Prediction of Ki-67 expression in gastrointestinal stromal tumors using radiomics of plain and multiphase contrast-enhanced CT

作者全名："Liu, Yun; He, ChangYin; Fang, Weidong; Peng, Li; Shi, Feng; Xia, Yuwei; Zhou, Qing; Zhang, Ronggui; Li, Chuanming"

作者地址："[Liu, Yun; He, ChangYin; Li, Chuanming] Chongqing Univ Cent Hosp, Chongqing Emergency Med Ctr, Med Imaging Dept, Chongqing, Peoples R China; [Fang, Weidong] Chongqing Med Univ, Affiliated Hosp 1, Dept Radiol, Chongqing, Peoples R China; [Peng, Li] Chongqing Univ Cent Hosp, Chongqing Emergency Med Ctr, Dept Pathol, Chongqing, Peoples R China; [Shi, Feng; Xia, Yuwei; Zhou, Qing] Shanghai United Imaging Intelligence Co Ltd, Shanghai, Peoples R China; [Zhang, Ronggui] Chongqing Univ Cent Hosp, Chongqing Emergency Med Ctr, Dept Urol, Chongqing, Peoples R China"

通信作者："Li, CM (通讯作者)，Chongqing Univ Cent Hosp, Chongqing Emergency Med Ctr, Med Imaging Dept, Chongqing, Peoples R China.; Zhang, RG (通讯作者)，Chongqing Univ Cent Hosp, Chongqing Emergency Med Ctr, Dept Urol, Chongqing, Peoples R China."

来源：EUROPEAN RADIOLOGY

ESI学科分类：CLINICAL MEDICINE

WOS号：WOS:001000489500004

JCR分区：Q1

影响因子：4.7

年份：2023

卷号：　

期号：　

开始页：　

结束页：　

文献类型：Article; Early Access

关键词：Gastrointestinal stromal tumors; Ki-67 antigen; Machine learning; Tomography; x-ray computed

摘要："ObjectiveTo study the value of radiomics models based on plain and multiphase contrast-enhanced CT to predict Ki-67 expression in gastrointestinal stromal tumors (GISTs).MethodsA total of 215 patients with GISTs were retrospectively analyzed, including 150 patients in one hospital as the training set and 65 patients in another hospital as the external verification set. The tumor at the largest level of CT images was delineated as the region of interest (ROI). The maximum diameter of the ROI was defined as the tumor size. A total of 851 radiomics features were extracted from each ROI by 3D Slicer Radiomics. After dimensionality reduction, three machine learning classification algorithms including logistic regression (LR), random forest (RF), and support vector machine (SVM) were used for Ki-67 expression prediction. Using a multivariable logistic model, a nomogram was established to predict the expression of Ki-67 individually.ResultsDelong tests showed that the SVM models had the highest accuracy in the arterial phase (Z value 0.217-1.139) and venous phase (Z value 0.022-1.396). For the plain phase, LR and SVM models had the highest accuracy (Z value 0.874-1.824, 1.139-1.763). For the delayed phase, LR models had the highest accuracy (Z value 0.056-1.824). For the combined phase, RF models had the highest accuracy (Z value 0.232-1.978). There was no significant difference among the above models for KI-67 expression prediction (Z value 0.022-1.978). A nomogram was developed with a C-index of 0.913 (95% CI, 0.878 to 0.956).ConclusionsRadiomics of both plain and enhanced CT images could accurately predict the expression of Ki-67 in GIST. For patients who were not suitable to use contrast agents, plain scan could be used as an alternative."

基金机构：Chongqing Natural Science Foundation of China [cstc2020jcyjmsxmX0044]

基金资助正文：This work has received funding by the Chongqing Natural Science Foundation of China to Chuanming Li (cstc2020jcyjmsxmX0044).