放疗定位CT影像组学对识别局部进展期直肠癌新辅助放化疗后病理无反应患者的预测价值  

Value of radiotherapy planning CT-based radiomics in predicting pathological non-response to neoadjuvant chemoradiotherapy in rectal cancer

作　　者：柳燕冬 李治斌 代明 陈龙[1] 秦颂兵[1] LIU Yandong;LI Zhibing;DAI Ming;CHEN Long;QIN Songbing(Department of Radiation Oncology,First Affiliated Hospital of Soochow University,Suzhou,Jiangsu 215006,China)

机构地区：[1]苏州大学附属第一医院肿瘤放疗科,江苏苏州215006

出　　处：《中华肿瘤防治杂志》2024年第22期1385-1392,共8页Chinese Journal of Cancer Prevention and Treatment

基　　金：国家自然科学基金(82073337);江苏省医学重点学科(ZDXK202235);中华医学会放射肿瘤治疗学分会免疫放疗研究基金(Z-2017-24-2108)。

摘　　要：目的建立基于放疗定位CT的影像组学模型,以此验证其对局部进展期直肠癌(LARC)新辅助放化疗后病理无反应患者的预测价值,为其临床诊治提供参考。方法回顾性分析2015-01-01-2022-12-31苏州大学附属第一医院行新辅助放化疗(nCRT)联合直肠全系膜切除术的151例LARC患者临床资料。根据术后病理肿瘤退缩分级(TRG)将TRG 3级定义为病理无反应组,TRG 0~2级定义为病理反应组。将2015-01-01-2020-12-31行定位CT的患者归为训练集(n=103),2021-01-01-2022-12-31行定位CT的患者归为验证集(n=48)。通过门静脉期增强CT图像手动勾画全部直肠肿瘤并提取影像组学特征,经弹性网络回归筛选出最佳变量后通过极端梯度提升机器学习法构建影像组学模型并计算影像组学评分Rad-score。通过多因素logstic回归筛选出与TRG 3级相关的临床危险因子并结合Rad-score通过决策树构建临床与影像组学联合模型以及单纯临床模型,受试者工作特征(ROC)曲线用于性能评估,Delong检验用于性能比较。结果1316个影像组学特征经过特征筛选后剩余9个最优特征用于模型构建,极端梯度提升机器学习模型训练集及验证集曲线下面积(AUC)分别为0.802(95%CI:0.719~0.867)及0.793(95%CI:0.655~0.909)。多因素logistic回归结果显示,治疗前肿瘤最大厚度(OR=1.119,95%CI:1.036~1.208,P=0.004)及癌胚抗原(OR=1.038,95%CI:1.007~1.069,P=0.016)是LARC患者出现TRG 3级的影响因素。通过决策树算法构建单纯临床模型和临床与影像组学联合模型,训练集AUC分别为0.727(95%CI为0.646~0.786)和0.858(95%CI为0.800~0.921),P<0.001;验证集AUC分别为0.704(95%CI为0.589~0.818)和0.843(95%CI为0.718~0.932),P=0.042。临床与影像组学联合模型灵敏度、特异度、准确度分别为0.923、0.686、0.750。通过特征重要性分析可见,Rad-score权重最大(0.538),肿瘤最大厚度次之(0.346),癌胚抗原权重最低(0.116)。结论基于放疗定位CT的影像组�Objective To establish an radiomics model based on radiotherapy localization CT to verify its predictive value for patients with locally advanced rectal cancer(LARC)who have no pathological response after neoadjuvant radiotherapy and chemotherapy,and provide reference for their clinical diagnosis and treatment.Methods A retrospective analysis was performed on 151 patients with LARC who underwent neoadjuvant chemo radiotherapy combined with total mesorectal ex-cision(TME)from 1st January 2015 to 31 December 2022.According to the postoperative pathological tumor regression grade(TRG),TRG 3 was defined as the pathological non-response group and TRG 0-2 was defined as the pathological response group.Patients who underwent planning CT from 1st January 2015 to 31 December 2020 were classified as the training cohort(n=103),and patients who underwent planning CT from 1st January 2021 to 31 December 2022 were clas-sified as the validation cohort(n=48).Primary tumors were manually delineated and radiomics features were extracted.Elastic net regression was applied to selected predictive features to build a radiomic signature for TRG 3 prediction(Rad-score)through extreme gradient boosting machine learning(Xgboost).The clinical risk factors related to TRG 3 were selected through multivariate logistic regression and constructed a clinical model(based on the clinical risk factors)through decision tree.At the same time,we constructed a combined model based on the Rad-score and clinical risk factors through decision tree.The performance of the model was evaluated and compared by the receiver operating characteristic curve(ROC).Results After feature screening,1316 radiomics features were selected,leaving 9 optimal features for model construction.The area under the curve(AUC)of the extreme gradient boosting machine learning model training and validation sets were 0.802(95%CI:0.719-0.867)and 0.793(95%CI:0.655-0.909),respectively.The results of multiple logistic regression showed that the maximum tumor thickness before treatment(OR=

关 键 词：直肠肿瘤 影像组学 新辅助放化疗 肿瘤消退分级 肿瘤最大厚度 

分 类 号：R735.37[医药卫生—肿瘤]