开发并验证术前基于瘤灶及瘤周的临床放射学列线图以预测肺结节病理浸润性  

Develop and Validate a Preoperative Clinical Radiological Nomogram Based on Intratumoral and Peritumoral CT Radiomics to Predict Pathology Invasiveness of Pulmonary Nodules

作　　者：夏志颖 刘子蔚 包陈政任 胡秋根 张榕 XIA Zhi-ying;LIU Zi-wei;BAOCHEN Zheng-ren;HU Qiu-gen;ZHANG Rong(Department of Radiology,Hainan Traditional Chinese Medicine Hospital,Haikou 570203,Hainan Province,China;Department of Radiology,Shunde Hospital of Southern Medical University(The First People's Hospital of Shunde),Foshan 528308,Guangdong Province,China;Department of Radiology,Shunde Hospital of Southern Medical University Affiliated Chencun Hospital(the First People's Hospital of Shunde Affiliated Chencun Hospital),Foshan 528313,Guangdong Province,China)

机构地区：[1]海南省中医院放射科,海南海口570203 [2]南方医科大学顺德医院(佛山市顺德区第一人民医院)放射科,广东佛山528308 [3]南方医科大学顺德医院附属陈村医院(佛山市顺德区第一人民医院附属陈村医院)放射科,广东佛山528313

出　　处：《中国CT和MRI杂志》2024年第6期61-64,共4页Chinese Journal of CT and MRI

基　　金：佛山市科技计划项目(2220001005383);南方医科大学顺德医院科研启动项目(SRSP2021021)。

摘　　要：目的 基于肺结节瘤灶及瘤周影像组学特征开发反向传播神经网络(BPNN)深度学习模型,结合CT影像学特征及临床信息,构建用于预测肺结节浸浸润性的列线图模型。方法 回顾性收集经手术切除的242例GGN患者,分为腺体前驱病变(AAH/AIS)和浸润性肺腺癌(MIA/IAC),比较两组患者影像征象、临床特征的差异性。从肺部CT图像勾画肺结节及瘤周靶区,通过单因素秩和检验及相关性分析,对影像组学特征进行筛选,采用反向传播神经网络深度学习算法构建预测模型,以受试者工作特征(ROC)曲线下面积(AUC)评估模型预测性能,在外部队列中对模型进行泛化性验证。结果 BPNN模型在训练组、内部验证组和外部验证组中的AUC分别为0.883(95%CI:0.830-0.929)、0.854(95%CI:0.786-0.909)和0.854(95%CI:0.786,0.909)。经过单因素及多因素分析得出月牙征、CT值、GGN长径是预测肺结节浸润性的独立危险因素(P<0.05),由此构建临床模型在训练组及验证组中的AUC分别为0.889(95%CI:0.835-0.934)、0.778 (95%CI:0.668-0.879)和0.9 01 (95%CI:0.856-0.940)。结合BPNN模型及临床模型构建联合模型,模型在训练组和验证组中的AUC分别为0.952(95%CI:0.920-0.977)、0.891 (95%CI:0.807-0.959)和0.939(95%CI:0.899-0.968)。相比其他两个模型,联合模型评估肺结节病理浸润性展示了更强的模型性能。结论 基于瘤灶及瘤周放射组学特征联合临床-放射学信息构建的列线图模型,表现出较好的预测性能,联合模型在外部队列中得到较好的泛化性验证,可辅助临床对肺结节诊疗提供参考意见。Objective To develop a backpropagation neural network(BPNN) deep learning model based on the focal and peritumoral imaging featu res of pulmonary nodules,and combine CT imaging features and clinical information to construct a nomogra m model for predicting the infiltration of pulmonary nodules.Methods242 cases of GGN were retrospectively collected and divided into glandular prodromal disease(AAH/AIS)and invasive lung adenoca rcinoma(MIA/IAC).The imaging signs and clinical features of the two groups were compared.Lung nodules and peritumoral target areas were characterized from lung CT images,and the image omics features were screened through single-factor rank sum test and correlation analysis.The back propagation neural network deep learning algorithm was used to construct a prediction model,and the predictive performance of the model was evaluated by area under receiver Operating characteristic(ROC) curve(AUC),and the model was generalized to verify in external validation cohort.Results The AUC of BPNN model in training cohort,internal validation cohort and external validation cohort were 0.883(95%CI:0.830-0.929),0.854(95%CI:0.786-0.909) and 0.854(95%CI:0.786-0.909),respectively.U nivariate and multiva riate analysis showed that crescent sign,CT value and GGN length diameter were independent risk factors for predicting pulmonary nodule invasion(P<0.05).The AUC of the constructed clinical model in training cohort,internal validation cohort and external validation cohort were 0.889(95%CI:0.835-0.934),0.778(95%CI:0.668-0.879) and 0.901(95%CI:0.856-0.940),respectively.Combined with the BPNN model and the clinical model,the AUC of the model in training cohort,internal validation cohort and external validation cohort were 0.952(95%CI:0.920-0.977),0.891(95%CI:0.807-0.959) and 0.939(95%CI:0.899-0.968),respectively.Compared with the other two models,the combined model demonstrated a stronger model performance in evaluating pulmonary nodule pathologic infiltration.Conclusion The nomogram model based on tumor foci and per

关 键 词：肺结节 反向神经传播网络 影像组学 深度学习 列线图 

分 类 号：R445[医药卫生—影像医学与核医学]