基于多参数磁共振的影像组学模型预测前列腺癌Ki67的表达

Prediction of Ki67 expression in prostate cancer by radiomics models based on multiparameter magnetic resonance imaging

目的基于多参数磁共振建立影像组学模型预测前列腺癌Ki67的表达。方法回顾性分析我院2020年12月1日~2023年6月30日经病理证实、磁共振数据完成的PCa患者176例,按照7:3的比例分配到训练组(n=140)及验证组(n=36)。从PACS工作站中导出患者T2加权成像、T2加权脂肪抑制成像、小视野扩散加权成像、表观扩散系数的DICOM图像,在4个序列图像中勾画出病变区域的三维感兴趣区,并提取其中的影像组学特征,使用Spearman相关系数和LASSO回归对特征进行降维和选择,利用筛选出的组学特征建立影像组学模型。利用绘制ROC曲线并计算曲线下面积(AUC),阐述模型的诊断意义,并通过验证组对诊断效能施行验证。结果共提取1834个影像组学特征,最终筛选得到20个组学特征与Ki67表达状态相关。在互相独立创建的8个影像组学模型中,分别为逻辑回归、支持向量机、K-近邻、随机森林、极度随机树、极致梯度提升、轻量级梯度提升机和多层感知机,此中轻量级梯度提升机模型效益最优,训练组的AUC值为0.948(95%CI:0.913~0.982),测试组的AUC值为0.832(95%CI:0.698~0.967)。结论基于多参数磁共振构建的影像组学模型可以预测Ki67表达状况,且轻量级梯度提升机模型最好。

Objective To establish a radiomics model to predict Ki67 expression of prostate cancer on multiparameter magnetic resonance imaging(mp-MRI).Methods A total of 176 prostate cancer patients confirmed by postoperative pathology in our hospital from December 1,2020 to June 30,2023 with complete magnetic resonance date were rest retrospectively analyzed,and the patients were divided to the training group(n=140)and validation group(n=36)in a 7:3 ratio.The DICOM images of patients T2-weighted imaging(T2WI),fat-suppression T2-weighted imaging(FS-T2WI),zoomed imaging technique with parallel transmission diffusion weighted imaging(ZOOMit-DWI),the apparent diffusion coefficient(ADC)were exported from the PACS workstation,the three-dimension volume region of interest of the tumor was manually delineated on the four sequential images,and radiomics features were extracted,and the Spearman correlation analysis and LASSO analysis were used to single out the most valuable radiomic features.The radiomics models were built using the radiomics features The diagnostic value of the model was analyzed by using ROC curve and calculating the AUC,and the diagnostic efficacy was verified in the validation group.Results A total of 1834 radiomics features were extracted from T2WI,FS-T2WI,ZOOMit-DWI,ADC and 20 features were selected,which were related to Ki67 status.Among the eight radiomics models established for Logistic Regression,Support Vector Machine,K-Nearest Neighbor,RandomForest,ExtraTrees,eXtreme Gradient Boosting,Light Gradient Boosting Machine,Multilayer Perceptron,The Light Gradient Boosting Machine model was optimal with an AUC of 0.948(95%CI:0.913-0.982)in the training group and an AUC of the test group of 0.832(95%CI:0.698-0.967).Conclusion The radiomics models based on mp-MRI can noninvasively predict the expression of Ki67,and the LightGBM model is the best.