基于影像组学和自动机器学习的PA患者肾上腺静脉取血结果的研究  

Prediction of Adrenal Venous Sampling Outcome in Patients with PA Using Radiomics and Automated Machine Learning

作　　者：谢薇[1] 陈涛[2] 罗国婷 王寒箫 舒炀 刘娟[1] 郑涛 孙怀强[1,3] XIE Wei;CHEN Tao;LUO Guoting;WANG Hanxiao;SHU Yang;LIU Juan;ZHENG Tao;SUN Huaiqiang(Department of Radiology,West China Hospital Sichuan University,Chengdu Sichuan 610041,China;Department of Endocrinology and Metabolism,West China Hospital Sichuan University,Chengdu Sichuan 610041,China;Huaxi MR Research Center,West China Hospital Sichuan University,Chengdu Sichuan 610041,China;IT Center,West China Hospital Sichuan University,Chengdu Sichuan 610041,China)

机构地区：[1]四川大学华西医院放射科,四川成都610041 [2]四川大学华西医院内分泌代谢科,四川成都610041 [3]四川大学华西医院临床磁共振研究中心,四川成都610041 [4]四川大学华西医院信息中心,四川成都610041

出　　处：《中国医疗设备》2024年第2期45-51,共7页China Medical Devices

基　　金：四川省科技厅重点研发项目(23ZDYF2910)。

摘　　要：目的建立基于高分辨率增强CT影像和自动机器学习技术的原发性醛固酮增多症(Primary Aldosteronism,PA)亚型术前预测模型。方法回顾性研究经肾上腺静脉取血(Adrenal Venous Sampling,AVS)结果亚型诊断的PA患者312例,其中,207例诊断为单侧优势(AVS-右∶AVS-左=93∶114),105例诊断为双侧优势。纳入患者初诊CT影像,基于薄层静脉期图像提取双侧肾上腺影像组学特征,并定义影像组学商值特征为双侧肾上腺对应影像组学特征的比值,再将特征向量输入自动机器学习进行模型训练。结果根据自动模型筛选,随机森林分类器在预测AVS结果方面取得了较好的整体性能,其中准确度为0.7500,召回率为0.7466,受试者工作特征曲线下面积为0.8792。结论本系统在预测PA患者的AVS结果方面展示出了一定的潜力,因此,机器学习模型可以在常规临床实践中辅助预测PA的亚型诊断。Objective To build a preoperative prediction model for the subtype classification of primary aldosteronism(PA)based on enhanced high-resolution CT and automated machine learning techniques.Methods A retrospective study was conducted on 312 patients with PA diagnosed by subtypes of adrenal venous sampling(AVS).Among them,207 were diagnosed with unilateral dominance(AVS right∶AVS left=93∶114),and 105 were diagnosed with bilateral dominance.Initial CT images were retrospectively included and radiomics features were extracted from bilateral adrenal based on thin layer venous phase images.The quotient radiomics features were defined as the left-right ratio of bilateral adrenals radiomics features,and then input feature vectors into automatic machine learning for model training.Results According to the automatic model screening,the random forest classifier achieved good overall performance in predicting AVS results,with an accuracy of 0.7500,a recall rate of 0.7466,and an area under operating receiver characteristic curve of 0.8792.Conclusion This system has shown certain potential in predicting AVS outcomes in PA patients.Therefore,the machine learning model can assist in predicting the subtype diagnosis of PA in routine clinical practice.

关 键 词：肾上腺静脉取血 原发性醛固酮增多 亚型诊断 影像组学 自动机器学习 

分 类 号：R197.39[医药卫生—卫生事业管理]