基于CT影像组学的非微卫星高度稳定II期结直肠癌术后化疗疗效预测研究

There is still a controversy over whether patients with stage II colorectal cancer could benefit from postoperative fluorouracil based adjuvant chemotherapy. Although it has been firmly proved that microsatellite instability-high stage II patients with cannot benefit from chemotherapy, for other majority (approximately 85%) of patients showed nonmicrosatellite instability-high (i.e., microsatellite stability or instability-low) expression, there is still lack of an effective predictor or tool for screening patients with survival benefit from adjuvant chemotherapy. Based on statistics and machine learning, the recently emerging radiomics can comprehensively characterize tumor heterogeneity by deeply mining the massive imaging features in conventional images, which can be used for the prediction of staging, response and prognosis, as well as providing a new method for predicting of these patients with survival benefit..In this study, we aim to develop a predictive model for personalized evaluation of survival benefit from adjuvant chemotherapy in nonmicrosatellite instability-high stage II colorectal cancer patients and help precision stratification treatment by following the next steps. First, extracting massive features based on the preoperative CT data through manual and deep learning methods; then debasing dimension of features and delivering a radiomics signature highly associated with prediction; at last fusing the radiomics signature with clinicopathological features associated with prediction of chemotherapy to model.

II期结直肠癌患者根治性术后能否从氟尿嘧啶为基础的辅助化疗中获益仍存有争议。虽然对于微卫星高度稳定状态患者化疗不能获益已得以证实，但对于大多数（约85%）非微卫星高度不稳定状态（微卫星稳定/微卫星低度不稳定）患者，目前尚缺乏有效的疗效预测因子或工具用于化疗获益人群的筛选。近年新兴的影像组学方法以统计学及机器学习为基础，通过深度挖掘常规影像图像中的海量特征，全面表征肿瘤异质性，可用于分期、疗效及预后等预测研究，为化疗生存获益预测提供了新的思路。本项目拟以术前CT图像为基础，结合手工及深度学习方法提取特征，筛选并建立与疗效高度相关的影像组学特征标签，并利用机器学习及统计学方法，构建融合影像组学与临床病理特征的疗效预测模型，为非微卫星高度不稳定II期结直肠癌患者术后化疗的个体化疗效预测提供有力的决策工具，辅助临床精准分层治疗。

本项目建立了结直肠癌患者数据库，并基于此数据库，成功构建了CT组学模型识别结直肠癌患者的微卫星不稳定性，并确立了三级淋巴结密度为结直肠癌患者的一种简单、可靠且有效的免疫指标，有潜力对结直肠癌患者进行更为精准的风险分层，有助于指导临床决策，为患者提供个性化治疗和预后预测。同时，本项目还构建了基于影像组学和深度学习的非微卫星高度稳定II期结直肠癌术后化疗疗效预测模型。目前建立的模型虽然在训练集和验证集预测准确性均达到90%，但由于样本数据量较少，缺乏外部验证和一定的泛化性。可通过增长随访时间和提高阳性病例比例进一步优化模型并完成外部验证，实现模型预测精度提升，从而为结直肠癌术后化疗的个体化疗效预测提供有力的决策工具，研究成果将有望辅助临床精准分层治疗，具有重要的理论意义。