联合病理图像和转录组数据预测结肠癌预后新方法研究

Pathological examinations are the gold standard tests for tumor diagnosis. However, the rich information in pathological images, related to cancer prognosis, has not yet been fully mined. Furthermore, with the rapid development of sequencing technologies, molecular data is increasingly used to predict cancer prognosis. Pathological images and molecular data convey tumor structural and functional information, respectively. Therefore, an important scientific question is how to integrate these two complementary types of data to provide improved predictions of cancer prognosis. However, there is a challenge of how to extract and select prognostic features from high-dimensional and multi-modal data. Increased incidence in colon cancer in recent years motivate us to develop a novel method to predict colon cancer prognosis by integrating digital pathological images and transcriptomic data. First, we use a deep convolutional neural network to accurately detect and classify cells in pathological images, then extract cellular topological features that are related to tumor microenvironment. Second, in order to select important features from multi-modal data (cellular topological features and transcriptomic data) to build a prognostic model, we develop a multi-modal feature selection algorithm by adding a priori survival information into the sparse canonical correlation analysis model. This study will support clinical decision making by improved stratification of patients and informed treatment decisions and provide a research tool for digital pathological image analysis of tumor microenvironment, which is of great scientific and clinical significance.

病理检查是肿瘤诊断金标准，但其蕴含的与癌症预后有关的丰富信息尚未被充分挖掘；另一方面，测序技术的快速发展使得分子数据也逐渐用于癌症预后判断。病理图像和分子数据分别反映肿瘤的表层结构和内在功能信息。因此，如何有效结合这两种模态的互补信息以提高预后预测精度是亟待研究的重要科学问题，但面临如何从高维多模信息提取、选择有效预后预测特征等挑战。近年来结肠癌发病率在我国快速上升，本项目提出一种联合病理图像和转录组数据预测结肠癌预后的新方法：利用深度卷积神经网络对病理图像中的细胞进行精确检测和分类，进而提取可描述肿瘤微环境的细胞拓扑分布特征；提出在稀疏典型性相关分析模型中引入生存先验的特征选择新方法，从高维多模数据（细胞拓扑特征和转录组数据）中选择可有效预测预后的重要特征并构建预后模型。本项目将为结肠癌患者的风险分层和治疗方案选择提供决策支持，也可为肿瘤微环境分析提供研究工具，具有重要的科学与临床意义。

病理检查是进行肿瘤良恶性诊断、病理分型和预后的一种最常用的手段。对病理图像进行计算分析能够从中挖掘出人眼难以察觉的、客观的、定量的图像特征，从而能更好地对疾病表型进行建模。本项目旨在提出一种对HE染色病理图像进行定量分析的特征提取方法，用于癌症的计算机辅助诊断和预后。主要包括以下研究内容：1）病理图像特征提取：细胞核分割、细胞核级的特征提取、细胞核级特征聚合成患者级特征；2）研究基于HE染色病理图像的TFE3易位型肾细胞癌鉴别诊断；3）研究病理图像特征与基因突变和预后的关联。本项目取得了优异的研究成果，主要包括：以第一或通讯作者发表学术论文10篇，其中SCI收录8篇（包括Nature Communications和Medical Image Analysis等国际顶刊），MICCAI会议论文1篇；申请专利1项。项目组取得的重要学术贡献如下：1）开发了病理组学特征提取方法，并进行了开源；2）在国际上首次提出基于HE染色病理图像的TFE3易位型肾细胞癌鉴别诊断方法，发现TFE3易位型肾细胞癌与透明细胞肾细胞癌在52个图像特征中存在显著差异，分类模型自动鉴别诊断的性能为AUC=0.894；3）系统分析了胃腺癌组织形态学特征与高频突变基因以及患者总生存之间的关联，并联合转录组数据挖掘预后模型中重要特征涉及的生物过程，帮助理解肿瘤组织形态的底层分子基础。