可视化影像决策模型在评估肺结节浸润程度中的价值
Value of decision model based on visualized clinical imaging in evaluating the infiltration of pulmonary nodule
张榕 1蔡宏杰 2梁演婷 3洪敏萍 4刘子蔚 1杨少民 5王华峰 1胡秋根1
作者信息
- 1. 528308 广东佛山,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
- 2. 310051 浙江杭州,浙江中医药大学第一临床医学院
- 3. 510080 广东广州,广东省人民医院(广东省医学科学院)放射科
- 4. 314000 浙江嘉兴,浙江省嘉兴市中医医院放射科
- 5. 528315 广东佛山,佛山市顺德区乐从医院放射科
- 折叠
摘要
目的:探讨基于临床资料、影像征象和影像组学特征构建的联合模型在术前对肺结节浸润程度的预测价值,并通过决策热图及Shapley算法对模型进行可视化分析.方法:回顾性搜集2018年1月—2022年3月在本院经病理确诊的179例肺结节患者的临床资料和术前CT图像(肺窗平扫).根据肺肿瘤新分类,分为腺体前驱病变组(78例)和浸润性肺腺癌组(101例).采用Deepwise软件,分别提取瘤灶、瘤周3 mm和5 mm区域的影像组学特征.使用单因素分析、相关性分析、Boruta算法和逐步logistic回归分析等特征筛选算法确定各区域的最佳组学特征,然后采用logistics方法分别构建3个单区域及2个多区域(肿瘤+瘤周3 mm及肿瘤+瘤周5 mm)共5个影像组学模型,分析各模型的预测效能并计算其影像组学评分(Radsocre).通过单因素和多因素logistic回归方法筛选相关临床指标和结节的主要CT征象,并采用XGBoost算法将筛选出的高危因素结合瘤灶+瘤周3 mm联合模型的影像组学得分构建临床影像联合模型.额外收集浙江省嘉兴市中医医院经病理证实的69例肺结节患者的临床和CT资料来完成联合模型的泛化性验证.利用决策热图和Shapley算法对模型分别进行可视化和特征贡献度分析.结果:相比单区域影像组学模型(训练集:AUC=0.740、0753、0.768;验证集:AUC=0.841、0.856、0.809),多区域影像组学模型在两个数据集中均显示出更高的预测效能(AUC=0.878和0.834).XGBoost联合模型的预测效能得到进一步地提高(AUC=0.948和0.886).Shap-ley分析显示影像组学得分、CT值和结节长度为预测肺结节浸润程度的最重要的3个特征.决策热图算法实现了对浸润性预测推演过程的可视化.结论:XGBoost模型对肺结节浸润性的评估具有较高的准确性和泛化性.决策热图实现了可解释机器学习算法的可视化从而保障了模型的实用性,为肺结节的临床处理及管理提供了一种无创性的辅助诊断工具.
Abstract
Objective:To explore the predictive value of the combined model based on clinical da-ta,imaging signs and radiomics features in the infiltration degree of preoperative pulmonary nodule,and to perform visualized analysis of the model by decision heat map and Shapley algorithm.Methods:Clinical data and preoperative CT images (lung window plain scan)of 179 patients with pulmonary nodules diagnosed pathologically in our hospital from January 2018 to March 2022 were retrospectively collected.According to the new classification of lung tumor,all subjects were divided into glandular prodromal disease group (n=78)and invasive lung adenocarcinoma group (n=101).Deepwise soft-ware was used to extract the imaging features of tumor foci,peritumor-3mm and-5mm regions.Uni-factor analysis,correlation analysis,boruta algorithm and stepwise logistic regression analysis feature screening algorithm were used to determine the best radiomics features of each region.Then,three sin-gle-region models and two multi-regions (tumor+peritumor-3mm,and tumor+peritumor-5mm)mo-dels were constructed by logistics method.A total of five radiomics models were used to analyze the predictive efficacy of each model and calculate its radiomics score (Radsocre).The relevant clinical in-dicators and major CT signs of nodules were screened by univariate and multivariate logistic regression methods,and the combined clinical image model was constructed by combining the selected high risk factors and Radscore of tumor+peritumor-3mm using XGBoost algorithm.Additional clinical and CT data of 69 patients with pulmonary nodules confirmed by pathology in Jiaxing Hospital of Traditional Chinese Medicine of Zhejiang Province were collected to complete the generalization verification of the combined model.The decision heatmap and Shapley algorithm were used to visualize the model and ex-plain the feature's contribution,respectively.Results:Compared with the three single-region radiomics models (training set:AUC=0.740,0753 and 0.768;validation set:AUC=0.841,0.856 and 0.809),the multi-region radiomics models showed better performance in the two datasets (AUC=0.878 and 0.834).The performance of the XGBoost combined model was further improved (AUC=0.948 and 0.886).Shapley algorithm analysis showed that Radscore,CT value and nodule length were the three most important characteristics for predicting pulmonary nodule invasiveness.The decision heatmap al-gorithm realizes the visualization prediction and deduction process of tumor infiltration.Conclusion:The clinical-radiomic combined XGBoost model has high accuracy and generalization in the evaluation of pulmonary nodule invasiveness.The decision heatmap can realize the visualization of interpretable machine learning algorithms ensures the practicability of the model and provides a non-invasive tool for clinical treatment and management of pulmonary nodules.
关键词
肺结节/浸润程度/瘤周/影像组学/Boruta算法/XGBoost算法/Shapley算法Key words
Pulmonary nodules/Infiltration/Peritumor/Radiomic/Boruta algorithm/Extreme gradient boosting algorithm/Shapley additive explanations引用本文复制引用
基金项目
佛山市科技计划(2220001005383)
南方医科大学顺德医院科研启动计划(SRSP2021021)
佛山市高层次医学人才培养项目(医学骨干人才)()
出版年
2024
NETL
NSTL
万方数据