CT影像组学在胃肠道间质瘤危险度分级预测中的价值

The Value of CT Radiomics in Risk Grading Prediction of Gastrointestinal Stromal Tumors

目的探讨基于CT影像组学在术前预测胃肠道间质瘤(GIST)危险度分级的价值。方法回顾性搜集经手术切除病理确诊为GIST且具有完整术前CT增强扫描图像及临床资料的127例患者,依据病理结果将其分为低危组(极低风险和低风险)和高危组(中风险和高风险)。使用达尔文平台在每例患者的增强CT动脉期图像上提取影像组学特征,然后以7∶3的比例随机分为训练组(n=88)和验证组(n=39)。通过最小绝对收缩降维和选择算子算法(LASSO),利用Logistic回归方法建立预测模型。诊断医师于PACS系统中观察两组病例的主观CT特征和临床资料,利用统计显著性特征构建预测模型。采用受试者工作特征(ROC)曲线评价模型的预测效能。结果从CT动脉期图像中提取的939个影像组学特征参数中经筛选最终选择5个特征参数,构建影像组学预测模型。对于训练组,曲线下面积(AUC)为0.923(95%CI:0.846~0.969);验证组的AUC为0.917(95%CI:0.783~0.981)。两组在肿瘤最大径和肿瘤形态方面存在显著差异,用于建立临床-CT征象预测模型,训练组的AUC为0.865(95%CI:0.775~0.928);验证组的AUC为0.887(95%CI:0.745~0.966)。结论基于CT增强图像摄取影像组学特征构建的预测模型比临床-CT征象模型能更好地预测术前GIST的风险等级,可作为指导术前临床决策的有效工具。

Objective To evaluate the value of CT radiomics in preoperative prediction of risk grade of gastrointestinal stromal tumor.Methods A retrospective collection of 127 patients diagnosed with GIST by surgical resection and pathological diagnosis in our hospital with complete preoperative CT enhanced scan images and clinical data were included in the study,and they were divided into low-risk group(very low and low risk)and high-risk group(medium and high risk)according to pathological results.The Darwinian platform was used to extract radiologic features from each patient’s enhanced CT arterial phase images.Then they were randomly divided into the training group(n=88)and the verification group(n=39)in a ratio of 7∶3.The minimum absolute shrinkage reduction and selection operator algorithm(LASSO)were used to establish the prediction model by Logistic regression method.Diagnostic physicians observed subjective CT features and clinical data of the two groups in the PACS system,statistical significance features were used to construct a prediction model.ROC curve was used to evaluate the prediction efficiency of the model.Results Five feature parameters were selected from 939 feature parameters extracted from CT arterial phase images to construct a radiomic prediction model.For the training group,the area under curve(AUC)was 0.923(95%CI:0.846-0.969).The corresponding value for the validation group was 0.917(95%CI:0.783-0.981).There were significant differences in tumor maximum diameter and tumor morphology between the two groups,which could be used to establish a clinical-CT prediction model.The area under curve(AUC)of the training group was 0.865(95%CI:0.775-0.928).The corresponding value for the validation group was 0.887(95%CI:0.745-0.966).Conclusion Compared with clinical-CT model,the predictive model constructed based on the radiologic features of enhanced CT image intake can better predict the risk level of preoperative gastrointestinal stromal tumor,and can be used as an effective tool to guide preoperative clinical decision-making.