波谱学杂志 ›› 2022, Vol. 39 ›› Issue (1): 43-55.doi: 10.11938/cjmr20212908

• 研究论文 • 上一篇    下一篇

基于影像组学的直肠癌术前T分期预测

王楠1,王远军1,*(),廉朋2,*()   

  1. 1. 上海理工大学 医学影像工程研究所, 上海 200093
    2. 复旦大学附属肿瘤医院 大肠外科, 复旦大学上海医学院 肿瘤学系, 上海 200032
  • 收稿日期:2021-04-15 出版日期:2022-03-05 发布日期:2021-07-14
  • 通讯作者: 王远军,廉朋 E-mail:yjusst@126.com;lianpeng_crcc@163.com

Prediction of Preoperative T Staging of Rectal Cancer Based on Radiomics

Nan WANG1,Yuan-jun WANG1,*(),Peng LIAN2,*()   

  1. 1. Institute of Medical Imaging Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
    2. Department of Colorectal Surgery, Fudan University Shanghai Cancer Center; Department of Oncology, Shanghai Medical College, Fudan University, Shanghai 200032, China
  • Received:2021-04-15 Online:2022-03-05 Published:2021-07-14
  • Contact: Yuan-jun WANG,Peng LIAN E-mail:yjusst@126.com;lianpeng_crcc@163.com

摘要:

直肠癌T分期对患者的术前评估有重要作用.然而,传统的放射科医生根据患者磁共振图像直接判断分期的方法效果欠佳.本文提出使用影像组学的方法预测直肠癌T分期,首先获取105例直肠癌患者影像数据,根据病理报告中的T分期结果将T1、T2期患者划分为未突破肌层组,将T3、T4期患者分为突破肌层组,整理数据得到未突破肌层组31例,突破肌层组74例.在患者的轴向位T2WI图像中勾画病灶区域,并在病灶上使用pyradiomics工具包提取影像组学特征,使用最小绝对值收敛和选择算子(LASSO)对高维特征做特征选择,得到与T分期高度相关的特征数据,使用随机森林、支持向量机(SVM)、逻辑回归、梯度提升树(GBDT)分别建模,进行交叉验证调参,评估模型性能.每层图像提取100维特征,经LASSO特征选择后得到7个与T分期高度相关的特征,使用4种模型分别建模,其中SVM算法表现最优,平均受试者操作特征曲线下面积(AUC)、准确率、灵敏度、特异度分别为0.968 5、0.886 4、0.962 5、0.899 2,测试集准确率达到了0.904 7.结果表明,使用影像组学方法可以提高直肠癌T分期预测的准确率.

关键词: 磁共振成像, 直肠癌分期, 机器学习, 支持向量机(SVM)

Abstract:

T staging plays an important role in the preoperative evaluation of rectal cancer. However, the traditional stage judging method directly based on the patients' MRI images is not effective. In this paper, we proposed to predict the T-stage of rectal cancer by using radiomics. First, the imaging data of 105 patients with rectal cancer were obtained, the patients in T1 and T2 stages were classified as non-breakthrough muscular layer group (31 cases), and the patients in T3 and T4 stages were classified as breakthrough muscular layer group (74 cases). In the axial T2WI image of patients, the region of interest (ROI) was segmented, and the radiomics features were extracted using the pyradiomics toolkit. The high-dimensional features were selected using least absolute shrinkage and selection operator (LASSO), and the feature data highly related to T stage were obtained. Four machine learning methods including logistic regression, support vector machine (SVM), gradient boosting decision tree (GBDT) and random forest were used in modeling respectively. Cross validation was performed to evaluate the performance of each model. 100 dimensional features were extracted from each image layer, and 7 features highly related to T stage were obtained after lasso feature selection. Among the four machine learning methods, SVM performed best. The average area under curve (AUC), accuracy, sensitivity and specificity of SVM method were 0.968 5, 0.886 4, 0.962 5 and 0.899 2 respectively, and the accuracy of verification set reached about 0.904 7. The result proved that radiomics can greatly improve the accuracy of T-stage prediction of rectal cancer.

Key words: magnetic resonance imaging, rectal cancer staging, machine learning, support vector machine (SVM)

中图分类号: