基于CT放射组学列线图预测肝细胞癌TACE术后肿瘤进展:一项回顾性研究

Predicting tumor progression after TACE surgery for hepatocellular carcinoma based on CT radiomic nomogram:A retrospective study

目的开发并验证基于动态对比增强CT的放射组学列线图,用于术前无创预测肝细胞癌经动脉化疗栓塞(TACE)术后肿瘤反应。方法选择2016年1月—2022年12月在本院接受TACE治疗的肝细胞癌患者100例为研究对象,将患者分为TACE缓解组(CR和PR病人)以及非TACE缓解组(SD和PD病人),其中TACE缓解组包含49例患者,非TACE缓解组包含51例患者,按照8︰2的比例将其随机分配至训练组和验证组两组,训练组80例,验证组20例。TACE治疗前两周内进行动态对比增强CT检查,在TACE术后2个月左右按照改良的实体肿瘤反应评估标准(mRECIST)判断疗效。使用最小绝对收缩和选择算子(LASSO)算法进行放射组学标签筛选,然后使用10折交叉验证的逻辑回归(LR)、轻量级梯度提升机(LightGBM)和多层感知机(MLP)算法去构建临床、放射组学、列线图和堆叠模型。使用AUC、校准曲线和决策曲线来进行评估模型的诊断性能、校准性和临床净收益。结果在训练组中TACE缓解组与非TACE缓解组患者的最大肿瘤平均直径比较,差异有统计学意义(P<0.05);与此同时,验证组中TACE缓解组与非TACE缓解组患者非光滑肿瘤边缘比较,差异有统计学意义(P<0.05)。最大肿瘤直径和非光滑肿瘤边缘是TACE术后肿瘤复发的独立预测因子。使用临床独立预测因子(最大肿瘤直径和不光滑肿瘤边缘)和放射组学标签去构建列线图,列线图和堆叠模型的最佳模型为MLP模型,在训练组中列线图的诊断性能与堆叠模型相差不大(AUC为0.874和0.878),在验证组中列线图的诊断性能相比于堆叠模型有所改善(AUC=0.889和0.879)。校准曲线和决策曲线表明列线图具有良好的校准性和临床净收益。结论基于动态增强CT构建地放射组学列线图,能够很好地预测肝细胞癌TACE术后的肿瘤反应,可以为肝细胞癌患者的临床治疗决策做出进一步指导。

Objective To develop and validate a radiomic nomogram based on dynamic contrast-enhanced CT for non-invasive preoperative prediction of tumor response to transarterial chemoembolization(TACE)in patients with hepatocellular carcinoma.Methods One hundred patients with hepatocellular carcinoma treated with TACE at our institution from January 2016 to December 2022 were studied.Patients were categorized into TACE response(CR and PR)and non-response(SD and PD)groups,with 49 and 51 patients respectively.They were randomly allocated to training(80 patients)and validation(20 patients)sets in an 8︰2 ratio.Dynamic contrast-enhanced CT was performed within two weeks before TACE,and treatment response was assessed using modified Response Evaluation Criteria In Solid Tumors(mRECIST)approximately two months post-TACE.The least absolute shrinkage and selection operator(LASSO)algorithm was utilized for radiomic feature selection,followed by construction of clinical,radiomic,nomogram,and stacked models using logistic regression(LR),LightGBM,and multi-layer perceptron(MLP)algorithms with 10-fold cross-validation.Model performance was evaluated using AUC,calibration curves,and decision curves to assess diagnostic accuracy,calibration,and clinical utility.Results In the training set,there was a statistically significant difference in the mean maximum tumor diameter between the TACE response and non-response groups(P<0.05).Similarly,in the validation set,non-smooth tumor margins showed significant differences between the two groups(P<0.05).Maximum tumor diameter and non-smooth tumor margins were identified as independent predictors of post-TACE tumor recurrence.A nomogram incorporating these clinical predictors and radiomic features was constructed,with the MLP model proving optimal among the nomogram and stacked models,showing similar AUCs of 0.874 and 0.878 in the training set,and improved AUC to 0.889 for the nomogram in the validation set compared to 0.879 for the stacked model.Calibration and decision curves indicated good model calibration and clinical net benefit.Conclusions A radiomic nomogram based on dynamic contrast-enhanced CT can effectively predict tumor response following TACE in hepatocellular carcinoma,offering valuable guidance for clinical treatment decisions.