基于MRI征象及影像组学的列线图预测腮腺良性与恶性肿瘤的价值

Differentiation of benign and malignant lesions of the parotid gland by MRI based imaging features and radiomics nomogram

目的建立并验证基于MRI征象和影像组学的列线图鉴别腮腺良性与恶性肿瘤的效能。方法回顾性收集2015年1月至2020年5月青岛大学附属医院86例经手术病理证实的腮腺肿瘤患者为训练集,收集2013年1月至2020年1月香港大学深圳医院35例患者为独立外部验证集。采用logistic回归基于临床及MRI征象建立临床诊断模型。基于术前平扫T1WI和预饱和脂肪抑制T2WI(fs-T2WI)进行影像组学特征提取,建立影像组学诊断模型。基于影像组学评分及临床诊断模型,通过logistic回归建立影像组学+临床联合诊断模型及列线图。采用受试者操作特征(ROC)曲线评价各模型诊断腮腺良性与恶性肿瘤的效能,ROC曲线下面积(AUC)的比较采用DeLong检验。结果Logistic回归结果显示,腮腺深叶受累(OR值为3.285,P=0.040)和周围组织结构侵犯(OR值为15.919,P=0.013)是腮腺恶性肿瘤的独立影响因素,将二者构建临床诊断模型。基于平扫T1WI和fs-T2WI,共提取19个特征构建影像组学诊断模型。联合影像组学评分以及腮腺深叶受累、周围组织结构侵犯2个常规影像学特征建立联合诊断模型及列线图。临床诊断模型、影像组学诊断模型、联合诊断模型在训练集和验证集中诊断腮腺良性与恶性肿瘤的AUC分别为0.758、0.951、0.953和0.752、0.941、0.964。在训练集和验证集中,影像组学诊断模型、联合诊断模型的AUC均高于临床诊断模型(训练集:Z=3.95、4.31,P均<0.001;验证集:Z=2.16、2.67,P=0.031、0.008),影像组学诊断模型、联合诊断模型间AUC差异无统计学意义(训练集:Z=0.39,P=0.697;验证集:Z=1.10,P=0.273)。结论本研究所建立的MRI影像组学模型以及由腮腺深叶受累、周围组织结构侵犯、MRI影像组学特征组成的联合诊断模型,能有效鉴别腮腺良恶性肿瘤,具有较高的预测效能。

Objective To develop and validate a MRI-based radiomics nomogram combining with radiomics signature and clinical factors for the preoperative differentiation of benign parotid gland tumors(BPGT)and malignant parotid gland tumors(MPGT).Methods From January 2015 to May 2020,86 patients with parotid tumors confirmed by surgical pathology in the Affiliated Hospital of Qingdao University were enrolled as training sets,and 35 patients in the University of Hong Kong-Shenzhen Hospital from January 2013 to January 2020 were enrolled as independent external validation sets.The logistic regression was used to establish a clinical-factors model based on demographics and MRI findings.Radiomics features were extracted from preoperative T1WI and fat-saturated T2WI(fs-T2WI),a radiomics signature model was constructed,and a radiomics score(Rad-Score)was calculated.A combined diagnostic model and nomogram combining with the Rad-score and independent clinical factors was constructed using multivariate logistic regression analysis.The receiver operating characteristic(ROC)analysis was used to evaluate the performance of each model and DeLong test was used for comparison of area under the ROC curve(AUC).Results The logistic regression results showed that deep lobe involvement(OR=3.285,P=0.040)and surrounding tissue invasion(OR=15.919,P=0.013)were independent factors for MPGT and constructed the clinical-factors model.A total of 19 features were extracted from the joint T1WI and fs-T2WI to build the radiomics signature model.The combined diagnostic model and nomogram incorporating deep lobe involvement,surrounding tissue invasion and Rad-score were established.The AUCs of the clinical-factors model,radiomics signature model and combined diagnostic model for differentiating BPGT from MPGT for the training and validation sets were 0.758,0.951,0.953 and 0.752,0.941 and 0.964 respectively.The AUCs of the radiomics signature model and the combined diagnostic model were significantly higher than those of the clinical-factors model for both training and validation sets(training set:Z=3.95,4.31,both P<0.001;validation set:Z=2.16,2.67,P=0.031,0.008).There was no statistical difference in AUCs between the radiomics signature model and combined diagnostic model(training set:Z=0.39,P=0.697;validation set:Z=1.10,P=0.273).Conclusions The MRI-based radiomics signature model and radiomics nomogram incorporating deep lobe involvement,surrounding tissue invasion,and Rad-score showed favorable predictive efficacy for differentiating BPGT from MPGT.