Automatic spatiotemporal alignment of large-scale 3D+t point clouds

多维荧光显微镜已成为世界各地生物实验室的常用技术，因为它代表了在细胞分辨率上研究空间和时间（3D+t）早期胚胎发育的宝贵工具。自动分割和跟踪算法用于从潜在的tb级3D+t图像数据集中提取数千个细胞运动轨迹，这些数据集为详细分析个体间差异提供了可能。然而，在获得这种跟踪点云之后，仍然存在一个基本问题，那就是在多次重复中对单个实验进行比较，以证实生物学假设。目前，这种3D+t比对问题缺乏全自动解决方案，全胚胎分析仅限于简单标本、早期时间点或人工分析。该项目的目标是开发大型3D+t点云的自动时空对齐新方法。由于复杂生物体通常缺乏可用于注册的一对一细胞对应关系，因此该项目的一个基本部分将是开发通用描述符，以使用经典和基于机器学习的方法识别不同发育阶段的各种解剖区域。我们计划使用一个全面的模拟平台来生成机器学习方法的综合训练数据，该平台可以在多个细节水平上模拟不同标本的胚胎发育。我们明确地希望使用内存高效的点云表示，以便能够使用像图像这样的密集数据结构来对齐甚至覆盖多个tb的数据集。确定的解剖标志将使3D+t数据集的精确对齐成为可能，我们将开发一种模块化的相似性度量，可以将这些时空描述符的异构集的信息结合起来。我们设想了一种从粗到精的策略，首先使用动态时间扭曲等方法对数据集进行时间对齐，然后继续对单个帧进行刚性空间配准，最后使用适应性弹性配准技术将点云表面相互扭曲，形成胚胎发育的3D+t图谱。研究结果将成为3D+t点云数据集时空对齐的通用方法、开源实现以及在斑马鱼和果蝇胚胎的大规模光片显微镜实验中的应用。结合现有的自动分割和跟踪方法，期望的对准框架将代表朝着以全自动方式分析大规模成像实验的最终目标又迈出了一大步，这对于跟上不断发展的3D+t荧光显微镜技术至关重要。

Multidimensional fluorescence microscopy has become a common technique in biology labs all over the world as it represents an invaluable tool to study early embryonic development in space and time (3D+t) at cellular resolution. Automatic segmentation and tracking algorithms are used to extract thousands of cell movement trajectories from potentially terabyte-scale 3D+t image data sets that offer the possibility for a detailed analysis of inter-individual differences. A fundamental problem that remains after having obtained such tracked point clouds, however, is the comparison of individual experiments to confirm biological hypotheses in multiple repeats. The lack of fully automated solutions to this 3D+t alignment problem currently limits whole-embryo analyses to simple specimens, early time points or manual analyses.The aim of the proposed project is the development of new methods for automated spatiotemporal alignment of large 3D+t point clouds. As complex organisms usually lack one-to-one cell correspondences that could be used for registration, a fundamental part of the project will be the development of generic descriptors to identify various anatomical regions at different developmental stages using both classical and machine learning-based approaches. We plan to generate synthetic training data for the machine learning approaches using a comprehensive simulation platform that allows mimicking embryonic development of different specimens at multiple levels of detail. We explicitly want to use memory-efficient point cloud representations to be able to align even data sets that would cover multiple terabytes using a dense data structure like images. The identified anatomical landmarks will enable precise alignment of 3D+t data sets and we will develop a modular similarity metric that can combine information from a heterogeneous set of such spatiotemporal descriptors. We envision a coarse-to-fine strategy, which first temporally aligns the data sets using methods like dynamic time warping, then continues with a rigid spatial registration of the individual frames and finally uses adapted elastic registration techniques to warp the point cloud surfaces onto each other to form a 3D+t atlas of embryonic development. The results will be general methods for spatiotemporal alignment of 3D+t point cloud data sets, open-source implementations and the application to large-scale light-sheet microscopy experiments of zebrafish and fruit fly embryos. Combined with existing automated segmentation and tracking methods, the aspired alignment framework will represent a large step further towards the ultimate goal of analyzing large-scale imaging experiments in a fully automated fashion, which is essential to keep pace with the continuously advancing 3D+t fluorescence microscopy techniques.