基于多参数MRI及影像组学建立机器学习模型诊断临床显著性前列腺癌

Establishment of machine learning models for diagnosis of clinically significant prostate cancer based on multi-parameter MRI and radiomics

目的建立基于多参数MRI(mpMRI)和影像组学特征的机器学习模型,评价其诊断临床显著性前列腺癌(CSPC)的价值。方法结合纹理分析、MR动态增强定量分析、前列腺影像报告与数据系统(PI-RADS)评分和部分临床资料建立Logistic回归(LR)、逐步回归(SR)、经典决策树(cDT)、条件推断树(CIT)、随机森林(RF)和支持向量机(SVM)模型,运用ROC曲线和决策曲线分析法(DCA)评价上述模型和变量的重要性。结果验证组中RF模型诊断CSPC的AUC大于SVM、cDT、SR模型(P均<0.05),RF模型与LR、CIT模型诊断CSPC的AUC差异无统计学意义(P均>0.05),其余各模型间诊断CSPC的AUC差异无统计学意义(P均>0.05)。概率阈值为16%~91%时,RF模型的净获益最大,优于其他模型;概率阈值为23%~91%时,SVM模型的净获益仅次于RF模型而优于其他模型。前列腺特异性抗原密度(PSAD)和部分纹理分析参数的重要性较高。结论RF模型诊断CSPC优于其他模型,SVM模型次之。PSAD和纹理分析相关参数诊断CSPC的重要性高于PI-RADS评分和动态增强MRI定量参数。

Objective To establish machine learning models based on multi-parameter MRI(mpMRI)and radiomics features,and to investigate their value for diagnosis of clinically significant prostate cancer(CSPC).Methods Logistic regression(LR),stepwise regression(SR),classical decision tree(cDT),conditional inference tree(CIT),random forest(RF)and support vector machine(SVM)models were established with combining of texture analysis,dynamic contrast enhanced MRI(DCE-MRI),prostate imaging reporting and data system(PI-RADS)score and part of clinical data.ROC curve and decision curve analysis(DCA)were used to evaluate the models and the importance of variables.Results AUC of RF model for diagnosing CSPC in verification group was larger than that of SVM,cDT and SR model(all P<0.05).There was no statistically significant difference for diagnosing CSPC in AUC of RF model and LR,CIT model(P>0.05),nor of AUC for diagnosing CSPC among the other models in validation group(all P>0.05).When the probability threshold was 16%—91%,the net benefit of RF model was the largest,better than other models.When the probability threshold was 23%—91%,the net benefit of SVM model was second only to the RF model,but better than other models.Prostate specific antigen density(PSAD)and some texture analysis parameters were of high importance.Conclusion RF model is superior to other models in diagnosis of CSPC,SVM model comes second.PSAD and some texture analysis parameters are more important than PI-RADS score and DCE-MRI quantitative parameters for diagnosis of CSPC.