融合多模态数据的机器学习模型在鉴别胆囊肿瘤性息肉中的价值-东南大学学报医学版

融合多模态数据的机器学习模型在鉴别胆囊肿瘤性息肉中的价值

单位：成都市中铁二局集团医院超声科, 四川成都 610031

分类号：R657.4;R195.1

出版年·卷·期（页码）：2026·45·第三期（406-417）

摘要：

目的: 探讨融合超声影像特征与血清肿瘤标志物的多模态机器学习模型在鉴别胆囊肿瘤性息肉中的诊断效能及临床应用价值。方法: 回顾性选取2021年9月至2024年9月期间,于成都市中铁二局集团医院行手术治疗的311例胆囊息肉样病变(GPLs)患者为研究对象,其中肿瘤性息肉组93例,非肿瘤性息肉组218例。采用R软件将所有数据集按7∶3比例随机划分为训练集(n=217)和测试集(n=94)。基于训练集采用逻辑回归(LR)、极端梯度提升(XGBoost)和随机森林(RF)构建胆囊肿瘤性息肉的鉴别模型。基于测试集采用曲线下面积(AUC)进行评价,并采用决策曲线分析(DCA)评估最优模型的临床净收益。最后,应用沙普利加和解释法(SHAP)对最优模型的预测结果进行可视化解释。结果: 模型性能评估结果显示,基于息肉的蒂结构(宽基底)、癌胚抗原(CEA)、糖类抗原19-9(CA19-9)构建的XGBoost算法模型在测试集上对胆囊肿瘤性息肉鉴别的AUC为0.947 0,优于LR(AUC=0.749 7)及RF模型(AUC=0.770 1)。XGBoost模型训练集在阈值0.1时获得最大净收益(0.281 6),测试集在阈值为0.2时获得最大净收益(0.271 3)。SHAP可解释性分析显示,CA19-9、息肉的蒂结构(宽基底)及CEA依次是影响模型预测的重要特征。结论: XGBoost模型在鉴别胆囊肿瘤性息肉方面具有最优的判别效能与临床实用性,且CA19-9、息肉的蒂结构(宽基底)及CEA依次是其预测的关键因子。

Objective: To evaluate the diagnostic performance and clinical utility of a multimodal machine learning model integrating ultrasound imaging features and serum tumor markers for differentiating neoplastic gallbladder polyps(NGPs) from non-neoplastic lesions. Methods: A retrospective cohort of 311 patients with gallbladder polypoid lesions(GPLs) who underwent surgical resection at China Railway Erju Group Hospital(Chengdu) between September 2021 and September 2024 was included. Of these, 93 had histologically confirmed neoplastic polyps and 218 had non-neoplastic polyps. The dataset was randomly splitted into training set(n=217) and test set(n=94) in a 7∶3 ratio using R software. Three machine learning models—Logistic regression(LR), extreme gradient boosting(XGBoost), and random forest(RF) were developed on the training set to discriminate NGPs. Model performance was assessed on the test set using the area under the curve(AUC). Decision curve analysis(DCA) was employed to quantify the clinical net benefit of the best-performing model, and Shapley additive explanations(SHAP) were used to provide interpretable visualizations of its predictions. Results: The XGBoost model incorporating three key predictors—broad-based stalk morphology, carcinoembryonic antigen(CEA), and carbohydrate antigen 19-9(CA19-9) achieved the highest diagnostic accuracy, with an AUC of 0.947 0 on the test set, significantly outperforming LR(AUC=0.749 7) and RF(AUC=0.770 1). DCA demonstrated that the XGBoost model yielded the greatest net clinical benefit at a probability threshold of 0.2 in the test set(net benefit=0.271 3), comparable to 0.281 6 at a threshold of 0.1 in the training set. SHAP analysis identified CA19-9, broad-based stalk morphology, and CEA as the top three features driving model predictions, in descending order of importance. Conclusion: The XGBoost-based multimodal model demonstrates superior discriminative ability and clinical applicability for identifying NGPs. Serum CA19-9, broad-based stalk morphology on ultrasound, and CEA are the most influential predictors.

参考文献：

[1] WANG K,XU Q,XIA L,et al.Gallbladder polypoid lesions:current practices and future prospects[J].Chin Med J,2024,137(14):1674-1683.
[2] KAMAYA A,FUNG C,SZPAKOWSKI J L,et al.Management of incidentally detected gallbladder polyps:society of radiologists in ultrasound consensus conference recommendations[J].Radiology,2022,305(2):277-289.
[3] ZHAO H,CHEN Y,ZHANG Z,et al.Ultrasound-guided percutaneous microwave ablation of gallbladder polyps:a case report[J].Medicine,2023,102(51):e36622.
[4] LIU K,LIN N,YOU Y,et al.Risk factors to discriminate neoplastic polypoid lesions of gallbladder:a large-scale case-series study[J].Asian J Surg,2021,44(12):1515-1519.
[5] WANG Y,PENG J,LIU K,et al.Preoperative prediction model for non-neoplastic and benign neoplastic polyps of the gallbladder[J].Eur J Surg Oncol,2024,50(2):107930.
[6] YUAN H X,WANG C,TANG C Y,et al.Differential diagnosis of gallbladder neoplastic polyps and cholesterol polyps with radiomics of dual modal ultrasound:a pilot study[J].BMC Med Imag,2023,23(1):26.
[7] 李文静,杨杰,陈冬梅.WGCNA与机器学习算法识别结直肠癌发病机制中潜在PANoptosis关键基因[J].东南大学学报(医学版),2023,42(5):688-696.
[8] 叶菊欣,徐邦红,孟慧,等.基于机器学习构建传染性单核细胞增多症患儿持续发热预测模型[J].现代医学,2026,54(2):191-200.
[9] MA N Q,LV H Y,BI J,et al.A scoring system for gallbladder polyps based on the cross-sectional area and patient characteristics[J].Asian J Surg,2022,45(1):332-338.
[10] KIM J S,LEE J K,KIM Y,et al.US characteristics for the prediction of neoplasm in gallbladder polyps 10 mm or larger[J].Eur Radiol,2016,26(4):1134-1140.
[11] LEVY A,POPOVICI T,BORIES P N.Tumor markers in pancreatic cystic fluids for diagnosis of malignant cysts[J].Int J Biol Markers,2017,32(3):291-296.
[12] QIU H,CHEN K,QIU Y,et al.Preoperative prediction of microvascular invasion in Intra-hepatic cholangiocarcinoma using serum tumor markers integrated with inflammatory and liver function indices[J].BMC Cancer,2025,26(1):27.
[13] DOU C,HAN Y,LIN L,et al.Development and validation of a comprehensive risk prediction model for polypoid lesions of the gallbladder[J].Clin Exp Pharmacol Physiol,2025,52(4):e70028.
[14] TSILIMIGRAS D I,CHATZIPANAGIOTOU O P,RUZZENENTE A,et al.Recurrence after gallbladder cancer resection:prognostic impact of CA19-9 and CEA levels at baseline and during surveillance[J].J Surg Oncol,2025,132(3):497-502.
[15] YOON S J,KIM S H,KIM H,et al.Development of an artificial intelligence-based prediction platform for early recurrence of resectable pancreatic cancer after curative surgery-toward future use as an indication for neoadjuvant treatment:a retrospective multicenter cohort study[J].Ann Surg Treat Res,2026,110(2):76.
[16] 苏明,赵欣.肿瘤标记物CA19-9、CEA、AFP和CA125在胆囊癌诊断中的应用价值[J].武警医学,2008,19(10):906-908.
[17] KIM E Y,HONG T H.Bile cholesterol and viscosity,the keys to discriminating adenomatous polyps from cholesterol polyps by a novel predictive scoring model[J].BMC Gastroenterol,2020,20(1):268.
[18] 李起,豆明辉,贾蓬勃,等.基于随机森林算法构建胆囊肿瘤性息肉发生的预测模型[J].中华肝脏外科手术学电子杂志,2025,14(1):46-52.

服务与反馈：

【文章下载】【发表评论】【查看评论】【加入收藏】

提示：您还未登录，请登录！点此登录