MSCT影像组学对BorrmannⅣ型胃癌与原发性胃淋巴瘤的鉴别诊断价值

The Value of MSCT Based Radiomics in Differential Diagnosis of BorrmannⅣGastric Cancer and Primary Gastric Lymphoma

【目的】探讨CT影像组学模型鉴别BorrmannⅣ型胃癌与原发性胃淋巴瘤的价值。【方法】回顾性收集2008年6月至2018年4月经手术或内镜病理确诊的BorrmannⅣ型胃癌或原发性胃淋巴瘤的病例共186例,其中BorrmannⅣ型胃癌132例,原发性胃淋巴瘤54例。采用计算机软件对CT动脉期、静脉期图像提取放射组学特征,再选择Lasso Logistic回归方法进行特征筛选,并建立CT影像组学模型,此外亦筛选CT主观征象构建CT主观征象模型,并同时利用CT主观征象及组学特征构建联合模型。应用受试者工作特征(ROC)曲线评估CT主观征象模型、组学模型以及联合模型的性能。【结果】浆膜亮线征、浆膜侧不规则结节状外突两个征象进入了CT主观征象模型,而在组学特征中分别选出9个静脉期特征、8个动脉期特征以及14个动静脉期组合特征与病灶相关,分别构建相应组学模型。鉴别BorrmannⅣ型胃癌与原发性胃淋巴瘤,CT主观征象模型的截断值为0.188时,曲线下面积(AUC)为0.846,敏感性为61.9%,特异性为81.7%,准确性为76.5%;组学模型中的动脉期、静脉期以及动静脉期的截断值分别为-0.315、-0.669和-0.858,AUC分别为0.864、0.955和0.890,敏感性分别为71.4%、95.2%和81.0%,特异性分别为85.0%、88.3%和80.0%,准确性分别为81.5%、90.1%和80.3%;联合模型中的动脉期、静脉期以及动静脉期的截断值分别为0.257、0.556和0.497,AUC分别为0.883、0.956和0.918,敏感性分别为71.4%、90.5%和71.4%,特异性为85.0%、93.3%和90.0%,准确性为81.5%、92.6%和85.2%。经统计检验,模型诊断效能:联合模型>组学模型>CT主观征象模型(P<0.001),CT静脉期图像对于鉴别诊断2种肿瘤效能更好。【结论】基于CT动、静脉期的影像组学模型能够较准确地鉴别BorrmannⅣ型胃癌与原发性胃淋巴瘤。

【Objective】To explore the predictive value of CT based radiomics model in differentiating Borrmann typeⅣgastric cancer(GC)from primary gastric lymphoma(PGL).【Methods】A total of 186 cases(Borrmann typeⅣGC:132;PGL:86)pathologically diagnosed by surgical resection and/or endoscopic biopsy were enrolled from June 2008 to April 2018 retrospectively.Radiomics features were extracted from CT arterial phase and venous phase images by comput⁃ed algorithm,and selected by least absolute shrinkage and selection operator(Lasso)logistic regression,and then the CT based radiomics models were established.CT subjective signs were reviewed to build CT subjective signs model,while CT subjective signs and radiomics signature were assembled to build combined model.The receiver operating characteristic(ROC)curve was used to evaluate the performance of CT subjective sign model,radiomics model and the combined model.【Results】Two signs(the bright line sign of serosa and the irregular nodular protrusion on the serosa side)were selected in⁃to the CT subjective sign model.Among the radiomics features,9 venous phase features,8 arterial phase features and 14 arteriovenous combination features related to tumor classification were selected,and the corresponding radiomics models were constructed respectively.When the cut-off value of CT subjective sign model was 0.188,the area under curve(AUC)was 0.846,the sensitivity was 61.9%,the specificity was 81.7%,and the accuracy was 76.5%.The cut-off values of arteri⁃al phase,venous phase and arteriovenous phase radiomics model were-0.315,-0.669 and-0.858,respectively,and the AUCs were 0.864,0.955 and 0.890,the sensitivity were 71.4%,95.2%and 81.0%,the specificity were 85.0%,88.3%and 80.0%,the accuracy were 81.5%,90.1%and 80.3%,respectively.The cut-off values of arterial phase,venous phase and arteriovenous phase in the combined model were 0.257,0.556 and 0.497,respectively,and the AUCs were 0.883,0.956 and 0.918,the sensitivity was 71.4%,90.5%and 71.4%,the specificity was 85.0%,93.3%and 90.0%and the ac⁃curacy were 81.5%,92.6%and 85.2%,respectively.The diagnostic performance of the models from high to low were the combined model,radiomics model and CT subjective finding model(P<0.001),and CT venous phase images were more effective in the differential diagnosis of the two tumors.【Conclusions】The radiomics model based on the arterial and ve⁃nous phases CT images could differentiate Borrmann typeⅣgastric carcinoma from primary gastric lymphoma effectively.