基于IVIM-DWI定量参数的影像组学预测鼻咽癌短期疗效

Radiomics analysis based on IVIM-DWI quantitative parameters to predict the short-term therapeutic effect of nasopharyngeal carcinoma

目的基于治疗前鼻咽癌体素内不相干运动扩散加权成像(intravoxel incoherent motion diffusion-weighted imaging,IVIM-DWI)影像组学定量特征,建立用于预测鼻咽癌短期疗效的预测模型。材料与方法回顾性收集2019年1月至2021年8月期间于海南省人民医院放疗科接受治疗的首程病理确诊鼻咽癌患者80例。在治疗前均行头颈部MRI平扫+增强检查及11个b值(区间0~800 s/mm2)IVIM-DWI检查,在接受以放疗为主的综合治疗后每3个月进行头颈部常规MRI随访,依据实体肿瘤的治疗反应评价标准1.1版在治疗后6个月的MRI随访图像上行疗效评价,将患者分为完全缓解组(n=62)及非完全缓解组(n=18)。IVIM-DWI经双指数模型后处理计算得到真实扩散系数(true diffusion coefficient,D)、灌注相关扩散系数(perfusion related diffusion coefficient,D^(*))和灌注分数(perfusion fraction,f)的定量参数图像。使用ITK-SNAP软件在IVIM-DWI的S0图像上逐层勾画病灶的感兴趣区(region of interest,ROI),鼻咽癌常规及增强MRI图像作为定位参照。应用3D slicer软件分别在D、D^(*)和f定量参数图像上相应的ROI区域提取影像组学特征,包括直方图特征、纹理特征以及基于形态学的组学特征。使用最小绝对收缩和选择算子算法筛选出与疗效高度相关的影像组学特征,采用逻辑回归方法分别构建基于D、D^(*)、f和联合参数的影像组学预测模型,预测性能采用ROC曲线、曲线下面积(area under the curve,AUC)及校正曲线评估,并使用决策曲线分析(decision curve analysis,DCA)评价预测模型的临床实用性。10次十折交叉验证被使用于模型内部验证,计算平均敏感度及特异度。结果共计851个影像组学特征被提取,经过特征筛选后,筛选出2个D值特征,构建的影像组学模型的敏感度为60.0%,特异度为79.6%,AUC值为0.734;筛选出2个f值特征,构建的影像组学模型敏感度为66.1%,特异度为76.3%,AUC值为0.747;筛选出1个D^(*)值特征,构建的影像组学模型敏感度为76.1%,特异度为75.9%,AUC值为0.726;联合以上3种参数共5个影像组学特征,构建的联合影像组学模型敏感度为81.7%,特异度为80.6%,AUC值为0.827。校正曲线显示各模型均具有良好的拟合优度,DCA显示4种模型均具有良好的临床效益,而IVIM联合模型的临床效益最高。结论基于IVIM-DWI参数建立的影像组学模型能够在治疗前较好地预测鼻咽癌患者的治疗反应性。其中,效能最高的是IVIM-DWI联合参数模型,该模型可以为患者的临床诊疗决策提供帮助。

Objective:To establish a predictive model based on quantitative characteristics of intravoxel incoherent motion diffusion-weighted imaging(IVIM-DWI)radiomics to predict short-term treatment efficacy of nasopharyngeal carcinoma before treatment.Materials and Methods:A retrospective study was conducted to collect 80 patients with nasopharyngeal carcinoma diagnosed pathologically at the first visit who were treated in the Radiotherapy Department of Hainan Provincial People's Hospital from January 2019 to August 2021.Before treatment,all subjects underwent MRI plain scan+enhanced examination and 11 b-value(interval 0-800 s/mm2)IVIM-DWI examination.After receiving comprehensive treatment based on radiotherapy,routine MRI follow-up of head and neck was conducted every 3 months.Use MRI follow-up images taken 6 months after the end of treatment for efficacy evaluation.According to respond evaluation criteriain solid tumors,version 1.1 standard,the patients were divided into complete remission group(n=62)and incomplete remission group(n=18).The real diffusion coefficient(D),perfusion related diffusion coefficient(D^(*))and perfusion fraction(f)were obtained by post-processing IVIM-DWI with a double exponential model.Itk-snap was used to delineate the region of interest(ROI)of the lesion layer by layer on the S0 image of IVIM-DWI,and the conventional and enhanced MRI images of nasopharyngeal carcinoma were used as the positioning reference.3D slicer software was used to extract radiomics features,including histogram features,texture features and morphology features,from the corresponding ROI regions of D,D^(*)and f quantitative parameter images.The least absolute shrinkage and selection operator algorithm was used to screen out the radiomics features that were highly correlated with the treatment effect.Logistic regression was used to construct radiomics prediction models based on D,D^(*),f,and joint parameters,and predictive performance was evaluated using ROC curves,area under the curve(AUC),and calibration curves.Decision curve analysis(DCA)was used to evaluate the clinical utility of the prediction models.A 10-fold cross-validation was used for internal model validation,and the average sensitivity and specificity were calculated.Results:A total of 851 radiomics features were extracted,and after feature selection,two D-value features were selected to construct a radiomics model with a sensitivity of 60.0%,specificity of 79.6%,and an AUC value of 0.734.Two f-value features were selected to construct a radiomics model,with a sensitivity of 66.1%,specificity of 76.3%,and an AUC value of 0.747.One D^(*)-value feature was selected to construct a radiomics model,with a sensitivity of 76.1%,specificity of 75.9%,and an AUC value of 0.726.The sensitivity,specificity and AUC of the radiomics model based on the three types of IVIM-DWI radiomics features were 81.7%,80.6%and 0.827 respectively.The calibration curves showed good goodness-of-fit for all models,and the DCA demonstrated good clinical utility for all four models,with the IVIM joint model showing the highest clinical benefit.Conclusions:The radiomics model based on IVIM-DWI parameters can predict the therapeutic response of nasopharyngeal carcinoma patients before treatment.Among them,the most effective model is the IVIM-DWI combined parameter model,which can assist in clinical decision-making for patients.