基于CT影像组学的深度学习模型预测胃癌隐匿性腹膜转移的价值  

The value of deep learning models based on CT radiomics in predicting occult peritoneal metastasis of gastric cancer

作　　者：牛冉冉 陈基明 任梦婷 姚琪 NIU Ran-ran;CHEN Ji-ming;REN Meng-ting(Medical Imaging Central,the First Affiliated Hospital of Wannan Medical College,Anhui 241001,China)

机构地区：[1]皖南医学院第一附属医院影像中心,安徽芜湖241001

出　　处：《放射学实践》2025年第4期485-492,共8页Radiologic Practice

摘　　要：目的:探讨基于CT影像组学构建的逻辑回归模型及深度学习模型预测胃癌隐匿性腹膜转移(OPM)的价值。方法:回顾性将2016年1月-2023年8月在本院经组织病理学检查证实的133例胃癌患者(OPM组68例,非OPM组65例)纳入本研究,并按照7∶3的比例随机分为训练集(n=94)和验证集(n=39)。所有患者术前行腹部CT平扫及多期增强扫描。基于静脉期CT增强图像,分别对肿瘤和腹膜下脂肪组织(SAT)逐层手动勾画ROI并融合成相应容积感兴趣区(VOI),采用Python软件提取手工影像组学(HCR)特征和深度学习影像组学(DLR)特征。然后,依次对肿瘤和SAT的HCR、DLR及HCR-DLR特征进行降维、建立影像组学标签并构建模型。采用多因素logistic回归分析基于组间比较P<0.05的临床资料、CT特征构建临床-CT征象模型,并基于临床-CT征象模型及表现最优的影像组学标签(HCR-DLR标签)分别构建肿瘤和SAT的联合模型,绘制最优联合模型(SAT)的列线图。利用受试者工作特征(ROC)曲线评价模型的预测效能,应用决策曲线分析(DCA)评估模型的临床应用价值。结果:临床-CT征象模型、肿瘤-HCR模型、肿瘤-DLR模型、肿瘤-HCR-DLR模型、SAT-HCR模型、SAT-DLR模型、SAT-HCR-DLR模型和SAT联合模型在训练集中的AUC分别为0.78(95%CI:0.68~0.87)、0.88(95%CI:0.82~0.95)、0.90(95%CI:0.84~0.96)、0.92(95%CI:0.87~0.98)、0.88(95%CI:0.81~0.95)、0.91(95%CI:0.85~0.96)、0.92(95%CI:0.87~0.97)和0.94(95%CI:0.89~0.98);在验证集中的AUC分别为0.77(95%CI:0.62~0.93)、0.83(95%CI:0.69~0.97)、0.88(95%CI:0.78~0.99)、0.89(95%CI:0.78~1.00)、0.84(95%CI:0.71~0.97)、0.86(95%CI:0.74~0.98)、0.88(95%CI:0.76~1.00)和0.89(95%CI:0.78~0.99)。DCA显示,SAT联合模型的临床获益高于临床-CT征象模型。结论:基于肿瘤和SAT的CT影像组学模型对胃癌患者隐匿性腹膜转移均具有较高的预测效能;除HCR模型外的各种影像组学模型的预测效能均显著优于临床-CT征象模型,以SAT�Objective:The purpose of this study was to investigate the value of logistic regression model and deep learning model based on CT radiomics in predicting occult peritoneal metastasis(OPM)of gastric cancer.Methods:133 patients with gastric cancer confirmed by histopathology(68 cases of OPM group and 65 cases of non-OPM group)were retrospectively analyzed.All subjects underwent plain and multiphase enhanced CT scanning of the abdomen.Patients were randomly assigned to two groups,including the training set(n=94)and the validation set(n=39)at a ratio of 7∶3.Based on the venous phase of the enhanced CT images,regions of interest(ROIs)of the tumor and subperitoneal adipose tissue(SAT)were manually segmented slice by slice,and then merged into corresponding volumes of interest(VOIs),and python software was used to extract the hand-crafted radiomics(HCR)features and deep learning radiomics(DLR)features.Sequentially,dimensionality reduction was performed on the HCR,DLR and HCR-DLR features of the tumor and SAT,followed by the establishment of radiomics signatures and the construction of models.The multivariate logistic regression analysis was performed to construct a clinical-CT signs model based on clinical data and CT features with P<0.05 from intergroup comparison.The multivariate logistic regression analysis was also performed to construct the combined models for the tumor and SAT based on the clinical-CT signs model and the optimal radiomics signature(HCR-DLR signature).A nomogram for the optimal combined model(SAT)was also developed.The predictive performance of the models was evaluated using receiver operating characteristic(ROC)curves,and decision curve analysis(DCA)was used to assess the clinical applicability of the models.Results:In the training set,the AUCs of the clinical-CT signs model,the tumor-HCR model,the tumor-DLR model,the tumor-HCR-DLR model,the SAT-HCR model,the SAT-DLR model,the SAT-HCR-DLR model,and the combined model of SAT were 0.78(95%CI:0.68~0.87),0.88(95%CI:0.82~0.95),0.90(95%CI:0.84~0.96),0

关 键 词：胃肿瘤 腹膜转移 腹膜下脂肪组织 影像组学 深度学习 

分 类 号：R814.42[医药卫生—影像医学与核医学] R735.2[医药卫生—放射医学]