基于MRI影像组学预测宫颈癌Ki-67表达水平

MRI Radiomic models for prediction of Ki-67 expression in cervical cancer

目的:开发并验证基于多参数MRI图像特征的影像组学特征预测模型对术前宫颈癌症患者的Ki-67指数状态的预测。方法:回顾性分析来自两个不同机构的91例宫颈癌患者的MRI影像及病理结果。根据术后免疫组化结果,将Ki-67指数分为高表达组(>60%)及低表达组(≤60%)。从每位患者的T_(2)/SPAIR、ADC和CE T_(1)WI图像中共提取3390个影像学特征。单变量分析和最小绝对收缩选择算子(LASSO)对影像组学特征进行降维处理,最终筛选出关键特征。采用Logistic回归、决策树、支持向量模型(SVM)方法构建模型。采用受试者操作特征(ROC)曲线分析影像组学特征的预测准确性,计算曲线下面积(AUC)。结果:91例患者中,27例Ki-67低表达,64例Ki-67高表达。最终从T_(2)/SPAIR、CE T_(1)WI、ADC图像中分别筛选出4、6、5个影像学特征。对Ki-67状态的预测,三个序列对应的模型构建方法为Logistic回归、SVM、Logistic模型,最终获得的训练组AUC分别为0.801、0.856、0.819;验证组AUC分别为0.716、0.731、0.719。结论:MRI影像学特征可作为一种无创方法评估Ki-67状态,为患者术前制定个体化治疗方案、化疗敏感性评估提供了重要的信息。

Objective:To develop and validate a radiomics model based on multiparameter MRI features to predict the Ki-67 index in cervical cancer patients.Methods:A total of 91 consecutive patients with cervical cancer from two centers were enrolled in the retrospective study.According to the results of postoperative immunohistochemistry,Ki-67 PI was divided into high expression group(>60%)and low expression group(≤60%).A total of 3,390 imaging features were extracted from T_(2)WI-SPAIR,ADC,and contrast enhanced(CE)T_(1)WI images.Univariate analysis and Least Absolute Shrinkage Selection Operator(LASSO)performed dimensionality reduction on the radiomic features to screen out important features.Models are constructed using logistic regression,decision trees,and support vector methods(SVM).The prediction accuracy of the radiomics signature was quantified by the receiver operating characteristics curve(ROC)of the training and validation groups.The area under curve(AUC)was calculated.Results:Among the 91 patients,27 had low Ki-67 expression and 64 had high Ki-67 expression.Through radiometric feature selection,4,6 and 5 features were finally selected based on T_(2)/SPAIR,CET_(1)WI and ADC images.For the prediction of Ki-67 status,3 sequence images of T_(2)/SPAIR,CE T_(1)WI,and ADC were used respectively.The model construction methods corresponding to each sequence were logistic regression,SVM,and logistic regression model.The final AUC of the training group was 0.801,0.856 and 0.819,respectively;The AUC of validation group was 0.716,0.731 and 0.719,respectively.Conclusion:MRI radiomic signature can be used as a non-invasive method to assess Ki-67 status,providing important information for individualized treatment plan and evaluate chemotherapy sensitivity before surgery.