CT影像组学机器学习模型鉴别肺炎型黏液腺癌与机化性肺炎
Preliminary study of machine learning model based on CT radiomics with in differentiating pneumonia-type mucinous adenocarcinoma from organizing pneumonia
杜自宏 1伍志发 2刘静 2李新春 2刘红艳3
作者信息
- 1. 650032 云南昆明,云南省第一人民医院放射科/昆明理工大学附属医院
- 2. 510120 广东广州,广州医科大学附属第一医院放射科
- 3. 511500 广东清远,广州医科大学附属第六医院放射科
- 折叠
摘要
目的:评估CT影像组学结合机器学习方法鉴别原发肺炎型黏液腺癌(PTMA)与机化性肺炎(OP)的价值.方法:回顾性分析2010年1月-2020年1月在本院经病理证实的51例PTMA患者与50例OP患者的临床及影像学资料.分别在平扫及CT增强图像上提取病灶的影像组学特征,通过线性相关性分析和L1正则化方式进行特征的筛选和降维.对两组的临床特征、CT形态学特征及影像组学特征进行统计学分析,将3类特征中有统计学意义者分别或联合构建机器学习预测模型,共获得4个预测模型(临床、CT形态学征象、影像组学合联合模型).采用ROC曲线分析评估各类模型的诊断效能.结果:临床特征中的性别、咳白黏痰、癌胚抗原和糖类抗原153、CT形态学征象中的小结节、空泡/假空洞征、血管造影征和重力分布在PTMA组与OP组之间的差异均有统计学意义(P＜0.05).二元logistic回归分析显示性别、小结节、空泡/假空洞征和血管造影征是鉴别PTMA与OP的独立预测因素(P＜0.05).在训练集和验证集中各类机器学习模型的AUC:影像组学模型为0.997和0.946,临床模型为0.869和0.814,CT形态学特征模型为0.919和0.797,联合模型为0.999和0.972.Delong检验显示影像组学模型的诊断效能显著优于临床模型及CT形态学特征模型(P均＜0.05),与联合模型无显著差异(P＞0.05).结论:CT影像组学结合机器学习方法提取并分析多维度影像数据,可以有效鉴别PTMA与OP,辅助临床治疗决策.
Abstract
Objective:To evaluate the value of CT radiomics combined with machine learning in the differential diagnosis of primary pneumonia-type mucinous adenocarcinoma(PTMA)and organi-zing pneumonia(OP).Methods:The clinical and imaging data of 51 patients with PTMA and 50 pa-tients with OP confirmed by pathology from January 2010 to January 2020 in our hospital were retro-spectively analyzed.CT with/without contrast radiomics features of lesions were extracted based on pre-contrast and post-contrast enhanced CT images and subjected to feature filtering and dimensionali-ty reduction by linear correlation analysis and L1 regularization.The clinical,CT morphological and ra-diomics characteristics of the two groups were compared statistically,and the features from the three categories with statistically significant difference were used to build predictive models by machine learning respectively or combinedly.Four models were obtained,including clinical,CT features model,radiomics and combined model.ROC curve analysis was performed to evaluate the diagnostic efficacy of various models.Results:There were statistical differences(P＜0.05)between the PTMA group and the OP group in terms of gender,coughing up white and sticky sputum,carcinoembryonic antigen,car-bohydrate antigen 153,small nodules,vacuoles or pseudocavities,angiographic signs,and gravity dis-tribution in the imaging signs.Binary Logistic regression showed that gender,small nodules,vacuoles or pseudocavities,and angiographic signs were independent predictors(all P＜0.05).In training set and verification set,the AUCs of each type of machine learning model were as follows:0.997 and 0.946 for radiomics model,0.869 and 0.814 for clinical model,0.919 and 0.797 for CT morphological features model,and 0.999 and 0.972 for combined model.The Delong test showed that the AUC of the ra-diomics model was significantly higher than that of the clinical feature model and the CT morphologi-cal features model(both P＜0.05),and there was no difference in the AUC between the radiomics model and the combined model(P＞0.05).Conclusion:CT radiomics combined with machine learning methods to extract and analyze multi-dimensional image data can effectively identify PTMA and OP,and assist in clinical treatment decision-making.
关键词
肺肿瘤/肺炎型黏液腺癌/机化性肺炎/影像组学/机器学习/体层摄影术,X线计算机Key words
Lung tumor/Pneumonia-type mucinous adenocarcinoma/Organizing pneumonia/Radiomics/Machine learning/Tomography,X-ray computer引用本文复制引用
基金项目
市校联合资助项目基础与应用基础研究项目(202201020456)
广东省基础与应用基础联合基金青年项目(2019A1515111161)
出版年
2024
NETL
NSTL
万方数据