增强CT影像组学对声带良、恶性病变的鉴别诊断价值

Value of Enhanced CT-Based Radiomics in Differential Diagnosis of Benign and Malignant Vocal Cord Lesions

目的探讨增强CT影像组学在声带良、恶性病变鉴别诊断中的价值。方法回顾性分析经术后病理证实的140例声带病变患者(良性90例,恶性50例)的术前增强CT图像。所有病例按7∶3的比例随机分为训练组(n=97)和验证组(n=43)。使用IBEX软件勾画感兴趣区(ROI)并提取影像组学特征,通过信度测试、标准化、单因素分析及LASSO降维处理,筛选出有效的影像组学特征,构建影像组学模型,评价其对于声带良、恶性病变的鉴别诊断效能。同时加入性别、年龄及临床影像特征(包括病灶形态是否规则、病灶强化程度、结节样突起和/或弥漫增厚)与影像组学特征组成联合模型,分别与临床模型及影像组学模型进行对比,评价鉴别诊断效能。建立列线图(Nomogram),评估其对声带良、恶性病变的预测能力。结果最终提取出15个有效的影像组学特征。在4种分类器中,随机森林模型及支持向量机(SVM)模型的预测效能最佳。随机森林模型训练组及验证组的曲线下面积(AUC)分别为1和0.998,SVM训练组及验证组的AUC分别为1和0.990。同时,联合模型的诊断效能也明显优于临床模型[训练组:联合模型的受试者工作特征曲线(ROC)vs.临床模型的ROC,P=0.004;验证组:联合模型的ROC vs.临床模型的ROC,P=0.013]。除此以外,Hosmer-Lemeshow检验(P=0.723)和校准曲线证实列线图具有良好的工作效果。结论基于增强CT的影像组学模型在声带良、恶性病变的鉴别诊断中具有一定优势。同时,与临床参数组成的联合模型预测效能更高。此外,列线图能够为声带良、恶性病变的鉴别提供直观、可靠的参考。

Objective This study aimed to explore the value of enhanced CT radiomics in the differential diagnosis of benign and malignant vocal cord lesions.Methods The preoperative enhanced CT images of 140 cases with vocal cord lesions(including 90 benign and 50 malignant cases)confirmed by pathology,were retrospectively analyzed.All cases were divided into training group(n=97)and validation group(n=43),according to the ratio of 7∶3 randomly.Then,we delineated the ROI,and extracted the radiomics features by using IBEX software.After reliability analysis,standardization,single-factor analysis,and lasso dimensionality reduction,the effective radiomics features were filtered,and constructed as the radiomics model to evaluate the differential diagnosis efficiency for vocal cord lesions.Besides,the clinical data,such as gender,age,and the clinical imaging features(including shape,the degree of enhancement,nodular protrusion and/or diffuse thickening)combined with radiomics model to form a combined model.The combined model was compared with the clinical model or the radiomics model respectively,to evaluate the diagnostic efficiency.Finally,the Nomogram was established to evaluate its predictive power in vocal cords lesions.Results A total of 15 effective radiomics features were finally extracted.The prediction power of random forest model and SVM model showed the best performance in four classifiers.The average area under the curve(AUC)of the random forest model training group and the validation group were 1 and 0.998,and that of the SVM training group and the validation group were 1 and 0.990,respectively.Besides,the prediction efficiency of combined model had better performance than the clinical model(Training group:ROC of the combined model versus ROC of the clinical model,P=0.004;Test group:ROC of the combined model versus ROC of the clinical model,P=0.013).Moreover,Hosmer-Lemeshow test(P=0.723)and calibration curve confirmed that the Nomogram had good working effect.Conclusion Radiomics model based on enhanced CT has certain advantages in the differential diagnosis of benign and malignant vocal cord lesions.Meanwhile,the combined radiomics model with clinical data has better prediction efficiency.Furthermore,the Nomogram can provide an intuitive and reliable reference for the identification of benign and malignant vocal cord lesions.