面向非独立同分布大数据的学习方法及其在医学图像分析中的应用

Currently, most big data learning methods are usually required to satisfy the IID (independent and identically distributed) assumption, however, many real applications commonly follow the non-IID assumption. Although several efforts have been contributed to develop the non-IID methods for small-scale learning problems, these methods cannot be directly borrowed to big data learning. Therefore, we study the theory and method for non-IID big data learning from the perspectives of structure, amount and semantics. Specifically, for structure representation, we will focus on multi-scale sparse graph-based structure learning and parallelized Bayesian structure learning. For data reduction, we will develop the novel theory and method for both non-IID random sampling and non-IID subspace learning. For multi-modal data fusion, we will design new methods for both local metric ensemble and coupled metric analysis. Also, these aforementioned methods will be applied to three typical medical image analysis tasks, including the medical image synthesis and super resolution, histological image diagnosis and identification, as well as multi-modal MR image segmentation. Our research works aim to further enhance both the theory and the technologies for non-IID big data learning.

目前针对大数据进行学习的方法基本都基于独立同分布假设，然而在现实应用中非独立同分布假设往往更加普遍。尽管近年来在小规模数据学习中已经对非独立同分布学习模型进行了探索，但无法直接借用于大数据学习。因此，本项目将从“结构、数量、语义”三个角度研究面向非独立同分布大数据的学习理论与方法。特别地，针对结构表示，研究多尺度稀疏图结构学习与并行贝叶斯结构学习；针对数据约简，研究非独立同分布随机采样与子空间学习；针对多源融合，研究局部度量集成与耦合度量分析。针对上述研究内容，将分别应用于三个具体医学图像分析任务，包括医学图像虚拟合成与超分辨率、病理学图像异常定位与识别、跨模态MR图像分割。通过本项目的研究，将进一步加强非独立同分布大数据学习研究的理论深度和应用广度。

本项目研究非独立同分布数据学习理论与方法，并将所提方法应用于医学图像分析及其他相关机器视觉应用。具体地，针对结构表示，研究结构数据的表示与学习；针对数据约简，研究非独立同分布数据的采样与子空间学习，针对多源融合，研究局部度量集成与耦合度量分析。在迁移学习与域适应学习、多源数据融合、度量学习等方面进行突破。基于所研究成果，我们在若干国际重要期刊和会议上发表论文28篇，其中CCF-A/B类与IEEE 杂志论文21篇，申请专利4项，培养(或者联合培养)博士硕士研究生21名。项目成果完成当时申请指标。