桥本甲状腺炎性结节与甲状腺微小乳头状癌鉴别诊断:基于MRI影像组学机器学习的应用  

Differentiation Between Hashimoto's Thyroiditis Nodule and Papillary Thyroid Microcarcinoma:Application of MRI Radiomics-Based Machine Learning

作　　者：王庆军[1] 程流泉[1] 符永瑰 梁晓晶 洪柳 李梦露[1] WANG Qingjun;CHENG Liuquan;FU Yonggui;LIANG Xiaojing;HONG Liu;LI Menglu(Department of Radiology,the 6th Medical Center of Chinese PLA General Hospital,Beijing 100048,China;不详)

机构地区：[1]解放军总医院第六医学中心放射诊断科,北京100048 [2]解放军总医院第六医学中心病理科,北京100048

出　　处：《中国医学影像学杂志》2023年第3期213-219,共7页Chinese Journal of Medical Imaging

基　　金：2021年北京市海淀区卫生健康发展科研培育计划立项项目(HP2021-32-80501)。

摘　　要：目的研究基于甲状腺MRI多序列图像影像组学机器学习分类诊断模型鉴别诊断桥本甲状腺炎性结节与甲状腺微小乳头状癌的价值。资料与方法回顾性纳入2015年6月—2022年4月于解放军总医院第六医学中心行MRI检查并经病理证实的114个桥本甲状腺炎性结节和76个甲状腺微小乳头状癌,MRI检查序列包括T1WI、T2WI、扩散加权成像(b=0、800、2000 s/mm^(2))、表观扩散系数和增强扫描T1WI。基于MRI图像对两组病灶进行分割、配准、影像组学特征提取和特征选择,经有监督机器学习建立并超参数调优6个常用的单模型分类诊断模型:逻辑回归、线性判别分析、支持向量机、K邻近算法、高斯朴素贝叶斯和分类回归树,比较6个分类模型对两组病变的鉴别诊断能力。结果每个序列图像提取960个特征,共提取6720个影像组学特征,最终经特征选择筛选出30个表观扩散系数原始图像形状和一阶统计量特征。在6个分类诊断模型中,支持向量机和逻辑回归模型鉴别桥本甲状腺炎性结节和甲状腺微小乳头状癌的效果最优,曲线下面积均为0.97。结论基于甲状腺MRI影像组学机器学习分类模型鉴别诊断桥本甲状腺炎性结节与甲状腺微小乳头状癌具有一定的应用价值。Purpose To explore the diagnostic value of MRI multi-sequences radiomics-based machine learning in distinguishing Hashimoto's thyroiditis nodule from papillary thyroid microcarcinoma.Materials and Methods A total of 114 Hashimoto's thyroiditis nodules and 76 papillary thyroid microcarcinomas confirmed by pathology were retrospectively enrolled from June 2015 to April 2022 in the 6th Medical Center of Chinese PLA General Hospital.The MRI sequences included T1WI,T2WI,diffusion weighted imaging(b=0,800 and 2000 s/mm^(2)),apparent diffusion coefficient and contrast-enhanced T1WI.Nodules were segmented,registered and radiomics features were extracted and selected.The frequently-used six single diagnostic models,including Logistic regression,linear discriminant analysis,support vector machine,K-nearest neighbors,Gaussian naive Bayes and classification and regression tree,were built and hyper-parameters were tuned based on the selected radiomic features by supervised machine learning.The ability of differential diagnosis for the two diseases was compared between the six models.Results A total of 6720 radiomic features were extracted(960 features for each MRI sequence)and 30 features were finally selected.Among the six diagnostic models,the support vector machine and Logistic regression had the superior performance in classification between Hashimoto's thyroiditis nodules and papillary thyroid microcarcinomas with area under curve value of 0.97,respectively.Conclusion MRI radiomics-based machine learning has a certain application value for differentiating Hashimoto's thyroiditis nodules from papillary thyroid microcarcinomas in clinical practice.

关 键 词：甲状腺炎 甲状腺癌 乳头状 磁共振成像 影像组学 机器学习 诊断 鉴别 

分 类 号：R445.2[医药卫生—影像医学与核医学] R736.1[医药卫生—诊断学] R581.4[医药卫生—临床医学]