探讨基于磁共振表观扩散系数（ADC）图像的影像组学模型对鉴别前列腺癌（PCa）和前列腺增生（BPH）的诊断价值。

回顾性分析2018年12月至2019年5月在南通大学附属建湖医院，经手术病理证实且能确定病理分级的42例患者（PCa 21例、BPH 21例）的ADC图像。应用ITK-SNAP软件勾画兴趣区（ROI），将ADC图像导入Analysis-Kinetics分析软件，进行影像特征提取。采用Lasso回归分析进行特征降维。通过Lasso降维筛选出的特征和相应加权系数乘积的线性组合来建立鉴别PCa和BPH的模型，绘制ROC曲线评价模型鉴别PCa和BPH的预测效能。

共提取396个影像组学特征，通过特征筛选后最后筛选出5个影像组学特征。建模后影像组学特征对鉴别PCa和BPH具有较好的预测效能，预测模型在训练组中鉴别效能的曲线下面积、准确度、敏感度、特异度、阳性预测值和阴性预测值分别为0.89、90%、80%、100%、100%和83%；在验证组中的曲线下面积、准确度、敏感度、特异度、阳性预测值和阴性预测值分别为0.96、83%、83%、83%、83%和83%。

基于磁共振ADC图像的影像组学模型对诊断前列腺癌和前列腺增生具有一定价值。

To develop and validate an apparent diffusion coefficient (ADC)-based radiomics predictive model in distinguishing the prostate cancer from benign prostate hyperplasia.

The ADC imaging data of 42 patients (21 cases of prostate cancer, and 21 cases of benign prostate hyperplasia) with histologically confirmed from December 2018 to May 2019 were analyzed retrospectively. The software ITK-SNAP was used to draw the region of interest (ROI), and the radiomic features based on ADC map were generated automatically from Analysis-Kinetics(GE Healthcare). Lasso regression model was used for data dimension reduction. The linear combination of the features selected by Lasso dimensionality reduction screening and the corresponding weighted coefficient product was used to establish prediction model. The model performance was assessed with respect to discrimination using the area under the curve(AUC) of ROC analysis.

396 radiomics features were extracted automatically by software and 5 features were left after redundancy reduction step. Radiomics features after modeling have a good predictive effect for the identification of prostate cancer and benign prostate hyperplasia. The prediction model showed good discrimination in both primary dataset(AUC=0.89, 95%; Accuracy=90%, sensitivity=80%, specificity=100%, positive predictive value=100%, negative predictive value=0.83) and independent validation datset(AUC=0.96, 95%; sensitivity=83%, Accuracy=90%, specificity=83%, positive predictive value=83%, negative predictive value=83%).

The radiomics model could provide important reference in differentiation between prostate cancer and benign prostate hyperplasia.