基于小样本甲状腺高b值扩散加权成像影像组学的机器学习在鉴别诊断结节型桥本甲状腺炎与甲状腺微小乳头状癌中的应用

Machine Learning-Based Small Sample High b-Value Diffusion Weighted Imaging Radiomics in the Differential Diagnosis of Nodular Hashimoto's Thyroiditis and Papillary Thyroid Microcarcinoma

目的探讨小样本甲状腺高b值扩散加权成像(DWI)影像组学分类模型鉴别诊断结节型桥本甲状腺炎(NHT)与甲状腺微小乳头状癌(PTMC)的价值。资料与方法回顾性纳入经病理证实的40例46个NHT和48例61个PTMC甲状腺高b值DWI,对结节进行分割、特征值提取,并使用3种特征值标准化、2种降维、3种特征值筛选方法确定影像组学特征值,采用受试者工作特征(ROC)曲线和曲线下面积(AUC)分析10种分类模型对两种病变的鉴别诊断效能。结果46个NHT和61个PTMC共提取了107个影像组学特征值。当选择Normalize to unit(NormUnit)进行标准化、主成分分析进行降维、ANOVA进行特征值筛选,在15个特征值时,线性判别分析分类模型鉴别诊断NHT与PTMC效果最优,准确度为85.71%,敏感度为80.00%,特异度为100.00%,AUC值为0.925。结论小样本甲状腺高b值DWI影像组学分类模型对于NHT和PTMC具有鉴别诊断价值。

Purpose To evaluate the role of machine-learning model based small sample high b-value diffusion weighted imaging(DWI)radiomics in differentiation between nodular Hashimoto's thyroiditis(NHT)and papillary thyroid microcarcinoma(PTMC).Materials and Methods Thyroid high b-value DWI of 46 NHT in 40 patients and 61 PTMC in 48 patients confirmed by pathology were selected into this retrospective study.After nodules segmentation and radiomics features extraction,3 features'standardizations,2 features'dimensionality reductions,3 features'selections and 10 classification models were used to select radiomics features and comparison analysis on the 10 classification models for differential diagnosis between NHT and PTMC was carried out,the receiver operator characteristic(ROC)curve and the area under curve(AUC)were further drawn and the diagnostic efficiency was compared.Results A total of 107 radiomics features were extracted from the 46 NHT and 61 PTMC.When features standardization of Normalize to unit(NormUnit),features dimensionality reduction of principal component analysis(PCA)and features selection of analysis of variance(ANOVA)were used and 15 radiomics features were selected,the classification model of linear discriminant analysis had the best diagnostic efficiency in differentiating PTMC from NHT with accuracy of 85.71%,sensitivity of 80.00%,specificity of 100.00%and AUC of 0.925 in the test set,respectively.Conclusion The machine-learning model based small sample high b-value DWI radiomics contributes to differentiation between NHT and PTMC.