多参数MRI影像组学在直肠黏液腺癌新辅助治疗疗效预测中的应用价值

The value of multi-parametric MRI radiomics in the prediction of neoadjuvant therapy for rectal mucinous adenocarcinoma

目的探讨基线期MRI多序列影像组学在预测直肠黏液腺癌(RMAC)新辅助放化疗(NCR)疗效的应用价值。方法回顾性分析2012年8月至2018年10月中山大学附属第六医院在NCR前行MRI检查的RMAC患者的临床和影像资料。共纳入79例患者,男52例,女27例,年龄20~78岁,中位年龄52岁。根据病理消退分级标准,将所有患者分为NCR有效组(n=31)和NCR无效组(n=48)。分别提取基线期MRI的T2WI、扩散加权成像(DWI)和增强T1WI图像的701个影像组学特征,并通过可重复性分析和特征降维筛选出特征子集构建影像组学预测模型。比较NCR有效组和无效组基线期MRI影像特征,将P<0.05的特征与影像组学结合构建模型。以病理为金标准,采用受试者操作特征(ROC)曲线评价预测模型的诊断效能,并计算曲线下面积(AUC)、95%可信区间、灵敏度和特异度,并采用DeLong法比较不同预测模型的诊断效能。结果NCR有效组和无效组常规影像表现比较,淋巴结分期和有无黏液结节差异有统计学意义(χ^²=6.040、5.870,P<0.05)。基于T2WI、DWI、增强T1WI影像组学ROC曲线的AUC分别为0.816、0.821和0.819,均高于常规特征(淋巴结分期、黏液结节状态)的AUC(0.607),且差异均有统计学意义(Z=-2.391、-2.580、-2.717,P<0.05)。T2WI、DWI及增强T1WI影像组学特征联合常规特征预测模型中,DWI联合常规特征的AUC最大,为0.843,且3种联合模型之间差异均无统计学意义(P均>0.05)。结论基线期T2WI、DWI、增强T1WI影像组学构建模型可以预测RMAC行NCR后疗效,优于常规特征预测效能;且联合常规特征后可进一步提升预测效能。

Objective To investigate the application value of baseline MRI multi-parametric imaging radiomics in prediction of neoadjuvant chemoradiotherapy(NCR)efficacy of rectal mucinous adenocarcinoma(RMAC).Methods Retrospective analysis was performed in the Sixth Affiliated Hospital of Sun Yat-sen University from August 2012 to October 2018.A total of 79 patients were included in this study,including 52 males and 27 females,aged 20-78 years(median age 52 years).According to the classification criteria of pathological regression,all patients were divided into NCR responsiveness group(n=31)and nonresponsiveness group(n=48).And 701 imaging features of T2WI,diffusion weighted imaging(DWI)and enhanced T1WI images of baseline MRI were extracted,and feature subsets were selected by repeatability analysis and feature dimensionality reduction to construct the radiomics prediction model.The tumor features from baseline MRI between the NCR responsiveness group and the nonresponsiveness group were compared,and the features of P<0.05 were combined with the radiomics to construct a model.Using pathology as the gold standard,the receiver operating characteristic(ROC)curve was used to evaluate the diagnostic efficiency of the prediction model,and the area under the curve(AUC),95%confidence interval,sensitivity and specificity were calculated,and the DeLong test was used to compare the diagnostic efficacy of different prediction models.Results By comparing the conventional tumor imaging characteristics of the NCR responsiveness group and the nonresponsiveness group,the differences in lymph node stage and mucinous nodule status between the two groups were statistically significant(χ^²=6.040,5.870,P<0.05).The AUC of ROC curves based on T2WI,DWI,and enhanced T1WI radiomics were 0.816,0.821,and 0.819,respectively,which were higher than those of conventional tumor characteristics(lymph node staging,mucinous nodule status)(AUC=0.607),and the differences were statistically significant(Z=-2.391,-2.580 and-2.717,P<0.05).Among the joint prediction models of T2WI,DWI and contrast-enhanced T1WI radiomics and conventional tumor features,the DWI combined model had the largest AUC(0.843),and there was no statistically significant difference between the three combined models(all P>0.05).Conclusion The baseline T2WI,DWI,and contrast-enhanced T1WI radiomics model can be used to predict the NCR efficacy of RMAC,which is better than the predictive efficacy of conventional features,and the combination with conventional features can further improve the predictive efficacy.