基于CT的直肠癌新辅助化疗后病理完全缓解预测模型的初步探索--DACCA数据库的联合研究  

Preliminary study on prediction model based on CT for pathological complete response of rectal cancer after neoadjuvant chemotherapy

作　　者：李芊 周逸菲 李峥艳[4] 汪晓东[1] 高绍兵 李立[1] LI Qian;ZHOU Yifei;LI Zhengyan;WANG Xiaodong;GAO Shaobing;LI Li(Department of Gastrointestinal Surgery,West China Hospital,Sichuan University,Chengdu 610041,P.R.China;West China School of Medicine in Sichuan University,Chengdu 610041,P.R.China;School of Computer Science,Sichuan University,Chengdu 610065,P.R.China;Department of Radiology,West China Hospital,Sichuan University,Chengdu 610041,P.R.China)

机构地区：[1]四川大学华西医院胃肠外科,成都610041 [2]四川大学华西临床医学院,成都610041 [3]四川大学计算机学院,成都610065 [4]四川大学华西医院放射科,成都610041

出　　处：《中国普外基础与临床杂志》2020年第5期606-611,共6页Chinese Journal of Bases and Clinics In General Surgery

基　　金：四川大学大学生创新创业计划项目(项目编号:C2019104739)。

摘　　要：目的初步探索基于CT影像组学特征建立的决策树模型对局部进展期直肠癌(LARC)患者行单纯新辅助化疗疗效的评估价值。方法回顾性分析四川大学华西医院肠癌数据库(DACCA)中2016年10月至2019年3月期间符合本研究纳入和排除标准的244例单纯新辅助化疗后行根治性手术的LARC患者的临床及CT检查的DICOM格式图像资料。利用ITK-SNAP软件选取肿瘤最大层面并对影像感兴趣区域进行勾画。使用计算机随机分配软件将200例患者纳入训练集,44例患者纳入测试集。利用MATLAB软件读取DICOM格式图像并提取和筛选影像组学特征,进而用降维后得到的影像组学特征进行机器学习并建模。通过绘制受试者操作特征曲线并计算曲线下面积(AUC)来评估模型预测单纯新辅助化疗后对病理完全缓解(pCR)的效能。结果根据术后病理肿瘤退缩分级(TRG)患者被分为pCR组(TRG0,28例)和非pCR组(TRG1~TRG3,216例)。最终获得13个影像组学特征即6个灰度特征(均值、方差、标准差、偏离度、峰态、能量)、3个纹理特征(对比度、相关性、同质化)及4个形状特征(边界长度、直径、面积、形状参数)。基于CT的决策树模型预测LARC患者新辅助化疗后效果的AUC值为0.772 [95%CI(0.656,0.888)],对于非pCR的预测准确度较高(97.2%),但对于pCR的预测准确度较低(57.1%)。结论在本次初步探索中,基于CT的决策树模型在判断LARC患者新辅助化疗后pCR上较同质研究的预测效能低,通过分析后将进一步从优化算法、继续拓展数据集、挖掘更多影像组学特征值等方面优化模型,从而最终实现pCR的精准预测。Objective To explore the value of a decision tree(DT) model based on CT for predicting pathological complete response(pCR) after neoadjuvant chemotherapy therapy(NACT) in patients with locally advanced rectal cancer(LARC). Methods The clinical data and DICOM images of CT examination of 244 patients who underwent radical surgery after the NACT from October 2016 to March 2019 in the Database from Colorectal Cancer(DACCA) in the West China Hospital were retrospectively analyzed. The ITK-SNAP software was used to select the largest level of tumor and sketch the region of interest. By using a random allocation software,200 patients were allocated into the training set and44 patients were allocated into the test set. The MATLAB software was used to read the CT images in DICOM format and extract and select radiomics features. Then these reduced-dimensions features were used to construct the prediction model. Finally,the receiver operating characteristic(ROC) curve,area under the ROC curve(AUC),sensitivity,and specificity values were used to evaluate the prediction model. Results According to the postoperative pathological tumor regression grade(TRG) classification,there were 28 cases in the pCR group(TRG0) and 216 cases in the nonpCR group(TRG1–TRG3). The outcomes of patients with LARC after NACT were highly correlated with 13 radiomics features based on CT(6 grayscale features: mean,variance,deviation,skewness,kurtosis,energy;3 texture features:contrast,correlation,homogeneity;4 shape features: perimeter,diameter,area,shape). The AUC value of DT model based on CT was 0.772 [95% CI(0.656,0.888)] for predicting pCR after the NACT in the patients with LARC. The accuracy of prediction was higher for the non-PCR patients(97.2%),but lower for the pCR patients(57.1%). Conclusions In this preliminary study,the DT model based on CT shows a lower prediction efficiency in judging pCR patient with LARC before operation as compared with homogeneity researches,so a more accurate prediction model of pCR patient will be optimized th

关 键 词：直肠癌 新辅助化疗 影像组学 人工智能 决策树 病理完全缓解 

分 类 号：R735.37[医药卫生—肿瘤] R730.44[医药卫生—临床医学]