前列腺癌弥散加权成像与非影像学指标结合的Logistic回归预测模型

Estimation of prostate cancer probability by logistic regression using diffusion-weighted imaging and non-imaging parameters

目的建立前列腺癌弥散加权成像（DWI）与非影像学指标结合的Logistic回归预测模型,评估患者前列腺癌发病的可能性,以便后续诊断方法的选择和分析。方法回顾性分析114例2012年1月—2013年6月在复旦大学附属肿瘤医院先后行DWI和经直肠超声引导下前列腺穿刺活检患者的资料,其中前列腺癌患者70例、非肿瘤患者44例。应用DWI影像评价患者前列腺癌发病率,按危险高低行前列腺影像报告和数据系统（PI-RADS）评分（由低到高依次为1~5分）。随机抽取一半患者分入建模组,其余分入验证组。分析建模组中良恶性病例年龄、前列腺特异性抗原（PSA）、游离PSA与总PSA比值（f/t PSA）、肛门指检（DRE）、PI-RADS分级等指标的差异,选取有统计学差异的指标为考察变量,以穿刺病理是否为前列腺癌为应变量,建立两项Loglstic回归模型。A：非影像学指标＋PI-RADS;B：只含非影像学指标。利用验证组病例数据建立Loglstic回归模型和PI-RADS预测概率的受试者工作特征（ROC）曲线,分析其相应的灵敏度、特异度,确定最佳界值。结果良恶性患者PSA、f/t PSA、DRE及PI-RADS分级有显著统计学意义（P〈0.05）,纳入考察变量。前列腺癌Logistic回归预测模型建立如下：A：Logit P=-6.18＋0.0006PSA＋1.73DRE-8.01 f/t PSA＋1.3 PI-RADS;B：Logit P=-0.095＋0.0467PSA＋1.88DRE-7.3959 f/t PSA。应用A模型预测验证组的穿刺结果,ROC曲线的曲线下面积（AUC）为0.902,明显高于单独使用DWI PI-RADS分级（0.835）和临床指标（0.751）。A、B模型最佳灵敏度和特异度分别为86.4%、80.0%,90.9%、62.9%;PI-RADS分级最佳灵敏度和特异度为63.6%和94.3%。结论联合PSA、f/t PSA、DRE及PI-RADS这4个指标,利用Logistic回归预测模型有助于评估个体患前列腺癌的可能性,准确率较高,为可疑前列腺癌患者行前列腺穿刺提供了更充分的依据。

Objective To develop a logistic regression model using diffusion-weighted imaging （DWI） and non-imaging parameters to predict the probability of prostate cancer and facilitate the selection and interpretation of subsequent diagnostic tests.Methods From January 2012 to June 2013, a total of 114 males with elevated prostate-specific antigen （PSA） who underwent DWI and prostate biopsy in Fudan University Shanghai Cancer Center were reviewed retrospectively.Prostate cancer was detected in 70 cases, and the other 44 proved benign.The DWI results were applied to a prostate imaging reporting and data system （PI-RADS） and scored on a five-point scale （1-5 from normal to malignant）. Half of the cases were randomly put into train sample and the other test sample. The patients’ age, PSA, free to total PSA ratio （f/t PSA） , digital rectum examination （DRE） and PI-RADS score were compared between the benign and prostate cancer patients.Logistic regression model was performed with the parameters which were significantly different between the two groups to estimate prostate cancer probability.The logistic regression was established with two protocols： A （non-imaging parameters ＋PI-RADS） and B （non-imaging parameters only）. The logistic regression and PI-RADS score were applied in the test group. The receiver operating characteristic （ROC） curve was constructed to assess the diagnostic efficiency, and relative sensitivity and specificity were calculated.Results Independent predictors of a positive biopsy result included PSA, f/t PSA, DRE and PI-RADS score （P〈0.05）.Predictive multivariate model was developed as A： LogitP=-6.18＋ 0.0006PSA＋1.73DRE-8.01 f/t PSA ＋ 1.3 PI-RADS; B： LogitP= -0.095＋ 0.0467PSA ＋1.88DRE -7.3959 f/t PSA. When applied to the test sample to predict the result of biopsy, the analysis of ROC curve showed a higher area under curve （AUC） of 0.902 than protocol B （0.751） and PI-RADS （0.835） （P〈0.05）. The protocol A revealed a sensitivity of 86.4% and a specificity of 80.0%. The protocol B revealed a sensitivity of 90.9% and a specificity of 62.9%. PI-RADS score showed a sensitivity of 63.6% and a specificity of 94.3%.Conclusion Incorporation of PI-RADS score, f/t PSA, PSA and DRE into logistic regression model significantly improved the prediction of prostate cancer, and may serve as an aid to reduce unnecessary prostate biopsies.