国际肿瘤学杂志

国际肿瘤学杂志 ›› 2020, Vol. 47 ›› Issue (10): 593-597.doi: 10.3760/cma.j.cn371439-20200527-00084

• 论著 • 上一篇    下一篇

基于磁共振高分辨T2WI影像组学预测直肠癌新辅助治疗后病理完全反应的研究

陆海迪, 沈浮, 陆建平, 郝立强()   

  1. 海军军医大学附属长海医院医学影像科,上海 200433
  • 收稿日期:2020-05-27 修回日期:2020-07-13 出版日期:2020-10-08 发布日期:2020-11-20
  • 通讯作者: 郝立强 E-mail:hao_liqiang@139.com

Research on prediction of pathological complete response after neoadjuvant therapy for rectal cancer based on MRI high-resolution T2WI images

Lu Haidi, Shen Fu, Lu Jianping, Hao Liqiang()   

  1. Department of Medical Imaging, Changhai Hospital Affiliated to Naval Military Medical University, Shanghai 200433, China
  • Received:2020-05-27 Revised:2020-07-13 Online:2020-10-08 Published:2020-11-20
  • Contact: Hao Liqiang E-mail:hao_liqiang@139.com

RichHTML

51

PDF

359

摘要：

目的 探讨基于磁共振高分辨T2WI影像组学方法对预测直肠癌新辅助治疗后病理完全反应(pCR)的价值。方法 回顾性分析我院2018年1月至2019年3月新辅助治疗前接受磁共振高分辨T2WI成像检查并经病理证实的80例直肠癌患者,在高分辨T2WI图像上手动勾画病灶容积感兴趣区(VOI)后提取影像组学特征,采用最小绝对值收缩算子(LASSO)算法进行降维,筛选对肿瘤pCR有价值的特征,利用Random算法将数据随机分为训练集(n=64)与测试集(n=16)进行机器学习,建立决策树(DT)、逻辑回归(LR)、随机森林(RF)、极限梯度增强树(XGBoost)4种机器学习模型并绘制ROC曲线,分别计算AUC、敏感性、特异性及95%CI,采用DeLong检验比较ROC曲线差异。结果 80例直肠癌患者pCR 15例,占18.75%;非pCR 65例,占81.25%。共提取1 409个影像组学特征,经LASSO算法降维后筛选出8个最有价值的特征。测试集DT、LR、RF、XGBoost 4种分类器模型的AUC分别为0.870、0.801、0.912、0.945,其中XGBoost分类器模型的AUC最大,与DT、LR、RF分类器模型相比较,差异具有统计学意义(P=0.008;P=0.006;P=0.009);其他3种模型两两比较,差异均无统计学意义(PLR-RF=0.083;PDT-LR=0.113;PDT-RF=0.879)。4种分类器模型敏感性分别为78.57%、64.29%、78.57%、85.71%,特异性分别为95.38%、84.62%、92.31%、98.46%,95%CI分别为0.775~0.935、0.696~0.882、0.827~0.964、0.870~0.984。结论 基于高分辨T2WI图像的影像组学对直肠癌新辅助治疗后pCR有预测价值,其中XGBoost模型预测效能优于DT、LR、RF,可以用于辅助临床制定个体化治疗决策。

关键词: 直肠肿瘤, 磁共振成像, 新辅助治疗, 影像组学

Abstract:

Objective To explore the value of MRI high-resolution T2WI based-radiomics in predicting pathologic complete response (pCR) after neoadjuvant therapy for rectal cancer. Methods This retrospective study included 80 patients with rectal cancer confirmed by postoperative pathology, who underwent high-resolution imaging of rectal MRI before neoadjuvant therapy from January 2018 to March 2019 in our hospital. After manually delineating the volume of interest (VOI) of the lesion in the high-resolution T2WI image, the radiomics features were extracted, and the least absolute shrinkage and selection operator (LASSO) algorithm was adopted to reduce the dimension and select the features that were valuable for tumor pCR. Using Random algorithm, the data were randomly divided into training set (n=64) and test set (n=16) for machine learning, and 4 kinds of machine learning models including decision tree (DT), logistic regression (LR), random forests (RF) and extreme gradient boosting (XGBoost) were established and ROC curves were drawn. The area under the curve (AUC), sensitivity, specificity and 95%CI were respectively calculated, and the difference of ROC curves was compared with DeLong test. Results Among 80 patients with rectal cancer, there were 15 cases by pCR, accounting for 18.75%, and 65 cases were non-pCR, accounting for 81.25%. A total of 1 409 imaging features were extracted. After dimension reduction by LASSO algorithm, 8 most valuable features were selected. The AUC of DT, LR, RF and XGBoost in the test set group was 0.870, 0.801, 0.912, 0.945, the AUC of XGBoost was the largest, and the differences between XGBoost and DT, LR, RF were statistically significant (P=0.008; P=0.006; P=0.009), and the pairwise comparisons of DT, LR, RF showed no statistically significant difference (PLR-RF=0.083; PDT-LR=0.113; PDT-RF=0.879). The sensitivity was 78.57%, 64.29%, 78.57%, 85.71%, and the specificity was 95.38%, 84.62%, 92.31%, 98.46% respectively. The 95%CI was 0.775-0.935, 0.696-0.882, 0.827-0.964, 0.870-0.984. Conclusion The radiomics based on high-resolution T2WI images has predictive value for pCR after neoadjuvant treatment of rectal cancer. XGBoost model has better predictive efficiency than DT, LR and RF, and can be used to guide clinical individualized treatment and related interventions.

Key words: Rectal neoplasms, Magnetic resonance imaging, Neoadjuvant therapy, Radiomics