多视图重建

下载

多视图重建和BigDataViewer的集成可通过斐济更新程序实现。只需更新 Fiji，多视图重建管道即可在 Plugins › Multiview Reconstruction › Multiview Reconstruction Application 下使用。源代码已提供 on GitHub，也请在那里报告功能请求和错误。

要启用 GPU 硬件加速处理，您可能需要下载 本机 CUDA 代码：

• 可分离卷积：用于高斯差分分割，可在GitHub上找到
• 不可分离卷积：用于MultiView Deconvolution，可在GitHub上使用

引文

请注意，斐济提供的 SPIM 注册插件基于出版物。如果您成功地将其用于您的研究，请引用我们的工作：

• S. Preibisch、S. Saalfeld、J. Schindelin 和 P. Tomancak (2010)“用于选择性平面照明显微镜数据的基于珠的配准的软件”，《自然方法》，7(6)：418-419。 Webpage PDF Supplement
• S. Preibisch、F. Amat、E. Stamataki、M. Sarov、R.H. Singer、E. Myers 和 P. Tomancak (2014)“基于贝叶斯的高效多视图解卷积”，自然方法，11(6)：645-648。 Webpage

有关配准方法和 SPIM 成像的技术细节另请参阅SPIM Registration Method。

介绍与概述

多视图重建软件包使用户能够注册、融合、去卷积和查看多视图显微镜图像（第一个框）。该软件专为光片荧光显微镜（LSFM，第二个框）而设计，但适用于任何形式的三维或更高维成像模式，例如共焦时间序列或多色堆栈。

通过与 Tobias Pietzsch 的 BigDataViewer 集成，提供数据的交互式查看和注释。这两个项目共享一个通用的 XML 数据模型来描述多视图数据集。

历史

该软件包是 SPIM Registration 软件包的后续版本。虽然 SPIM 注册暂时将继续在斐济境内进行，但我们主要只为这个新软件包提供支持。它具有 SPIM 注册提供的所有功能，但更加灵活，支持更多类型的注册、融合和数据处理。

示例

“多视图”到底是什么意思？
当我们说到多视图数据集我们通常指的是每个时间点n图像堆栈，可能不同： 渠道 照明方向 旋转角度 请注意，即使n=1，即每个时间点只有一个堆栈，该软件可用于稳定时间序列（漂移校正）。

下面的 YouTube 视频显示了使用该软件包重建的多视图数据集的两个示例。两个数据集均使用基于珠的配准 (Nature Methods, 7(6):418-419) 和多视图解卷积 (Nature Methods, 11(6):645-648) 进行注册，它们是该软件包的一部分。请查看“引文”部分，了解如何引用该软件包的信息。

第一个视频显示了正在发育的 果蝇 胚胎，其所有细胞中都表达 His-YFP。整个胚胎发生过程是使用 Zeiss Demonstrator B 获得的。顶行显示了该七视图数据集的多视图解卷积，下排显示了基于内容的融合。

光片荧光显微镜
光片荧光显微镜于 2004 年进入现代生物学领域选择性平面照明显微镜 (SPIM）已发表。它允许在托托通过从多个角度获取具有高空间和时间分辨率的图像堆栈来对大型样本进行成像。许多令人印象深刻的变体和扩展已经发布，其中一些使用新的变体或全新的命名方案： 开放式SPIM 数字扫描激光光片荧光显微镜 (DSLM) DSLM 和结构照明 (DSLM-SI) SPIM 和 DSLM 的双光子版本 MuVi-SPIM 双光子光片显微镜 (2P-LSM) 贝塞尔光束照明和 DSLM 超显微术 正交平面荧光光学切片显微镜 (OPFOS) 多向选择性平面照明显微镜 (mSPIM) 薄片激光成像显微镜 (TSLIM) ... 在 Emmanuel Reynaud 组织的 2012 年都柏林 Lightsheet 荧光显微镜会议期间，投票总结了大部分这些进展，并命名为光片荧光显微镜 (LSFM).

第二个视频显示了固定的 C。处于 L1 阶段的线虫 幼虫表达 Lamin-GFP 并用 Hoechst 染色。四视图数据集是使用 Zeiss Lightsheet Z.1 显微镜获取的。它说明了通过多视图成像与多视图反卷积相结合可以实现分辨率的提高。顶行显示四个输入视图之一，底行显示多视图反卷积的结果。

详细教程

使用该软件包由几个步骤组成。请注意，该软件更加灵活，并且该顺序只是关于如何在或多或少的标准情况下使用它的建议。

• Dataset Definition
  • The first step in every reconstruction is to define the dataset and thereby create the XML file. This has to be done only once, all consecutive steps are based on this definition, i.e. the XML file.

• Resave the dataset as HDF5/TIFF
  • Once the dataset is defined, you might want to resave all the image data as HDF5 (to be able to view it using the BigDataViewer) or simply as TIFF to enable fast loading of the image data. Also note that those two formats are the only ones that allow to extend the dataset/XML with newly fused data.

• Detect Interest Points
  • Based on an existing dataset definition (XML file), the first step is typically to find interest points in the images that will be used for registration. In this step it is possible to look for multiple types of detections, e.g. fluorescent beads, nuclei or membrane markers. All of them can be stored in parallel.

• Interest Point Registration
  • Once interest points are detected, they can be used to register/align multiple views over time. To find corresponding detection the registration module supports the following modules:
    • Rotation-invariant matching using fluorescent beads (no prior transformation knowledge necessary)
    • Translation-invariant matching using any kind of detections (an approximate knowledge of the rotation is required, e.g. 45 degrees around the x-axis. Check the Tools-Section for how to provide approximate transformations.)
    • Precise matching using the Iterative-Closest Point (ICP) algorithm (the dataset needs to be aligned using for example any of the above methods).
  • The different views (or timepoints) can be aligned using Translation, Rigid or Affine transformation models. We also support regularized transformation models developed by Stephan Saafeld.
  • Alignment over time can be performed in different ways
    • If there is no drift, every timepoint can be simply registered individually. Please note that it is possible to register each timepoint and later on treat it as a rigid unit. In this way it is possible to first register each timpoint using an affine transformation model and consecutively stabilize over time just using a translation model.
    • A reference timepoint that is individually registered first can serve as basis for all other timepoints. This kind of alignment usually only works with external landmarks like fluorescent beads.
    • Alternatively, it is possible to align timepoints using a sliding window of +-n timepoints, in which all views are matched against each other. This will work on any kind of detections.
  • In general, it is possible to stack up as many rounds of transformations as you want. They will be summarized in a list transformations that are concatenated to one single affine transformation before fusion/deconvolution/viewing. A typical list of transformations looks like:
    • Apply calibration (difference in xy and z resolution per pixel)
    • Affine bead-based registration
    • ICP registration based on nuclei

• Fusion/Deconvolution
  • Once the dataset is entirely aligned it can be fused or deconvolved into a single image per timepoint and channel. Deconvolution requires the knowledge of point spread functions (PSF’s), which can be extracted from matched beads directly or can be provided by the user.
  • Alternatively, there is no need to fuse the data and you can interactively look at the data using the BigDataViewer if you resaved it as HDF5. You can still do that now after the registration is complete.

    工具

除了这个潜在的处理概要之外，还有许多工具可用于帮助处理多视图延时数据集。

• Apply Transformations
  • This tool can be used to apply any kind of transformations to individual views, or all views at once. This allows the user to specify a know rotation of different acquisition angles around an axis or to simply re-orient the entire dataset after the registration is complete.

• Duplicate Transformations
  • It can be used to apply transformations that have been computed (or defined) for a certain subset of views/timepoints to other views/timepoints. Typical scenarios where this is required are:
    • Fluorescent beads are only visible in one channel, but the user wants to do apply the same transformation for all other channels
    • The user registered one timepoint and wants to apply the same transformation for all other timepoints

• Display View
  • It allows to simply load and display one of the views as defined in the dataset

• Specify Calibration
  • This enables the user to change the calibration of individual or all views after having defined them. Note that all registrations need to be recomputed if it should reflect the new calibration.

• Visualize Detections
  • Detections as identified by Detect Interest Points can be visualized. It is possible to visualize all detections or only those that are found to be corresponding with other detections and were therefore used for registration. This helps to identify potential misalignments if corresponding detections are not equally distributed around the sample as they should be. One can also load the input view at the same time to overlay the detections with the image data

    视频教程和科学讲座

在EMBO Practical Course on Lightsheet Microscopy期间，我的两次谈话被记录下来：

演讲：多视图光片荧光显微镜 (LSFM) 图像的配准

Stephan Preibisch 的 30 分钟演讲涵盖了多视图光片显微镜数据配准背后的理论，并且还快速解决了多视图融合和反卷积问题。

教程：斐济多视图光片重建软件

Stephan Preibisch 的一小时教程涵盖了斐济多视图重建软件的基本用法。文档、源代码、错误报告和功能请求可以在on SourceForge中找到。

评论、错误和功能请求

GitHub 和SourceForge 上有一个专门针对该项目的网页。它包含本文档、源代码和其他相关内容的链接。所有评论、错误和功能请求均应发布在 GitHub issues 或 SourceForge discussion board 中。

下载示例数据集

有 果蝇 的 7 角度 SPIM 数据集可供下载here。其他数据集可根据要求提供。

系统要求

多视图 SPIM 数据集通常相当大，因此建议在具有大量 RAM 的计算机上使用注册插件。示例数据集的最低要求是 至少 4Gb 内存，但我们建议使用 16Gb+ 系统，理想情况下至少 64Gb 和支持 CUDA 的显卡。您可能需要通过转到Edit › Options › Memory & Threads来增加 Fiji 内存限制。

集群处理

请参阅专门的page，描述用于处理来自 MPI-CBG 集群上的 Lighsheet.Z1 和 OpenSPIM 的 SPIM 数据的自动化工作流程。