基于CT增强图像的影像组学模型对鉴别结直肠癌旁肿瘤沉积及转移淋巴结的诊断价值

Diagnostic value of radiomics model based on enhanced CT image for the identification of tumor deposition and metastatic lymph nodes in colorectal cancer

目的:探讨基于CT增强图像的影像组学模型对鉴别结直肠癌旁肿瘤沉积(TD)及转移淋巴结(MLN)的诊断价值。方法:回顾性分析经手术病理证实且CT增强图像上可辨识的45个结直肠癌旁TD和45个转移性淋巴结的影像学资料。所有患者在术前一周内行全腹部CT平扫和动静脉双期增强扫描。使用Mazda软件,在CT静脉期增强图像上提取病灶的影像组学特征,随后采用软件自带的FPM法及主成分分析法对纹理特征进行特征选择及抽取的降维方法,筛选出有预测价值的纹理特征。并将样本随机分为训练集与验证集,使用降维后的特征和临床高危因素建立随机森林、决策树、朴素贝叶斯和逻辑式回归共4种机器学习模型,计算4种模型在验证集的鉴别诊断符合率,进行ROC曲线分析并获得曲线下面积(AUC)。结果:共提取794个影像组学特征,经降维后获得有鉴别诊断意义的9个主成分特征参数。建立的随机森林、决策树、朴素贝叶斯和逻辑式回归机器学习模型在验证集中对鉴别TD与MLN的符合率分别为100%、100%、100%和92.86%,ROC曲线下面积分别为0.83、0.71、0.94和0.89。结论:基于CT增强静脉期图像的影像组学模型对结直肠癌旁肿瘤沉积及转移性淋巴结的鉴别诊断具有较大价值。

Objective:To investigate the diagnostic value of radiomics model based upon contrast-enhanced CT image in the identification of peripheral tumor deposits(TD)and metastatic lymph nodes(MLN)in colorectal cancer.Methods:A retrospective analysis was performed on 45 lesions of tumor deposition and 45 metastatic lymph nodes identified by CT enhanced scan,which were confirmed by surgery and pathology.Radiomics features were extracted from the enhanced axial CT images of the above lesions in the venous phase obtained by Mazda software.Then,FPM method and principal component analysis(PCA)were adopted to select the features and perform the dimensionality reduction,and then the valuable features were selected for postoperative pathological diagnosis.The samples were randomly divided into training sets and test sets.Four machine learning models including random forest,decision tree,naive bayes and logistic regression were established by using the characteristics and information of clinical interest after dimension reduction,and the accuracy was calculated to obtain the ROC curve and the area under the curve.Results:794 image features were extracted and 9 principal component feature parameters related to the identification of the two lesions were obtained.The diagnostic accuracy rates of the four machine learning models were 100%,100%,100%and 92.86%,respectively.The areas under the ROC curve were 0.83,0.71,0.94 and 0.89,respectively.Conclusion:The radiomics model based upon CT enhanced image is of great value in the differential diagnosis of the peripheral tumor deposition and metastatic lymph nodes in colorectal cancer.