基于ADC图的不同影像组学模型在预测局部进展期直肠癌新辅助放化疗疗效中的应用

Application of ADC imaging-based different radiomics models in predicting the efficacy of neoadjuvant chemoradiotherapy for locally advanced rectal cancer

目的:探究基于表观扩散系数(apparent diffusion coefficient,ADC)图不同影像组学模型预测局部进展期直肠癌新辅助放化疗(neo-adjuvant chemoradiotherapy,nCRT)疗效的有效性。方法:回顾并分析2017年3月—2019年5月43例非转移性局部进展期直肠癌患者的资料,患者均行nCRT,在治疗前后进行MRI检查。采用ITK-SNAP在治疗前ADC图上手动勾画感兴趣区(region of interest,ROI),通过影像组学软件Pyradiomics提取109个影像组学特征。43例患者采用配对差异分析方法(paired-difference analysis,PDA)增加样本量后共获得378个样本对,按照7∶3随机分成训练组和测试组。分别采用支持向量机(support vector machine,SVM)、自动编码器(auto-encoder,AE)、线性判别分类器(linear discriminant analysis,LDA)、随机森林(random forest,RF)、罗杰氏回归(logistic regression,LR)、LR-Lasso等模型进行筛选。根据模型在测试集上的准确率、灵敏度、特异度来决定一个模型的最优组合,使用受试者工作特征曲线分析评估不同模型的诊断性能。分析模型基于Sklearn和软件FeAture Explorer。结果:最终筛选出较为稳定的3个模型分别是SVM、RF及LR-Lasso模型,SVM模型的曲线下面积(area under curve,AUC)、准确率为0.934和98.4%,灵敏度和特异度为80.0%和100.0%,阴性预测值和阳性预测值为98.3%和100.0%。RF模型的AUC、准确率为0.998和98.4%,灵敏度和特异度为100.0%和98.3%,阴性预测值和阳性预测值为100%和83.2%。LR-Lasso模型的AUC、准确率为0.997和98.4%,灵敏度和特异度为100.0%和98.3%,阴性预测值和阳性预测值为100%和83.3%。结论:影像组学模型在预测局部进展期直肠癌疗效方面具有更高的准确率,采用RF方法建立的影像组学模型较其他组学模型诊断效能更高。

Objective:To investigate the effectiveness of predicting the efficacy of neo-adjuvant chemoradiotherapy(nCRT)for locally advanced rectal cancer based on different radiomics models based on apparent diffusion coefficient(ADC)map.Methods:A retrospective analysis of 43 non-metastatic locally advanced rectal cancer patients from Mar.2017 to May.2019 was carried.All patients underwent nCRT,and before and after treatment underwent MRI.ITK-SNAP was used to manually outline the region of interest(ROI)on the ADC map before treatment,and 109 imaging omics features were extracted by the imaging omics software Pyradiomics.Forty-three patients used paired-difference analysis(PDA)to increase the sample size,and a total of 378 sample pairs were obtained,which were randomly divided into training and test groups according to 7∶3.Support vector machine(SVM),auto-encoder(AE),linear discriminant analysis(LDA),random forest(RF),and logistic regression(LR)And LR-Lasso(logistic regression via Lasso).According to the accuracy,sensitivity,and specificity of the model on the test set,the optimal combination of a model is determined,and the receiver operating characteristic curve analysis is used to evaluate the diagnostic performance of different models.The analysis model was developed based on Sklearn and FeAture Explorer.Results:Finally,the three more stable models were screened:SVM,RF,and LR-Lasso models.The SVM model had an AUC,accuracy of 0.934 and 98.4%,sensitivity and specificity of 80.0%and 100.0%;negative predictive value and positive predictive value were 98.3%and 100.0%.The RF model had AUC,accuracy of 0.998 and 98.4%,sensitivity and specificity of 100.0%and 98.3%,negative predictive value and positive predictive value of 100.0%and 83.2%.The LR-Lasso model has AUC,accuracy of 0.997 and 98.4%,sensitivity and specificity of 100.0%and 98.3%,negative predictive value and positive predictive value of 100.0%and 83.3%.Conclusion:The Radiomics model has a higher accuracy in predicting the efficacy of locally advanced rectal cancer.The Radiomics model established by the RF method has higher diagnostic efficiency than other omics models.