基于深度随机场的跨域细胞行为检测方法研究

项目名称： 基于深度随机场的跨域细胞行为检测方法研究

项目编号： No.61472275

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 刘安安

作者单位： 天津大学

项目金额： 83万元

中文摘要： 细胞行为自动检测是当前高通量细胞生理周期定量分析亟待解决的问题。针对该问题，拟提出基于深度随机场的跨域细胞行为检测方法。该方法包括细胞行为候选序列提取、低层特征表示和细胞行为建模三个部分，本课题将重点研究基于深度随机场的细胞行为建模问题，具体包括：1) 模型构建：拟利用细胞生物学知识引导联合序列分类和标注的图模型设计，同时继承深度神经网络的非线性特征学习能力和随机场的序列结构学习能力，通过序列特征和序列结构的联合学习，实现细胞行为的联合识别和定位；2) 模型学习：拟构建深度随机场跨域学习的目标函数并设计相应参数优化算法，从而联合利用跨域数据增加模型的泛化能力；3) 模型推断：拟设计融合序列类别标记和段级结构标记的动态规划算法，从而实现二者的联合最优推断。本课题研究有利于跨域细胞行为自动检测系统构建，相关研究成果将有利于信息科学自身发展，并对生命科学新规律的探索具有重要意义。

中文关键词： 深度随机场；跨域学习；动态规划；细胞行为；细胞生物学

英文摘要： Automatic cell behavior detection is a challenging problem for quantitative analysis of high-throughput cell cycle. To handle this problem, we propose a cross-domain cell behavior detection method based on deep random fields. This method consists of three key steps, candidate sequence extraction which containing the potential cell behavior, low-level feature representation, and cell behavior modeling. We will mainly focus on the cell behavior modeling via deep random fields. This proposal contains three parts: 1) model formulation: leveraging the domian knowledge of cell biology to guide the formulation of graphical model, which inherits the ability of nonlinear feature learning by deep neural network and the ability of sequence structure learning by random fields,for joint learning of sequential features and structures and consequently realizing joint identification and localization of cell behaviors; 2) model learning: designing the objective function and the optimization method for deep random fields learning with cross-domain data to enhance the generalization ability of models; 3) model inference: developing the dynamic programming method by fusing both sequence category and structure to achieve jointly optimal inference. With the proposed methods, we will develop the automated cross-domain cell behavior detection system. Moreover, this research will not only facilitate the development of information science but also benefit the exploration of innovative theories of life science.

英文关键词： Deep Random Fields;Cross-domain Learning;Dynamic Programming;Cell Behavior;Cell Biology