摘要
目的:建立临床-CT影像组学列线图模型并验证其在术前预测胃肠道间质瘤(GIST)危险度分级的应用价值。方法:回顾性搜集266例经病理证实的GIST患者,并将其分为低恶性风险组和高恶性风险组。提取10个临床及CT图像特征(年龄、性别、肿瘤部位、大小、形态、边界、囊变或坏死、钙化、表面溃疡、强化方式)用于构建临床模型。对病例的CT图像分别特征提取,并分为平扫期(N)、动脉期(A)、静脉期(V)及影像3期(N+A+V)4组,对上述4组通过select percentile和最小绝对收缩与选择算子(LASSO)算法降维、筛选组学特征后,分别使用5种分类器(Random Forest、Logistic Regression、SVM、SGD、XGBoost)建立各期组学模型。采用受试者工作特征(ROC)曲线下面积(AUC)进行量化。使用Delong检验比较各模型间AUC值差异,得到最佳影像组学模型。然后组合临床模型和最佳影像组学模型建立临床-CT影像组学列线图模型。结果:肿瘤大小、部位、形态、边界、强化方式及有无囊变坏死在两组中有统计学差异(P<0.05),年龄、性别、有无钙化及溃疡在两组中无统计学差异(P>0.05)。肿瘤大小和强化方式2个临床特征LASSO权重系数最高分别为0.08、0.20,临床模型基于XGBoost分类器诊断效能最高,测试组AUC值为0.91(95%CI:0.83~0.96),平扫期组学模型基于SGD分类器诊断效能最高,测试组AUC值为0.84(95%CI:0.75~0.92),动脉期组学模型基于SVM分类器诊断效能最高,测试AUC值为0.87(95%CI:0.80~0.94),静脉期组学模型基于Random Forest分类器诊断效能最高,AUC值为0.84(95%CI:0.75~0.93),影像3期组学模型基于XGBoost分类器诊断效能最高,AUC值为0.88(95%CI:0.79~0.94),基于XGBoost分类器构建的临床-CT影像3期(N+A+V)的联合模型最佳,训练组和测试组的AUC分别达0.98(95%CI:0.97~0.99)、0.95(95%CI:0.89~0.98)。结论:临床-CT影像组学列线图模型在术前预测GIST的危险度分级上具有较高价值,从�
Objective:To establish a clinical-CT radiomics nomogram model and stratify risk grade of gastrointestinal stromal tumors(GIST)preoperatively.Methods:266 patients with GIST confirmed by pathology were retrospectively analyzed,and were divided into low-and high-risk group.Ten clinical and CT image features(age,gender,tumor location,size,contour,margin,cystic degeneration or necrosis,calcification,surface ulcer and enhancement mode)were included for the construction of clinical models.The CT images-based radiomics features were extracted and divided into four groups:plain scan phase(N),arterial phase(A),venous phase(V)and image phaseⅢ(N+A+V).After dimensionality reduction and radiomics feature selection by select percentile and least absolute shrinkage and selection operator(LASSO)algorithm,five classifiers(Random Forest、Logistic Regression、SVM、SGD、XGBoost)were used to establish the radiomics model of each period.The area under curve(AUC)of receiver operating characteristic(ROC)was used for quantification.Delong test was used to compare the differences of AUC values.Then the clinical model and the best radiomics model were combined to establish the nomogram model.Results:The tumor size,location,shape,boundary,enhancement mode and cystic necrosis were statistically different between two groups(P<0.05).The maximum LASSO weight coefficients of the two clinical features of tumor size and enhancement mode were 0.08 and 0.20,respectively.The clinical model based on XGBoost classifier had the highest diagnostic performance,with the AUC of 0.91(95%CI:0.83~0.96),and the model based on SGD classifier had the highest diagnostic performance,with the AUC value of 0.84(95%CI:0.75~0.92),and the arterial phase model based on SVM classifier had the highest diagnostic performance,with the AUC of 0.87(95%CI:0.80~0.94).The venous phase model based on Random Forest classifier had the highest diagnostic performance with the AUC value of 0.84(95%CI:0.75~0.93).The triple phase model based on XGBoost classifier had the highest dia
作者
贾济波
张万军
刘原庆
冯飞文
胡粟
胡春洪
JIA Ji-bo;ZHANG Wan-jun;LIU Yuan-qing(Department of Radiology,the Fist Affiliated Hospital of Soochow University,Jiangsu 215006,China)
出处
《放射学实践》
CSCD
北大核心
2023年第10期1269-1275,共7页
Radiologic Practice
