基于经直肠超声造影参数及临床资料建立机器学习模型诊断临床显著性前列腺癌

Establishment of a machine learning model for the diagnosis of clinically significant prostate cancer based on transrectal contrast-enhanced ultrasound parameters and clinical data

目的基于前列腺经直肠超声造影参数及临床相关资料建立机器学习模型,并探讨各模型诊断临床显著性前列腺癌的效率。方法回顾性分析重庆大学附属肿瘤医院2018年11月至2021年9月接受经直肠超声造影检查并行经直肠超声引导下穿刺活检的患者151例。采用VueBox软件绘制时间-强度曲线,定量分析上升时间、达峰时间、平均渡越时间、峰值强度、上升斜率等12个参数。将年龄、总前列腺特异抗原、游离前列腺特异抗原、游离前列腺特异抗原比值、体积、前列腺特异抗原密度及经直肠超声造影参数共18个特征参数,通过相关属性值及信息增益属性值进行特征分析及特征筛选。将筛选特征通过机器学习单一算法及集成算法进行模型训练及测试,后通过F1值及ROC曲线下面积(AUC)进行模型评价。结果相关属性值及信息增益属性值分别筛选出12个变量及5个变量建立机器学习模型,集成算法建立模型均优于单一算法,两种变量筛选方式基算法为决策树的Bagging集成算法模型AUC(0.810比0.789)及F1值(0.748比0.742)均为最高,其次AUC及F1值均依序为Logistic回归、支持向量机(SVM)。结论基于经直肠超声造影参数及临床资料,在决策树、SVM、Logistic回归及集成算法中,基算法为决策树的Bagging集成算法模型诊断临床显著性前列腺癌性能最优。

Objective To establish a machine learning model for the diagnosis of clinically significant prostate cancer based on transrectal contrast-enhanced ultrasound parameters and clinically relevant data.Methods A retrospective analysis was performed on 151 patients in Chongqing University Cancer Hospital who underwent transrectal contrast-enhanced ultrasonography and transrectal ultrasound-guided needle biopsy from November 2018 to September 2021.The time intensity curve was drawn using VueBox software and 12 parameters such as rise time,peak time,average transit time,peak intensity,and rising slope were quantitatively analyzed.Age,total prostate-specific antigen,free prostate-specific antigen,free prostate-specific antigen ratio,volume,prostate-specific antigen density,and transrectal contrast-enhanced ultrasonography parameters,a total of 18 characteristic parameters,were analyzed and screened through relevant attribute values and information gain attribute values.The screening features were trained and tested by the machine learning single algorithm and integrated algorithm,and then the model was evaluated by the F1 value and the area under the ROC curve(AUC).Results Using the related attribute value and the information gain attribute value,12 variables and 5 variables were screened out respectively to establish a machine learning model.The model established by the ensemble algorithm was better than the single algorithm.For the two variable selection methods,the AUC(0.810 vs 0.789)and F1 values(0.748 vs 0.742)of the Bagging ensemble algorithm model,which basic algorithm was decision tree,were the highest,followed by Logistic regression and support vector machine(SVM)in order of AUC and F1 values.Conclusions Based on transrectal contrast-enhanced ultrasound parameters and clinical data,the Bagging ensemble model based on decision tree has the best performance in diagnosing clinically significant prostate cancer.