能谱CT影像组学非小细胞肺癌淋巴结转移的预测模型构建

Construction of a predictive model for lymph node metastasis innon-small cell lung cancer based on spectral CT radiomics

目的:构建基于能谱CT影像组学对非小细胞肺癌(NSCLC)术前淋巴结转移(LNM)的预测模型,并分析其诊断效能。方法:收集2019年1月—2023年1月在我院行能谱CT检查的NSCLC且接受肺癌切除术及淋巴结清扫术的153例患者的影像图像、肿瘤标记物及临床资料,按照7∶3比例分为训练集(n=107,其中LNM 37例,无LNM 70例)和验证集(n=46,其中LNM 15例,无LNM 31例),分别用于预测模型的训练和验证,比较两组患者一般临床资料及能谱CT参数,在CT平扫及增强图像上手动勾画肺癌病灶和目标淋巴结的感兴趣区(ROI)。应用人工智能软件自动化提取ROI的纹理参数,并从中筛选出能够鉴别LNM的纹理参数。利用LASSO回归筛选影像组学特征并建立影像组学标签,纳入多因素Logistic回归构建基于肿瘤组织及目标淋巴结纹理参数与影像组学特征的联合预测模型,采用受试者工作特征(ROC)曲线下面积(AUC)来评估能谱参数模型、影像组学模型及联合模型对术前LNM的诊断效能。运用DeLong检验对比各预测模型AUC的差异。通过决策曲线分析(DCA)对各预测模型的临床获益度进行评估。P<0.05为差异有统计学意义。结果:训练集和验证集中,LNM患者静脉期标准化碘浓度(NIC)低于无LNM患者,淋巴结短径高于无LNM患者(P<0.05);共提取207个影像组学特征,经LASSO回归筛选,最终纳入5个影像组学特征,包括灰度大小区域矩阵、灰度游程矩阵各2个,灰度依赖矩阵1个;多因素Logistic回归分析显示,淋巴结短径、静脉期NIC、Rad-score是预测NSCLC LNM的独立影响因素(P<0.05);ROC曲线分析显示,能谱参数模型预测训练集和验证集LNM的AUC分别为0.746、0.739,影像组学模型预测训练集和验证集LNM的AUC分别为0.747、0.726,联合模型预测训练集和验证集LNM的AUC分别为0.847、0.813,经Delong检验联合模型预测LNM的AUC高于能谱参数模型和影像组学模型预测LNM的AUC(P<0.05)。结论:基于能谱CT影像组学对NSCLC术前LNM具有较好的预测价值。

Objective:To construct a predictive model for lymph node metastasis(LNM)in non-small cell lung cancer(NSCLC)based on spectral CT radiomics,and to test its diagnostic efficacy.Methods:The imaging findings,tumor markers,and clinical data of 153 NSCLC patients who underwent spectral CT imaging examination,lung cancer resection surgery and lymph node dissection in our hospital from January 2019 to January 2023 were obtained.Patients were divided into training set(n=107)including 37 cases of LNM and 70 cases of non-LNM,and validation set(n=46)including 15 cases of LNM and 31 cases of non-LNM in a 7∶3 ratio.Data were used for training and verification of the models.The general clinical data and imaging features were compared between two sets.Regions of interest(ROI)of lung cancer lesions and target lymph nodes were delineated manually on CT plain and enhanced images.Artificial intelligence software was applied for extraction of texture parameters for ROIs and the screening of texture parameters capable of identifying LNMs.LASSO regression was used to screen radiomics features and establish radiomics labels.A joint predictive model based on tumour tissue and target lymph node texture CT parameters and radiomic features was constructed using the Multivariate Logistic regression algorithm.The area under the receiver operating characteristic(ROC)curve(AUC)was used to evaluate the diagnostic efficacy of spectral CT parameter model,radiomic signature model,and joint model for preoperative LNM.The DeLong test was used to compare the AUC differences of various models.Decision curve analysis(DCA)was used to evaluate the clinical benefits of the prediction model.A P<0.05 indicateed a statistically significant difference.Results:In both training set and validation set,the standardized iodine concentration(NIC)in the venous phase of LNM patients was lower than that of non LNM patients,and the lymph node shortaxis diameter was higher than that of non LNM patients(P<0.05).A total of 207 radiomic features were extracted from images.After LASSO regression screening, 5 radiomics features were finally included, including 2 grayscale size zone matrix features,2 grayscale run length matrix features, and 1 grayscale dependence matrix feature. Multivariate Logistic regression analysis denotedthat short-axis diameter of the lymph node, NIC of venous phase and Rad-score were independent factors in predictingLNM of NSCLC (P<0.05). ROC curve revealed that the areas under the curve (AUCs) for predicting the occurrence of LNM inthe training set and verification set were 0.746 and 0.739 in spectral CT parameter model, the AUC were 0.747 and 0.726 inradiomic signature model, and were 0.847 and 0.813 in joint model. Delong test verified that the AUC of joint model waslarger than that of spectral CT parameter model and radiomic signature model(P<0.05). Conclusion: Spectral CT radiomics hasgood predictive value for preoperative LNM in NSCLC.