45 keV单能量能谱CT影像组学结合机器学习预测肝癌TACE短期疗效

45 keV Single Energy Spectrum CT Imaging Omics Combined with Machine Learning for Predicting Short-Term TACE Efficacy in Liver Cancer

目的 探讨术前45 keV单能量能谱CT影像组学结合机器学习预测肝癌经导管肝动脉化疗栓塞术(TACE)后短期疗效的应用价值。方法 回顾性分析104例经病理证实的肝细胞癌(HCC)患者,随机划分为训练集(71例)和验证集(33例)。搜集患者治疗前能谱CT单能量及常规增强CT(CECT)图像,提取病灶影像组学特征。使用LASSO算法结合十折交叉验证筛选能谱组学特征,用逻辑回归(LR)、随机森林(RF)、支持向量机(SVM)、极限梯度提升(XGBoost)、K最近邻(KNN)等5种机器学习方法构建能谱影像组学模型,基于最佳能谱组学模型的机器学习方法与相应特征构建CECT模型,利用多因素LR分析筛选临床资料变量构建临床模型。受试者工作特征(ROC)曲线用于评价模型效能。决策曲线分析(DCA)评价最佳影像组学模型、CECT模型及临床模型的临床决策能力。结果 通过筛选共纳入19种特征构建模型,5种能谱影像组学模型的曲线下面积(AUC)分别为训练集0.72,0.68,0.77,0.70,0.71,验证集0.55,0.58,0.74,0.74,0.65。基于多因素LR筛选出两项临床变量分别为甲胎蛋白(AFP)>400μg/L和Child分级,临床模型训练集和验证集的AUC分别为0.63和0.67,CECT SVM模型AUC为训练集0.69、验证集0.61。校准曲线提示能谱及常规SVM模型和临床模型拟合优度良好。DCA曲线显示能谱SVM模型具有更高的临床应用价值。结论 术前45 keV单能量能谱CT影像组学结合机器学习能有效预测肝癌TACE后的短期疗效,其中能谱SVM模型预测性能最佳,较其他模型更具决策作用。

Objective To explore the application value of preoperative 45 keV single energy spectral CT imaging combined with machine learning in predicting short-term efficacy of liver cancer after TACE.Methods Retrospective analysis of 104 patients with pathologically confirmed hepatocellular carcinoma,randomly divided into a training set(71 cases)and a validation set(33 cases).Collect pre treatment energy spectrum CT single energy and conventional CECT images of patients,and extract imaging omics features of lesions.Using LASSO algorithm combined with ten fold cross validation to screen spectral omics features,constructing spectral imaging omics models using five machine learning methods including LR,RF,SVM,XGBoost,KNN,etc.Based on the machine learning method of the best spectral omics model and corresponding features,constructing a CECT model,and using multi factor LR analysis to screen clinical data variables to construct a clinical model.The ROC curve is used to evaluate the effectiveness of the model.The DCA curve evaluates the clinical decision-making ability of the best imaging omics model,CECT model,and clinical model.Results A total of 19 feature construction models were selected and included.The AUC of the 5 imaging omics models were the training set(0.72,0.68,0.77,0.70,0.71)and the validation set(0.55,0.58,0.74,0.74,0.65),respectively.Based on multiple factor LR screening,two clinical variables were identified as AFP>400μg/L and Child grading.The AUC of the clinical model training set and validation set were 0.63 and 0.67,respectively.The AUC of the CECT SVM model was 0.69 in the training set and 0.61 in the validation set.The calibration curve indicates a good fit between the energy spectrum,conventional SVM model,and clinical model.The DCA curve shows that the energy spectrum SVM model has higher clinical application value.Conclusion The combination of preoperative 45 keV single energy spectral CT imaging omics and machine learning can effectively predict the short-term efficacy of liver cancer after TACE.Among them,the energy spectral SVM model has the best predictive performance and is more decisive than other models.