CAREER: Large Scale Learning for Complex Image-Omics Data Analytics

本提案旨在开发计算工具，以分析复杂的病理和放射图像数据以及基因组学数据。最近的技术创新使科学家能够从不同的角度获取复杂的成像和基因组数据。然而，主要的计算挑战是由于异构数据分析的前所未有的规模和复杂性。为了解决挖掘此类综合异构图像和基因组数据的关键和挑战性问题，PI建议开发新的大规模学习工具，并探索如何整合多个数据源的特征以进行临床结果预测。它将极大地支持精准医疗计划，该计划已成为一个国家目标，并由美国政府公布，作为一项旨在使医生能够选择个性化治疗的研究努力。这个项目将促进新的教育工具的发展，以加强一些现有的课程。PI提出了基于以下三个组成部分的综合研究和教育计划：(1)大图像分析和特征提取，提出了新颖的稀疏卷积核、稀疏可变形模型和定量拓扑测量来提取局部和全局特征，以充分表征整个图像；(2)大规模特征学习，提出了面向大规模图像标记发现的领域知识引导稀疏特征学习模型和非凸稀疏特征学习模型；(3)多源图像组学数据集成，其中针对大图像组学数据集成开发了稀疏多视图学习和二部图大规模学习，其中图像组学是指从同一患者采集的图像数据（病理图像或放射图像）和组学数据（基因组学、蛋白质组学或代谢组学）。该项目将推进从千兆像素图像中高效特征学习的研究，以及整合异构图像组学数据以进行结果预测和知识发现。该项目的成功将开创医学图像信息学和大数据的新范式。

This proposal aims to develop computational tools for analyzing complex pathology and radiology image data as well genomics data. Recent technological innovations are enabling scientists to capture complex imaging and genomic data from different views. However, the major computational challenges are due to the unprecedented scale and complexity of heterogeneous data analytics. To solve the key and challenging problems in mining such comprehensive heterogeneous image and genomic data, the PI proposes to develop novel large scale learning tools and explore ways to integrate features from multiple data sources for clinical outcome prediction. It will greatly support the Precision Medicine Initiative, which has become a national goal and was unveiled by the U.S. government as a research effort designed to enable physicians to select individualized treatments. This project will facilitate the development of novel educational tools to enhance several current courses. The PI proposes an integrated research and education plan based on the following three components: (1) big image analytics and feature extraction, in which novel sparse convolution kernels, sparse deformable models and quantitative topology measurements are proposed to extract local and global features to fully characterize whole images; (2) large scale feature learning, in which domain knowledge guided sparse feature learning models and non-convex sparse feature learning models are proposed for large scale image marker discovery; and (3) multi-source image-omics data integration, in which sparse multi-view learning and large scale learning with bipartite graph are developed for big image-omics data integration, where the image-omics refers to both image data (pathology images or radiology images) and omics data (genomics, proteomics or metabolomics) captured from the same patient. This project will advance research in efficient feature learning from giga-pixel images, and in integrating heterogeneous image-omics data for outcome prediction and knowledge discovery. The success of this project will create a new paradigm for medical image informatics and big data.