经直肠超声影像组学在术前预测直肠癌淋巴结转移中的应用  

Radiomic features of endorectal ultrasound for preoperative prediction of lymph node metastasis in rectal cancer

作　　者：李晋[1] 陈少娜 陈菲[1] 林云勇 卢文洁 吴怡雯 邱少东[1] Li Jin;Chen Shaona;Chen Fei;Lin Yunyong;Lu Wenjie;Wu Yiwen;Qiu Shaodong(Department of Ultrasonography,Second Affiliated Hospital of Guangzhou Medical University,Guangzhou 510260,China)

机构地区：[1]广州医科大学附属第二医院超声科,广东广州510260

出　　处：《广州医科大学学报》2021年第3期19-27,共9页Academic Journal of Guangzhou Medical University

基　　金：广州市科技计划项目(202102010049)。

摘　　要：目的:探讨经直肠超声的影像组学特征在术前预测直肠癌患者淋巴结转移状态中的诊断性能。方法:回顾性分析从2018年1月至2020年7月广州医科大学附属第二医院术前经直肠超声检查的35例直肠癌患者,根据组织病理学结果TMN分期将患者分为阳性组(N1-2)和阴性组(N0)。基于患者经直肠超声图像,提取直肠癌矢状面超声图像的影像组学特征。使用独立样本t检验和最小绝对收缩和选择算子回归(LASSO回归)进行特征选择。根据筛选后的特征构建随机森林模型,以评估患者的淋巴结转移状态。使用K折交叉验证法评估该模型的诊断性能,绘制基于该模型的ROC曲线。结果:共提取了1693个直肠癌纹理特征,经t检验后有109个特征差异具有统计学意义(P<0.05),LASSO回归后得到最优的11个纹理特征。其中,从N1-2期直肠肿瘤图像中提取的WLGC、WLGL、WHGD、WHGS、WLGSZ显着高于N0期,而ONS、L3MFE、L3MFT、WLGG、WLGSD、WLGSR则低于N0期。权重图显示WLGL对于区分患者N分期更为重要。基于这11个特征建立的随机森林分类模型的10次验证平均ROC曲线的AUC值为71%。结论:基于术前经直肠超声检查的直肠癌影像组学特征对判断淋巴结转移状态具有重要的参考价值。利用这些特征所建立的随机森林模型在术前判断淋巴结转移状态方面存在一定的价值,有助于制定不同的治疗策略。Objective:To investigate the diagnostic performance of radiomic features of endorectal ultrasound(ERUS)for preoperative prediction of lymph node metastasis in rectal cancer.Methods:Retrospectively included in this study were 35 patients with rectal cancer who underwent preoperative ERUS in Second Affiliated Hospital of Guangzhou Medical University between January 2018 and July 2020.According to histopathological findings,the patients were divided into the positive group(N1-2)and negative group(N0).Based on the patients’ERUS images,radiomic features were extracted from the sagittal ultrasound image of rectal cancer.Independent sample t-test and least absolute shrinkage and selection operator(LASSO)regression were used for feature selection.A random forest model was constructed based on the screened features to assess the lymph node metastasis in the patients.K-fold cross-validation was used to evaluate the diagnostic performance of the model.ROC curves of the model were generated.Results:In this study,1693 texture features were extracted from rectal cancers.Statistically significant differences were found in 109 features by t-test.LASSO regression identified 11 texture features to be optimal.Among them,the values of WLGC,WLGL,WHGD,WHGS,and WLGSZ extracted from images of stages N1-2 rectal cancer were significantly higher than those from images of stage N0 cancers,while the values of ONS,L3MFE,L3MFT,WLGG,WLGSD,and WLGSR were lower on images of stages N1-2 rectal cancer compared with stage N0.According to weight map,WLGL seemed to be more important for the N staging of patients.The AUC of mean ROC curve was 71%after 10-fold cross-validation of the random forest model constructed with these 11 features.Conclusion:The radiomic features of rectal cancer based on preoperative ERUS have important implications in determining lymph node metastasis.A random forest model established using these features can be useful in preoperative prediction of lymph node metastasis and can help formulate relevant treatment strategies.

关 键 词：经直肠超声 影像组学 直肠癌 淋巴结转移 

分 类 号：R445.1[医药卫生—影像医学与核医学]