结合瘤周影像组学建立磨玻璃结节手术切除预测模型

Establishment of Predictive Model for Surgical Resection of Ground Glass Nodules Combined with Peritumoral Radiomics

目的 融合肺磨玻璃结节(GGN)瘤内及瘤周影像组学特征,并与临床模型相结合建立GGN手术切除预测模型。方法 回顾性搜集311例肺GGN患者CT图像,包括良性/腺体前驱病变121例,肺腺癌(微浸润腺癌/浸润性肺腺癌)190例。对GGN行手动分割获得瘤内ROI,使用膨胀算法外扩3 mm获得瘤周ROI,分别提取影像组学特征。按照7∶3比例随机划分训练集(217例)与验证集(94例),使用支持向量机构建瘤内组学模型、瘤周组学模型及融合组学模型。选取其中表现最好的模型与临床模型相结合,建立GGN手术切除预测模型。使用曲线下面积(AUC)、准确率、敏感度、特异度评价各模型预测效能,DeLong检验用于比较各模型AUC差异,使用决策曲线评估各模型的临床应用。结果 瘤内组学模型训练集AUC值为0.805(95%CI:0.745~0.866),验证集AUC值为0.787(95%CI:0.696~0.878);瘤周组学模型训练集AUC值为0.727(95%CI:0.655~0.799),验证集AUC值为0.759(95%CI:0.653~0.866);融合组学模型训练集AUC值为0.827 (95%CI:0.772~0.882),验证集的AUC值0.858(95%CI:0.777~0.939)。将融合组学模型与临床模型相结合建立的列线图模型训练集AUC值为0.840 (95%CI:0.788~0.892),验证集AUC值为0.877(95%CI:0.804~0.950)。DeLong检验示列线图模型预测效能高于瘤内组学模型,在验证集中差异有统计学意义。决策曲线分析表明,列线图模型整体净效益比最高。结论 结合瘤周影像组学建立的GGN手术切除预测模型可帮助临床医师把握手术节点,减少过度治疗的发生。

Objective To combine the intratumoral and peritumoral imaging features of lung GGN with clinical models to establish a prediction model for surgical resection of lung GGN.Methods CT images of 311 patients with lung GGN were retrospectively collected,including 121 cases of benign/glandular precursor lesions and 190 cases of lung adenocarcinoma(MIA/IAC).The intratumoral ROI was obtained by manual segmentation of GGN rows,and the peritumoral ROI was obtained by outward expansion of 3 mm using the expansion algorithm,and the radiomics features were extracted respectively.The training set(217 cases) and validation set(94 cases) were randomly divided according to a 7:3 ratio.Support vector machine was used to construct the intratumoral,peritumoral and fusion radiomics models.The best performing model was selected and combined with the clinical model to establish the prediction model of GGN surgical resection.The AUC,accuracy,sensitivity and specificity were used to evaluate the prediction performance of each model.DeLong test was used to compare the differences in AUC of each model,and the decision curve was used to evaluate the clinical application of each model.Results The AUC value of intratumoral radiomics model was 0.805(95% CI:0.745-0.866) in the training set and 0.787(95% CI:0.696-0.878) in the validation set.The AUC of peritumoral radiomics model was 0.727(95% CI:0.655-0.799) in the training set and 0.759(95% CI:0.653-0.866) in the validation set.The AUC value of the fusion radiomics model was 0.827(95% CI:0.772-0.882) in the training set and 0.858(95% CI:0.777-0.939)in the validation set.The fusion radiomics model combined with the clinical model established a nomogram with an AUC value of 0.840(95% CI:0.788-0.892) in the training set and 0.877(95% CI:0.804-0.950) in the validation set.DeLong test showed that the predictive efficiency of the graph model was higher than that of the intratumoral omics model,and the difference was statistically significant in the verification set.Decision curve analysis showed that the nomogram had the highest overall net benefit ratio.Conclusion The prediction model of surgical resection of GGN combined with peritumoral radiomics can help clinicians grasp the operative nodes and reduce the occurrence of overtreatment.