基于胸部增强CT影像组学模型用于胸腺瘤分类的研究

Contrast-enhanced computed tomography-based radiomics models for the risk categorization of thymoma

背景与目的:胸腺瘤术前分类对治疗决策很有帮助,但获取存在一定困难。本研究旨在建立基于增强对比计算机体层成像(contrast-enhanced computed tomography,CECT)影像组学的胸腺瘤风险分类训练模型,并验证其性能、可靠性和泛化能力。方法:本回顾性队列研究分析了在上海交通大学医学院附属胸科医院2008年1月—2017年12月接受胸腺瘤切除手术的患者(Masaoka-KogaⅠ~Ⅲ期)的临床资料。将患者随机分层为训练组(80%)和测试组(20%),将术后病理学检查结果作为金标准。低风险组织学类型包括A、AB和B1。高风险组织学类型包括B2和B3。手动分割术前CECT图像相关肿瘤区域并提取影像组学特征。使用最小绝对收敛和选择算子(least absolute shrinkage and selection operator,LASSO)回归进行特征选择,并将临床特征添加到联合模型中。模型性能指标包括受试者工作特征(receiveroperating characteristic,ROC)曲线的曲线下面积(area under curve,AUC)、灵敏度和特异度。结果:共纳入478例患者(平均年龄51.3±12.3岁,男性占48.1%)。临床模型、基于CECT的影像组学模型和在测试集上使用临床和CT特征的模型AUC分别为0.666、0.831和0.850。性能最佳的模型的灵敏度为0.829,特异度为0.764。结论:基于术前CECT的影像组学模型在胸腺瘤风险分类中表现良好。

Background and purpose:Preoperative risk categorization of thymoma is useful for treatment decisions but remains challenging.This study focused on training radiomics models using contrast-enhanced computed tomography(CECT)images for thymoma risk categorization and validating the model's performance,reliability and generalizability in a relatively large cohort.Methods:This retrospective cohort study analyzed the clinical data of thymoma patients(Masaoka KogaⅠ-Ⅲ)who underwent thymectomy surgery at the Affiliated Chest Hospital of Shanghai Jiao Tong University School of Medicine from January 2008 to December 2017.The cohort was divided into a training group(80%)and a test group(20%)using stratified random selection.The gold standard for histologic types was based on surgically resected specimens.Low-risk histologic types included A,AB and B1.High-risk histologic types included B2 and B3.Radiomics features were extracted from manually segmented regions of interest on preoperative CECT images.Interobserver correlation and least absolute shrinkage and selection operator(LASSO)regression were used for feature selection.Model performance metrics included area under the curve(AUC)of receiver operating characteristic(ROC)curve,sensitivity and specificity.Clinical characteristics were added to the combined model.Results:A total of 478 patients(mean age 51.3±12.3 years,48.1%was male)were included.The AUC of the clinical model,the CECT-based model,and the model using both clinical and CECT features on the test set were 0.666,0.831 and 0.850,respectively.The best performing model had a sensitivity of 0.829 and a specificity of 0.764.Conclusion:CECT-based radiomics models showed good performance in risk categorization of thymomas.