评估不同特征筛选及分类组学方法在定量识别乳腺结节超声影像特征的价值  

Evaluation of the Performance of Different Ultrasonographic Radiomics Feature Screening and Classification Methods in the Quantitative Identification of Breast Lesions

作　　者：周琦[1] 李娜[1] 王娟[1] 姜珏[1] 余珊珊[1] 周茹 何鑫[1] Zhou Qi;Li Na;Wang Juan;Jiang Jue;Yu Shanshan;Zhou Ru;He Xin(Department of Ultrasound,The Second Affiliated Hospital of Xi'an Jiaotong University,Xi'an 710004,China)

机构地区：[1]西安交通大学第二附属医院医用超声研究室,西安市710004

出　　处：《中国超声医学杂志》2023年第10期1113-1116,共4页Chinese Journal of Ultrasound in Medicine

基　　金：陕西省重点研发计划项目(No.2023-YBSF-513)。

摘　　要：目的 比较超声影像组学不同特征筛选及分类组学方法在识别乳腺结节特征中的效能,构建最优乳腺超声图像特征个体化定量识别模型。方法 回顾性分析因乳腺结节行超声检查的患者697例,共703个结节,703张二维超声图像。通过不同特征筛选和分类组学方法定量识别乳腺结节特征并比较不同方法组合之间的效能。结果 不同特征筛选方法及分类组学方法的诊断效能存在差异(P<0.05)。在识别钙化、形态及边缘特征中,最优特征筛选百分比-逻辑回归方法组合均表现出最高的诊断价值,受试者工作特征(ROC)曲线与曲线下面积(AUC)分别为0.794、0.868、0.777;识别纵横比根据模型选择-分布式梯度增强库(XGBOOST)组合方法诊断价值最高,ROC AUC为0.700。结论 本研究构建了乳腺结节超声图像定量个性化特征识别模型,可有效辅助超声医师提高结节特征判断的准确度,为进一步研究超声影像组学奠定应用基础。Objective To compare the efficiency of different ultrasound radiomics feature screening and classification methods in identifying breast nodules,and to construct the optimal individualized quantitative recognition model of breast ultrasound image features.Methods A total of 697 patients with 703 breast nodules(7032D ultrasound images)who underwent ultrasonography in our hospital were analyzed retrospectively.Different combinations of feature screening and classification machine learning methods were used to identify the features of breast nodules,and the effectiveness of different combinations was compared.Results The diagnostic efficiency of different radiomics feature screening and classification methods was different(P<0.05).In identifying calcification,morphological and marginal features,"the optimal feature screening percentage-logistic regression"method showed the highest diagnostic value,with the AUC of 0.794,0.868 and 0.777,receiver operating characterstic(ROC)curve.In terms of aspect ratio,"selection based on the (model-XGBOOST)"method had the highest diagnostic value,with the AUC of 0.700.Conclusions A quantitative personalized feature recognition model of breast nodules on ultrasound images is constructed,which can effectively assist doctors to improve the accuracy of nodule feature recognition,and lay the application foundation for further research on ultrasound radiomics.

关 键 词：超声 乳腺结节 影像组学 特征识别 

分 类 号：R445.1[医药卫生—影像医学与核医学] R737.9[医药卫生—诊断学]