基于增强T_(1)WI影像组学预测直肠腺癌病理分级  

The contrast-enhanced T_(1)WI radiomics for predicting pathological grade in rectal adenocarcinoma

作　　者：汪博泉 郭小芳[1] 肖峰 聂婷婷 袁子龙[1] 刘玉林[1] WANG Boquan;GUO Xiaofang;XIAO Feng;NIE Tingting;YUAN Zilong;LIU Yulin(Department of Radiology,Hubei Cancer Hospital,Wuhan 430079,China;Department of Radiology,Zhongnan Hospital of Wuhan University,Wuhan 430071,China)

机构地区：[1]湖北省肿瘤医院放射科,湖北武汉430079 [2]武汉大学中南医院放射科,湖北武汉430071

出　　处：《实用放射学杂志》2024年第8期1286-1290,共5页Journal of Practical Radiology

基　　金：国家癌症中心攀登基金临床研究重点课题(NCC201917B05);湖北省肿瘤医院生物医学中心专项科研基金项目(2022SWZX06)。

摘　　要：目的探讨基于增强T_(1)WI影像组学预测直肠腺癌病理分级的可行性。方法回顾性分析127例直肠腺癌患者的MRI及病理资料,使用ITK-SNAP软件在轴位T_(1)WI增强图像上,手动绘制直肠癌整个瘤灶作为感兴趣区(ROI),使用Pyradiomics软件提取ROI内所有影像组学特征。将目标任务分解为2部分:任务1预测肿瘤的高分化与中-低分化(“高-非高”组);任务2预测中-低分化肿瘤中的中分化与低分化(“中-低”组)。采用最大相关最小冗余(mRMR)方法筛选特征,采用最小绝对收缩和选择算子(LASSO)、逻辑回归(LR)、朴素贝叶斯(NB)、随机森林(RF)、支持向量机(SVM)5种方法分别构建模型,并评估比较各模型效能。结果任务1中5种模型的曲线下面积(AUC)在训练集中分别为0.86、0.90、0.59、1.00、0.99,在测试集中分别为0.71、0.62、0.53、0.67、0.64,任务2中5种模型在训练集中AUC分别为0.93、0.85、0.67、0.92、0.89,在测试集中分别为0.86、0.80、0.50、0.78、0.71。2个任务中LASSO构建的模型均为优势模型。结合年龄、性别与优势模型影像组学评分(Radscore)组成的融合模型,得到测试集中任务1的AUC为0.80[95%置信区间(CI)0.63~0.96],准确性、敏感度、特异度分别为78.94%、77.78%、79.31%;任务2则分别为0.89(95%CI 0.74~1.00)、90.00%、95.65%、71.43%。校准曲线表明融合模型的拟合优度良好。结论通过2个二分类模型的建立,基于增强T_(1)WI影像组学方法在预测直肠腺癌的高、中、低分化程度方面具有可行性。Objective To investigate the feasibility of using contrast-enhanced T_(1)WI radiomics in predicting the pathological grade in rectal adenocarcinoma.Methods The MRI and pathological data of 127 patients with rectal adenocarcinoma were analyzed retrospectively.ITK-SNAP software was used to manually draw region of interest(ROI)in rectal cancer on axial T_(1)WI enhanced images.The radiomics features were extracted by the Pyradiomics software from ROI.The task was divided into two parts:task 1(“high&non-high”group)predicted the high-differentiation and moderate/low-differentiation of the tumor;task 2(“moderate&low”group)predicted the tumor’s moderate-differentiation and low-differentiation in“non-high”group.Maximum relevance and minimum redundancy(mRMR)method was used to screen features.The five methods including least absolute shrinkage and selection operator(LASSO),logistic regression(LR),naive Bayes(NB),random forest(RF),and support vector machine(SVM)were used to build the models,and the efficiency of each model was evaluated and compared.Results In task 1,the area under the curve(AUC)of five methods were 0.86,0.90,0.59,1.00,0.99 in the training cohort and 0.71,0.62,0.53,0.67,0.64 in the testing cohort.In task 2,the AUC of five methods in the training cohort were 0.93,0.85,0.67,0.92,0.89,and in the testing cohort were 0.86,0.80,0.50,0.78,0.71.The models constructed by LASSO in both tasks were the dominant models,the AUC of the fusion model in the testing cohort which combined with age,gender and the dominant Radiomics score(Radscore)was 0.80[95%confidence interval(CI)0.63-0.96]in task 1,and the accuracy,sensitivity and specificity were 78.94%,77.78%,and 79.31%respectively.They were 0.89(95%CI 0.74-1.00),90.00%,95.65%,and 71.43%,respectively in task 2.The calibration curves showed that the fusion models had a good goodness of fit.Conclusion Based on the establishment of two dichotomous models,the radiomics based on the contrast-enhanced T_(1)WI is feasible in predicting the high,moderate and low diffe

关 键 词：磁共振成像 影像组学 直肠腺癌 病理分级 

分 类 号：R445.2[医药卫生—影像医学与核医学] R445[医药卫生—诊断学] R735.37[医药卫生—临床医学]