基于深度聚类的三维线虫细胞协同分割与识别研究

C.elegans as the gold standard of cell lineage tracing, the measurement of its single-cell resolution transcriptomes under spatiotemporal environments is an important strategy and channel for understanding cell fate and behavior. How to realize the high-throughput, high-precision segmentation and recognition/identification of densely distributed cells in three-dimensional microscopic images is the key enabling technology to support the above mentioned biological researches, and it is also a problem to be studied urgently.. In previous studies, we have demonstrated that the efficient collaboration of segmentation and recognition processes within the clustering framework is feasible and superior to the traditional sequential processing methods. However, the bottleneck that cell appearance features are difficult to be captured and utilized effectively in the co-processing process greatly limits its performance under dense cell distribution scenario. In this proposal, by taking the virtue of high discriminative feature learning ability of deep neural network, we intend to make further expansion and innovation on the basis of our previous studies. Firstly, under the deep neural network architecture, two pixel-clustering based collaborative optimization mechanisms, named joint-optimization and iterative-optimization, will be explored and compared. Secondly, the network architecture that can effectively capture cellular features will be studied with efficient 3D convolution realization as the basis. Thirdly, the multi-task cost function that takes full advantage of appearance, spatial features and corresponding prior knowledge will be designed. Our ultimate goal is to build an efficient, robust and accurate open source 3D C. elegans cell segmentation and recognition system. The research results of this proposal will further expand the theory of multi-task collaboration, and provide solid support for the single-cell resolution complex traits developmental research.

线虫作为细胞谱系追踪金标准，在时空环境下测量其单细胞精度转录组是理解细胞命运和行为的重要策略和渠道。如何在三维显微图像中实现密集分布细胞的高通量、高精度分割和识别是支撑上述生物学研究的关键使能技术，同时也是亟待研究的难题。.申请者前期研究已证实：聚类框架下分割和识别的高效协同是可行的且优于传统序列处理方法。然而，细胞外观特征在协同中难以有效捕获并利用，这一瓶颈极大限制了其在密集细胞分布时的性能。借助深度网络高鉴别力的特征学习能力，本申请拟在前期研究基础上进一步创新，探索深度网络架构下基于像素聚类的联合优化和迭代优化两种分割识别协同处理机制；以高效3D卷积实现为基础，研究有效获取细胞特征的多尺度网络架构；设计能充分利用外观空间特征及先验的多任务代价函数；实现高效、鲁棒、准确的三维线虫细胞协同分割识别系统并开源代码。该研究将进一步扩展多任务协同理论并为单细胞精度的复杂性状发育研究提供有力支撑。

实现高效、鲁棒、准确的三维线虫细胞协同分割识别是在时空环境下进行线虫单细胞精度转录组测量进而探索细胞命运和行为的重要途径。在本项目的支持下，我们实现了从3D线虫数据集构建、线虫细胞图谱制作到线虫全身细胞分割识别协同框架设计的完整过程，同时构造了3D卷积的高效实现结构，提升了3D深度神经网络的运行效率。在此基础上，我们还探索了图像配准与分割识别任务的潜在关联，实现了以高精度配准为手段的典型模式动物脑功能区域分割与识别。项目主要研究成果包括：（1）构建了由580幅3D线虫图像组成的线虫细胞分割识别数据库，制作了线虫细胞统计图谱和参考图谱，该图谱包含了细胞的外观、位置特征等多种先验信息，可为线虫细胞的分割、识别、功能解析等提供有力的数据支撑；（2）构建了面向密集线虫细胞的分割、识别深度网络结构，设计了基于像素聚类和距离向量场的线虫细胞分割模块，提出了基于统计结构匹配的细胞识别策略，实现了高效、鲁棒、准确的三维线虫细胞分割识别系统；（3）设计了一种3D密集分离卷积模块，有效提升了线虫细胞分割网络的训练和推理效率。（4）设计了面向小鼠全脑功能解析的跨模态配准流程，实现了高分辨、大尺度、多模态小鼠全脑三维图像配准系统，为配准、分割、识别协同优化系统研发储备了技术和理论基础。在实现上述具体研究目标的过程中，研究成果主要以论文、专利及开源软件等形式发表、公开在相关学术期刊、学术会议和开源网站上，在具有影响力的国际、国内期刊与会议上发表论文15篇（包括Nature、Nature Methods等国际顶级期刊），授权发明专利2项，项目成果同时为两项安徽省高校协同创新项目、一项科技部2030科技创新重大专项提供了技术支撑。综上，本项目较好地完成了计划研究任务，研究成果实现了预期目标。