基于临床、影像组学开发并验证用于预测肺磨玻璃结节浸润性的Nomogram模型

Development and Validation a Nomogram Incorporating CT Radiomics Signatures and Radiological Features for Differentiating Invasive the Invasiveness of Pulmonary Ground Glass Nodules

目的探讨基于影像学征象、临床特征及临床影像组学联合模型预测术前GGN的病理浸润性的价值。方法回顾性收集经手术切除的219例GGN患者临床、影像及病理资料,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。提取所有GGN患者影像组学特征,采用弹性网络算法构建影像组学模型。通过单因素和逐步Logistic回归筛选临床重要危险因素,构建临床模型,并进一步结合影像组学模型构建联合模型。ROC曲线评估模型预测性能,采用Nomogram图可视化模型风险因素。结果AAH/AIS(n=99)组和MIA/IAC组(n=120)中月牙征、分叶征、毛刺征、胸膜牵拉征、GGN长度和CT值具有统计学意义(P<0.05)。影像组学模型在训练组和验证组中的AUC分别为0.806(95%CI:0.7350.874)、0.814(95%CI:0.7030.913)。月牙征、胸膜牵拉征和HU为临床独立危险因素,逐步回归分析得到最优临床模型,模型在训练组和验证组中的AUC分别为0.753(95%CI:0.6740.828)、0.742(95%CI:0.6210.851)。结合Radscore、月牙征、胸膜牵拉征和HU构建联合模型,模型在训练组和验证组中的AUC分别为0.843(95%CI:0.8040.898)、0.869(95%CI:0.8480.927),相比其他两个模型,联合模型评估GGN病理浸润性展示了更强的模型性能,此外利用Nomogram量化影像特征。结论基于临床特征及影像组学构建联合模型有助于术前无创预测GGN病理浸润性,Nomogram有助于可视化模型。

Objective To explore the value of imaging signs,clinical features and clinical radiomics combined model in predicting the pathological invasion of preoperative GGN.Methods The clinical,imaging and pathological data of 219 patients with GGN who underwent surgical resection were retrospectively collected and divided into glandular precursor lesions(AAH/AIS)and invasive lung adenocarcinoma(MIA/IAC).The differences of imaging signs and clinical features between the two groups were compared.The radiomics features of all GGN patients were extracted,and the radiomics model was constructed by elastic network algorithm.The clinical important risk factors were screened by single factor and stepwise logistic regression,and the clinical model was constructed,and the combined model was further combined with the radiomics model.ROC curve was used to evaluate the predictive performance of the model,and Nomogram was used to visualize the risk factors of the model.Results The crescent sign,lobulation sign,spiculation sign,pleural traction sign,GGN length and CT value in AA H/AIS(n=99)group and MIA/IA C group(n=120)were statistically significant(P<0.05).The AU C of the radiomics model in the training group and the validation group were 0.806(95%CI:0.7350.874)and 0.814(95%CI:0.7030.913),respectively.The crescent sign,pleural traction sign and HU were independent clinical risk factors.The optimal clinical model was obtained by stepwise regression analysis.The AU C of the model in the training group and the validation group were 0.753(95%CI:0.6740.828)and 0.742(95%CI:0.6210.851),respectively.Combined model was constructed by Radscore,crescent sign,pleural traction sign and HU.The AU C of the model in the training group and the validation group were 0.843(95%CI:0.8040.898)and 0.869(95%CI:0.8480.927),respectively.Compared with the other two models,the combined model showed stronger performance in evaluating GGN pathological invasion.In addition,Nomogram was used to quantify the image features.Conclusion The combined model based on clinical features and radiomics is helpful to predict the pathological invasion of GGN noninvasively before operation,and Nomogram is helpful to visualize the model.