基于CT增强扫描的深度迁移学习特征联合传统影像组学特征术前预测结直肠癌脉管侵犯的价值  

Integrating deep transfer learning and conventional radiomics from contrast-enhanced CT for preoperative prediction of lymphovascular invasion in colorectal cancer

作　　者：郝慧婷 瞿航 赵义[1] 周怡 闫晓辉 王苇[1] HAO Huiting;QU Hang;ZHAO Yi;ZHOU Yi;YAN Xiaohui;WANG Wei(Department of Imaging,Affiliated Hospital of Yangzhou University,Yangzhou 225000,China;Dalian Medical University,Dalian 116000,China)

机构地区：[1]扬州大学附属医院影像科,江苏扬州225000 [2]大连医科大学,辽宁大连116000

出　　处：《中国中西医结合影像学杂志》2025年第2期232-238,共7页Chinese Imaging Journal of Integrated Traditional and Western Medicine

基　　金：扬州市“十三五”科教强卫工程重点专科(ZDXK20186)。

摘　　要：目的:探讨基于CT增强扫描的深度迁移学习特征和传统影像组学特征的联合模型术前预测结直肠癌患者脉管侵犯状态的应用价值。方法:回顾性收集323例经手术病理证实的结直肠癌患者,按8∶2的比例随机分成训练集258例和验证集65例。从静脉期CT图像中提取与脉管侵犯状态相关的传统影像组学特征及深度迁移学习特征,采用最小绝对收缩和选择算子(LASSO)算法进行特征选择,采用极限梯度提升算法(XGBoost)、光照梯度增强机(LightGBM)和梯度提升算法(GB)构建传统影像组学模型及联合特征预测模型。采用ROC曲线评价各预测模型的诊断效能。采用DeLong检验比较各模型的预测能力。采用决策曲线分析(DCA)评估模型的临床实用性。结果:验证集中传统影像组学XGBoost模型的AUC为0.576(95%CI0.434~0.718),LightGBM模型AUC为0.628(95%CI0.491~0.766),GB模型AUC为0.625(95%CI0.488~0.763)。验证集中,联合4个传统影像组学特征与2个深度迁移学习特征构建的XGBoost模型AUC为0.737(95%CI0.611~0.863),LightGBM模型AUC为0.692(95%CI0.563~0.820),GB模型AUC为0.645(95%CI0.508~0.783)。DeLong检验显示,联合深度迁移学习特征和传统影像组学特征的XGBoost模型、LightGBM模型与传统影像组学XGBoost模型AUC比较,差异均有统计学意义(均P<0.05)。DCA显示,联合深度迁移学习特征和传统影像组学特征模型更具临床实用性。结论:联合深度迁移学习特征和传统影像组学特征预测模型对结直肠癌脉管侵犯的术前预测效能良好。Objective:To develop and validate an integrated predictive model combining deep transfer learning(DTL)features with conventional radiomics from contrast-enhanced CT for preoperative assessment of lymphovascular invasion in colorectal cancer.Methods:This retrospective study analyzed 323 patients with colorectal cancer,randomly allocated into training(258 cases)and validation(65 cases)cohorts(8∶2 ratio).Venous-phase CT images were processed to extract conventional radiomics features and DTL signatures.Feature selection was performed using least absolute shrinkage and selection operator(LASSO)algorithm.XGBoost,LightGBM,GB were implemented to construct the conventional radiomics models and the integrated models combining radiomics and DTL features.ROC curve was used to evaluate the diagnostic efficiency of each model.DeLong test was used to compare the predictive ability of the models,and decision curve analysis(DCA)was used to evaluate the clinical applicability of the models.Results:In the validation cohort,the conventional radiomics models demonstrated moderate performance,the AUCs of XGBoost,LightGBM,and GB models were 0.576(95%CI 0.434—0.718),0.628(95%CI 0.491—0.766),and 0.625(95%CI 0.488—0.763),respectively.While the integrated models showed improved performance,the AUCs of XGBoost,LightGBM,and GB models were 0.737(95%CI 0.611—0.863),0.692(95%CI 0.563—0.820),0.645(95%CI 0.508—0.783),respectively.DeLong test showed that the AUCs had significant differences between XGBoost integrated model and XGBoost conventional model,and between LightGBM integrated model and XGBoost conventional model(both P<0.05).DCA showed that the integrated models were more clinically useful.Conclusion:The integrated model combining DTL and conventional radiomics features demonstrates superior diagnostic performance for preoperative prediction for lymphovascular invasion in colorectal cancer.

关 键 词：结直肠癌 脉管侵犯 影像组学 深度学习 迁移学习 体层摄影术 X线计算机 

分 类 号：R73[医药卫生—肿瘤]