基于灰阶超声的不同影像组学模型鉴别乳腺肿块良恶性的价值

The clinical value of different radiomics based on gray scale ultrasound imaging in differentiating malignant from benign breast masses

目的探讨基于灰阶超声的不同影像组学模型鉴别乳腺肿块良恶性的临床价值。方法回顾性分析2018年10月至2020年10月皖南医学院附属太和县人民医院经病理证实的180例患者的乳腺肿块灰阶超声图像并提取影像特征,将肿块按照7∶3随机抽样,其中126个肿块作为训练组,54个肿块作为验证组。采用单因素方差分析及最小绝对收缩和选择算子(LASSO)筛选最优特征构建影像组学模型,使用5种方法构建模型。采用受试者工作特征(ROC)曲线评价各个模型的优劣,计算最优模型鉴别乳腺肿块良恶性的效能。结果最终选出8个影像特征构建影像组学模型。对于训练组,随机森林和支持向量机模型的表现能力略高于决策树和逻辑回归模型,集成算法模型表现最差;在验证组,随机森林和逻辑回归模型的表现能力较强。验证组逻辑回归模型鉴别乳腺肿块良恶性的准确性、灵敏度、特异度、阳性预测值和阴性预测值均高于随机森林,其值分别为83.33%、91.70%、83.33%、85.71%、81.82%和81.48%、83.33%、80.00%、76.92%、85.71%。结论基于灰阶超声构建的影像组学对乳腺肿块良恶性的鉴别具有较高的临床价值,且5种建模方法中以逻辑回归模型的表现能力最强。

Objective To investigate the clinical value of different radiomics based on gray scale ultrasound imaging in differentiating malignant from benign breast masses.Methods A retrospective analysis was performed for the breast mass gray scale ultrasound imaging of 180 patients who had been confirmed by pathology in the People’s Hospital of Taihe County Affiliated to Wannan Medical College from October 2018 to October 2020,and radiomics features were extracted from imaging.All masses were divided into training(n=126)and testing(n=54)groups in a ratio of 7∶3 randomly.Then one-way analysis of variance and Lasso were used to select the most important features to bulid radiomics model with five methods.Each model performance and the optimal model performance was assessed with respect to discrimination by the receiver operating characteristic curve(ROC).Results A total of eightoptimal features were selected to develop radiomics model.For the training group,the performance of random forest and support vector machine model wasslightly higher than that of decision tree and logistic regression model,and the integrated algorithm model was the worst,but in the testing group,the performance of random forest and logistic regression model was the strongest.In the testing group,the accuracy,sensitivity,specificity,positive predictive value and negative predictive value of logistic regression model in differentiating malignant from benign breast masses were higher than those of random forest,which was 83.33%,91.70%,83.33%,85.71%,81.82% and 81.48%,83.33%,80.00%,76.92%,85.71% respectively.Conclusions Radiomics based on gray scale ultrasound imaging has high clinical value in differentiating malignant from benign breast masses,and logistic regression model has the strongest performance among the five models.