摘要
目的探讨基于磁共振T2 mapping的影像组学特征在预测乳腺病灶良恶性中的应用价值。材料与方法回顾性分析经病理证实的113例患者(良性51例,恶性62例)乳腺磁共振T2 mapping图像,应用ITK-SNAP软件手动勾画磁共振T2 mapping图像中的病灶感兴趣区,利用A.K.软件(AnalysisKit,GE Healthcare)对影像组学特征进行提取。根据病理结果,将其分为两组。通过组内相关系数进行一致性检验。在良性组、恶性组中通过7∶3的比例随机分割训练集与测试集。通过Z-score标准化处理、Pearson相关系数法、递归特征消除法对训练集进行特征降维及选择,逻辑回归分类器进行分类建模,并进行5折交叉验证。分别在训练集及测试集中绘制受试者工作特征(receiver operating characteristic,ROC)曲线,以评估模型的诊断效能,通过临床决策曲线分析(decision curve analysis,DCA)评价其临床有效性。结果通过特征提取获得107个定量影像特征参数,通过特征降维及筛选最终保留6个特征参数,分别为original_shape_Sphericity、original_glcm_InverseVariance、original_glrlm_GrayLevelNonUniformityNormalized、original_glrlm_ShortRunEmphasis、original_glszm_GrayLevelNonUniformity Normalized以及original_ngtdm_Coarseness。ROC曲线在测试集的曲线下面积为0.895(95%可信区间:0.768~0.990),敏感度为94.7%,特异度为80.0%,准确度为88.2%。结论基于磁共振T2 mapping的影像组学特征可用于术前预测乳腺病灶的良恶性,且具有较高的准确度。
Objective:To investigate the diagnostic performance of radiomic features based on breast MRI T2 mapping in differentiating benign and malignant lesions.Materials and Methods:This retrospective study included T2 mapping images of breast MRI from 113patients(113 breast lesions:51 benign lesions,62 malignant lesions)confirmed by pathology.Breast lesions were segmented manually on the T2 mapping images,and radiomic features were then extracted and selected.They were divided into two groups according to the pathological results.The Kappa was measured by the intra-class correlation coefficients.The training set and test set were selected according to the ratio of 7∶3.Z-score,Pearson correlation coefficients,recursive feature elimination were used to select features in the training set.A radiomics-based predictive model using logistic regression was developed and calibrated with five-fold cross-validation.The receiver operating characteristic(ROC)curves were drawn in the training set and test set respectively to evaluate the diagnostic performance of the model.The model efficiency was evaluated using the clinical decision curve.Results:A total of 107 features were extracted from T2 mapping images for each patient.Finally,6 features(original_shape_Sphericity,original_glcm_InverseVariance,original_glrlm_Gray Level Non Uniformity Normalized,original_glrlm_ShortRunEmphasis,original_glszm_GrayLevelNonUniformityNormalized and original_ngtdm_Coarseness)were selected to construct the model for differentiating benign from malignant lesions.The area under the curve,sensitivity,specificity and accuracy of model in test set were 0.895(95%confidence interval:0.768-0.990),94.7%,80.0%and 88.2%.Conclusions:T2 mapping-based radiomics method can be used to preoperatively discriminate benign and malignant lesions with high accuracy.
作者
黄文平
王芬
刘鸿利
余雅丽
娄鉴娟
邹启桂
王思奇
蒋燕妮
HUANG Wenping;WANG Fen;LIU Hongli;YU Yali;LOU Jianjuan;ZOU Qigui;WANG Siqi;JIANG Yanni(Department of Radiology,the First Affiliated Hospital of Nanjing Medical University,Nanjing 210029,China)
出处
《磁共振成像》
CAS
CSCD
北大核心
2023年第2期50-55,共6页
Chinese Journal of Magnetic Resonance Imaging
