【作者】 张巍；

【导师】 宋彬；

【作者基本信息】 四川大学 ， 影像医学与核医学（专业学位）， 2021， 博士

【摘要】 目的:微卫星不稳定(microsatellite instability,MSI)的直肠癌患者预后较好,不能从基于5-FU的化疗中获益,且对免疫治疗反应良好。有癌周肿瘤沉积(tumor deposits,TDs)的直肠癌患者复发率和转移率更高,被视为是预后不良的重要因素。因此,MSI和TDs都是决定直肠癌患者的治疗反应以及预后的重要生物学标记,但现有检查手段存在有创、方法步骤繁琐导致难以普及应用以及无法避免肿瘤异质性影响等缺点。本研究拟分别采用影像组学和深度学习方法,对直肠癌患者术前临床指标和磁共振(magnetic resonance imaging,MRI)图像特征进行分别和联合建模,旨在术前无创性预测直肠癌患者的MSI状态或有无TDs存在,以期为直肠癌患者的精准治疗和预后评价提供参考。材料和方法:回顾性收集:1.491例经病理证实的直肠癌患者,根据其错配修复(mismatchrepair,MMR)蛋白的表达情况,分为MSI组51例,微卫星稳定(microsatellite stability,MSS)组440例。将全部入组病例随机分为训练组(共327例,其中MSS患者291例,MSI患者36例)和验证组(共164例,其中MSS患者149例,MSI患者15例);2.455例经病理证实的直肠癌患者,根据其术后癌周结节病理表现,分为TDs组78例,non-TDs组377例。将全部入组病例随机分为训练组(共364例,其中TDs患者60例,non-TDs患者304例)和验证组(共91例,其中TDs患者18例,non-TDs患者73例);收集上述患者性别、年龄、MR-T分期、CEA和CA19-9水平等指标用于构建临床指标模型;收集上述患者直肠高分辨率T2WI序列MRI图像用于构建影像模型。影像组学研究经过肿瘤分割、特征提取、一致性检验、特征的双样本t检验和LASSO回归降维之后,构建并验证临床指标模型、影像组学模型和整合了临床变量与影像组学特征的联合模型,绘制受试者操作特征曲线(receiver operating characteristic,ROC),并计算曲线下面积(area under the curve,AUC)值、敏感性和特异性等指标来评价上述模型的预测效能。深度学习研究经过肿瘤分割、图像强度归一化、影像数据增强、预训练及迁移学习之后,采用MobileNetV2 网络为骨架的深度学习模型来分别对直肠癌的MSI与MSS、TDs与non-TDs进行分类,构建并验证临床指标模型、影像组学模型和上述两者的联合模型,计算ROC曲线的AUC值、敏感性、特异性等指标来反应上述模型的预测效能。采用Delong检验分别对影像组学和深度学习两种方法之内以及方法之间的预测模型进行对比,了解各个模型预测效能的差异。结果:1.预测直肠癌MSI方面:(1).影像组学研究通过特征提取和降维后,筛选出6个关键性特征。临床指标模型、影像组学模型和联合模型在训练组的AUC值(95%置信区间)分别为 0.756(0.705-0.801)、0.945(0.914-0.967)、0.989(0.970-0.997),在验证组的 AUC 值(95%置信区间)分别为 0.685(0.608-0.755)、0.784(0.713-0.844)、0.895(0.838-0.938),敏感性分别为:60%、60%、66.70%,特异性分别为98%、97.30%、98.70%,Delong检验提示联合模型与临床指标模型存在显著性差异(P=0.015),与影像组学模型无显著性差异(P=0.204),影像组学模型与临床指标模型无显著性差异(P=0.446);(2).深度学习研究中临床指标模型、影像模型和联合模型在训练组的AUC值(95%置信区间)分别为0.756(0.705-0.801)、0.937(0.905-0.961)、0.955(0.926-0.975),在验证组的 AUC 值(95%置信区间)分别为 0.685(0.608-0.755)、0.823(0.756-0.878)、0.874(0.814-0.921),敏感性分别为:60%、60%、66.67%,特异性分别为 97.99%、96.64%、95.97%。Delong检验提示联合模型与临床指标模型之间有统计学差异(P=0.043),其余模型之间无统计学差异(P>0.05)。2.预测直肠癌TDs方面:(1).影像组学研究通过特征提取和降维后,筛选出15个关键性特征。临床指标模型、影像组学模型和联合模型在训练组的AUC值(95%置信区间)分别为 0.820(0.776-0.858)、0.858(0.818-0.893)、0.899(0.863-0.928),在验证组的 AUC 值(95%置信区间)分别为 0.731(0.627-0.818)、0.742(0.640-0.828)、0.820(0.725-0.892),敏感性分别为:72.22%、55.56%、66.67%,特异性分别为63.01%、84.93%、86.30%,Delong检验提示验证组中三组模型相互之间均无统计学差异(P>0.05)。(2).深度学习研究中临床指标模型、影像模型和联合模型在训练组的AUC值(95%置信区间)分别为0.820(0.776-0.858)、0.871(0.833-0.904)、0.894(0.858-0.924),在验证组的 AUC值(95%置信区间)分别为 0.731(0.627-0.818)、0.804(0.707-0.880)、0.848(0.757-0.915),敏感性分别为:72.22%、83.33%、77.78%,特异性分别为 63.01%、68.49%、94.52%,Delong检验提示验证组中三组模型相互之间均无统计学差异(P>0.05)。3.在预测直肠癌MSI和TDs中,Delong检验显示影像组学与深度学习两种方法构建的预测模型无显著统计学差异(P>0.05)。结论:基于MRI数据的影像组学和深度学习方法可以在术前有效分类直肠癌的MSI与MSS、TDs与non-TDs状态,结合了临床指标和影像数据的联合模型预测性能最好,该方法有希望为直肠癌患者的精准治疗和预后评价提供依据。更多还原

【Abstract】 Objective:Rectal cancer patients with micro satellite instability(MSI)have a better prognosis,do not respond to 5-Fu based chemotherapy,and respond well to immunotherapy.Rectal cancer patients with tumor deposits(TDs)have higher rates of recurrence and metastasis,which is regarded as an important factor for poor prognosis.Therefore,both MSI and TDs are important biomarkers that determine the treatment response and prognosis of patients with rectal cancer.However,existing screening methods have disadvantages such as invasive,complicated procedures,difficulty in popularization and application,and inability to avoid the effects of tumor heterogeneity.In this study,preoperative prediction models are intended to be established using radiomics and deep learning methods combined with the clinical variables and imaging features of patients with rectal cancer,respectively,aiming to noninvasively predict the MSI status or TDs of patients with rectal cancer preoperatively,in order to provide reference for accurate treatment and prognosis evaluation of patients with rectal cancer.Materials and Methods:Using a retrospective study.1.491 patients with rectal cancer confirmed by pathology were divided into MSI group(n=51)and microsatellite stability(MSS)group(n=440)according to the expression of MMR protein.They were randomly divided into training group(n=327,including 291 cases of MSS and 36 cases of MSI)and validation group(n=164,including 149 cases of MSS and 15 cases of MSI).2.According to the postoperative pathological features of pericancerous nodules,455 patients with pathologically confirmed rectal cancer were divided into TDs group(n=78)and non-TDs group(n=377).In addition,the patients were randomly divided into training group(n=364,60 patients with TDs and 304 patients with non-TDs)and validation group(n=91,18 patients with TDs and 73 patients with non-TDs).Gender,age,MR-T stage,CEA and CA19-9 levels of the above patients were collected to construct a clinical model.The rectal high-resolution T2WI sequence MRI images of the above patients were collected to construct image models.After tumor segmentation,feature extraction,consistency test,characteristic two-sample t-test and LASSO regression-based dimension reduction of radiomics features,the clinical model,radiomics model and combined model integrating clinical variables and radiomics features were constructed and validated,and the area under the receiver operating characteristic(ROC)curve,sensitivity and specificity were calculated to reflect the prediction efficiency of the above models.After tumor segmentation,image intensity normalization,image data enhancement,pre-training and migration learning of deep learning research,the MSI and MSS as well as TDs and non-TDs of rectal cancer were classifies using the deep learning model based on MobileNetV2 network as the framework,the clinical model,radiomics model and their combined model were constructed and validated,and the AUC,sensitivity and specificity were calculated to reflect the prediction efficiency of the above models.Delong’s test was performed to compare the prediction models within and between the radiomics and deep learning respectively to explore the difference in the prediction efficiency of each model.Results:In the aspect of predicting MSI of rectal cancer:(1).After feature extraction and dimension reduction of radiomics research,6 key features were screened out.The AUC[95%confidence interval(CI)]of the clinical model,radiomics model and combined model was 0.756(0.705-0.801),0.945(0.914-0.967)and 0.989(0.970-0.997)in the training group,and 0.685(0.608-0.755),0.784(0.713-0.844)and 0.895(0.838-0.938)in the validation group,respectively.The sensitivity was 60%,60%and 66.70%,and the specificity was 98%,97.30%and 98.70%,respectively.Delong test showed a significant difference between the combined model and the clinical model(P=0.015,Z=2.421).But no significant difference between the combined and the radiomics model(P=0.204)as well as between the radiomics and the clinical model(P=0.446).(2).In deep learning research,the AUC(95%CI)of the clinical model,radiomics model and combined model was 0.756(0.705-0.801),0.937(0.905-0.961)and 0.955(0.926-0.975)in the training group,and 0.685(0.608-0.755),0.823(0.756-0.878)and 0.874(0.814-0.921)in the validation group,respectively.The sensitivity was 60%,60%and 66.67%,and the specificity was 97.99%,96.64%and 95.97%,respectively.Delong’s test revealed a significant difference between the combined model and the clinical model(P=0.043),but no significant difference between the rest models(P>0.05).In the aspect of predicting TDs of rectal cancer:(1).After feature extraction and dimension reduction of radiomics research,15 key features were screened out.The AUC(95%CI)of the clinical model,radiomics model and combined model was 0.820(0.776-0.858),0.858(0.818-0.893)and 0.899(0.863-0.928)in the training group,and 0.731(0.627-0.818),0.742(0.640-0.828)and 0.820(0.725-0.892)in the validation group,respectively.The sensitivity was 72.22%,55.56%and 66.67%,and the specificity was 63.01%,84.93%and 86.30%,respectively.Delong test showed no significant difference between the rest models(P>0.05).(2).In deep learning research,the AUC(95%CI)of the clinical model,radiomics model and combined model was 0.820(0.776-0.858),0.871(0.833-0.904)and 0.894(0.858-0.924)in the training group,and 0.731(0.627-0.818),0.804(0.707-0.880)and 0.848(0.757-0.915)in the validation group,respectively.The sensitivity was 72.22%,83.33%and 77.78%,and the specificity was 63.01%,68.49%and 94.52%,respectively.Delong’s test showed no significant difference between the rest models(P>0.05).In the prediction of MSI and TDs of rectal cancer.Delong’s test showed that there was no significant difference between the models constructed by radiomics and deep learning methods(P>0.05).Conclusion:Radiomics and deep learning methods based on MRI data can effectively classify MSI and MSS as well as TDs and non-TDs of rectal cancer before surgery.The combined model integrating clinical and imaging data has the best prediction performance.This method is expected to provide help for individualized treatment planning and prognostic evaluation of patients with rectal cancer.更多还原