组患者的性别和年龄差异。采用特征提取软件MaZda(Version 4.6)提取病灶最大层面的纹理特征参数,通过Standardization的方式对数据进行标准化处理,随后采用Univariate_Logistic、LASSO和MultiVariate_Logistic算法对数据进行降维,保留2组间差异明显的图像纹理特征,用以构建和筛选最佳诊断模型。将数据集按7∶3的比例分为训练组和验证组,采用6种机器学习算法对数据集进行处理,并根据验证组的准确度、ROC曲线下面积(Area Under Curve,AUC)、特异度和敏感度择最佳分类器。结果共提取病灶最大层面纹理特征参数306个,其中特征值之间差异明显的图像纹理特征共有114个,最终保留16个最佳图像纹理特征以构建预测模型,Logistic回归模型在验证集测试中的准确度最高,为本研究的最佳分类器。该模型在训练组的具体参数为AUC 0.826,其准确率、特异度、敏感度分别为75.7%、72.7%、78.7%;在验证组的具体参数为AUC 0.817,其准确率、特异度、敏感度分别为74.9%、74.1%、75.6%。结论CT影像特征提取方法结合机器学习算法建立的分析诊断模型在肺ADC和肺SCC病理分型的预测中具有一定的研究价值。

Objective To explore the application value of feature extraction combined with machine learning in the differential diagnosis of pathological typing of lung adenocarcinoma(ADC)and lung squamous cell carcinoma(SCC).Methods The data of 1013 patients with lung ADC or SCC confirmed by surgery and pathology were retrospective analyzed.According to the pathological results,patients with lung ADC were divided into group 1(n=515)and patients with lung SCC were divided into group 2(n=498).The gender and age differences of patients between the two groups were compared.The feature extraction software MaZda(Version 4.6)was used to extract the texture feature parameters at the largest layer of the lesion on CT images,and the data was standardized by Standardization.The Univariate_Logistic,LASSO and MultiVariate_Logistic algorithms were used to reduce the dimensionality of the data,and the texture features with obvious differences between groups were retained to construct and screen the optimal machine learning model.The dataset was divided into training and validation groups in a ratio of 7∶3.A total of 6 kinds of machine learning algorithms were used to process the dataset,and the optimal classifier was selected according to the accuracy,area under the receiver operating characteristic curve,specificity,and sensitivity of the verification group.Results A total of 306 texture feature parameters were extracted from the largest layer of lesions,of which 114 image texture features were obvious differences between feature values.Then 16 optimal radiomics features were retained to construct the prediction model.Logistic regression model had the highest accuracy in the validation set test and was the optimal classifier.The specific parameters of the model in the training group were that AUC was 0.826,and its accuracy,specificity,and sensitivity were 75.7%,72.7%and 78.7%respectively.In the validation group,the AUC was 0.817 and the accuracy,specificity,and sensitivity were 74.9%,74.1%and 75.6%respectively.Conclusion The analytical diagnostic model established by feature extraction combined with machine learning has certain research value in the prediction of pathological types of lung ADC and SCC.