基于增强CT影像组学预测非肌层浸润性膀胱癌的病理分级

Predicting the pathological grade of non-muscle-invasive bladder cancer based on enhanced CT radiomics

目的探讨基于增强CT影像组学建立的非肌层浸润性膀胱癌(non-muscle-invasivebladder cancer,NMIBC)病理分级预测模型的诊断价值。方法回顾性分析2015年1月至2018年12月嘉兴市第二医院病理确诊的81例NMIBC患者的临床资料,患者术前接受增强CT检查,收集其皮髓期和实质期影像资料,对膀胱肿瘤轮廓进行勾勒,提取一阶变量、纹理变量、形状特征、小波变换变量,总计1980个特征变量。采用最大相关最小冗余(max-relevance and min-redundancy,mRMR)算法与最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)算法进行特征选择,最后利用多种机器学习算法结合有意义的特征变量建立预测模型,用于比较其预测NMIBC病理分级的敏感度、特异性和准确率。结果运用mRMR联合LASSO筛选出19个特征变量,使用K近邻分类(K-nearest neighbor,KNN)、神经网络(neural networks,NNET)、随机森林(random forest,RF)、支持向量机(support vector machines,SVM)4种机器学习算法建立模型并验证。4种模型建立成功且结果相对一致,其中RF模型表现相对稳定,在验证集中准确率为91.4%。在测试集中准确率为70.0%。结论通过术前增强CT影像组学结合机器学习算法可精准预测NMIBC病理分级,对推动膀胱癌个性化治疗具有科学意义。

Objective To explore the diagnostic value of pathological grade prediction model of non-muscle-invasive bladder cancer(NMIBC)based on enhanced CT radiomics.Methods The clinical data of 81 patients with NMIBC who were pathologically diagnosed in Jiaxing Second Hospital from January 2015 to December 2018 were retrospectively analyzed.The patients underwent enhanced CT examination before surgery,and the image data of the cortex and medulla stage and parenchyma stage were collected.The contour of the bladder tumor was outlined,and first-order feature variables,texture variables,shape characteristics and wavelet transform variables were extracted,totaling 1980 feature variables.The max-relevance and min-redundancy(mRMR)algorithm and least absolute shrinkage and selection operator(LASSO)algorithm were used for feature selection.Finally,multiple machine learning algorithms were combined with meaningful feature variables to build a prediction model,which was used to compare the sensitivity,specificity and accuracy of predicting NMIBC pathological grade.Results mRMR and LASSO were used to screen out 19 characteristic variables,and K-nearest neighbor(KNN),neural networks(NNET),random forest(RF)and support vector machines(SVM)were used to established and verified the model.The four models were established successfully and the results were relatively consistent,among which the RF model was relatively stable,with an accuracy of 91.4%in the verification set.In the test set,the accuracy was 70.0%.Conclusion Preoperative enhanced CT radiomics combined with machine learning algorithm can accurately predict the pathological grade of NMIBC,and it is of scientific significance to promote personalized treatment of bladder cancer.