基于增强CT影像组学联合语义特征的支持向量机可解释模型术前预测胃肠道间质瘤Ki-67表达的研究

A Preliminary Study of Preoperative Prediction of Ki-67 Expression in Gastrointestinal Stromal Tumors Using a Machine Learning Interpretable Model Integrating Enhanced CT Radiomics and Semantic Features

目的构建并验证基于增强CT影像组学特征联合传统CT特征的机器学习SHAP模型术前预测胃肠道间质瘤(GIST)患者Ki-67表达状态。方法回顾性搜集我院149例GIST患者临床、影像及病理资料。根据术后病理将患者分为Ki-67低表达组和高表达组。分析术前增强CT图像中的传统CT特征并从静脉期图像中提取影像组学特征。采用组内相关系数(ICCs)、最大相关最小冗余(MRMR)和最小绝对收缩和选择算子(LASSO)方法筛选影像组学特征并构建影像组学标签。然后采用SVM机器学习算法对影像组学特征并联合传统CT语义特征进行模型构建,以受试者工作特征(ROC)曲线评估机器学习模型对GIST患者Ki-67表达的预测效能,并使用SHAP方法分析并研究不同变量的贡献度及风险阈值。结果在训练集中和验证集中,Ki-67高表达和Ki-67低表达患者的Radscore分别为(5.50±8.27)vs(-2.16±5.56)和(2.15±1.71)vs(-3.43±6.90),差异均有统计学意义(P<0.001)。Radscore在预测GIST患者Ki-67表达在训练集和验证集中的AUC分别为0.749和0.729。联合影像组学特征和传统CT特征SVM分类模型显示训练集和验证集的AUC分别为0.812和0.791。SHAP分析结果显示Radscore和肿瘤直径对模型具有高度正贡献。个性化特征归因结果显示Radscore>-0.1175、肿瘤直径>5.5 cm对GIST患者Ki-67高表达的风险预测能力越大。结论基于增强CT影像组学特征和传统语义特征的可解释SVM模型可以术前个性化预测GIST患者Ki-67表达状态,可为临床个体化治疗决策提供可靠的影像学生物标志物。

Objective To construct and validate a machine learning SHAP model based on enhanced CT radiomics features combined with traditional CT features for preoperative prediction of Ki-67 expression status in GIST patients.Methods A retrospective collection of clinical,imaging,and pathological data was performed on 149 GIST patients in our hospital.Patients were divided into low expression and high expression groups based on postoperative pathology.Traditional CT features were analyzed from preoperative enhanced CT images,and radiomics features were extracted from the venous phase images.ICCs,MRMR,and LASSO methods were employed to select radiomics features and construct radiomics labels.Subsequently,the SVM machine learning algorithm was used to build a model incorporating radiomics features and statistically significant traditional CT features.The predictive performance of the machine learning model for Ki-67 expression in GIST patients was evaluated using ROC curves.The SHAP method was utilized to analyze and investigate the contribution and risk threshold of different variables.Results In the training set and validation set,the Radscores for high and low Ki-67 expression in GIST patients were(5.50±8.27)vs(-2.16±5.56)and(2.15±1.71)vs(-3.43±6.90),respectively,with statistically significant differences(P<0.001).The Radscore had AUCs of 0.749 and 0.729 for predicting Ki-67 expression in GIST patients in the training set and validation set,respectively.The SVM classification model integrating radiomics features and traditional CT features showed AUCs of 0.812 and 0.791 in the training set and validation set,respectively.The SHAP analysis results demonstrated that the Radscore and tumor diameter made significant positive contributions to the model.Personalized feature attribution results indicated that a Radscore>-0.1175 and tumor diameter>5.5 cm corresponded to a greater risk prediction ability for high Ki-67 expression in GIST patients.Conclusion An interpretable SVM model based on enhanced CT radiomics features and traditional semantic features can provide individualized preoperative prediction of Ki-67 expression in GIST patients,offering reliable imaging biomarkers for clinical personalized treatment decisions.