Scalable image processing techniques for quantitative analysis of volumetric biological images from light-sheet microscopy

Abstract

Here we describe an image processing pipeline for quantitative analysis of terabyte-scale volumetric images of SHIELD-processed mouse brains imaged with light-sheet microscopy. The pipeline utilizes open-source packages for destriping, stitching, and atlas alignment that are optimized for parallel processing. The destriping step removes stripe artifacts, corrects uneven illumination, and offers over 100x speed improvements compared to previously reported algorithms. The stitching module builds upon Terastitcher to create a single volumetric image quickly from individual image stacks with parallel processing enabled by default. The atlas alignment module provides an interactive web-based interface that automatically calculates an initial alignment to a reference image which can be manually refined. The atlas alignment module also provides summary statistics of fluorescence for each brain region as well as region segmentations for visualization. The expected runtime of our pipeline on a whole mouse brain hemisphere is 1-2 d depending on the available computational resources and the dataset size.Recently, LSFM images of a whole mouse brain have been used to create a single-cell mouse brain atlas 4 . The pipeline consisted of a heterogeneous mix of MATLAB, Python, and C++ software as well as expensive computer hardware, including a dedicated image processing server equipped with four NVIDIA graphics processing units (GPU). In order to handle individual volumetric images that are larger than the amount of available memory, images were processed slice-by-slice for cell detection and rescaled to a manageable size for atlas alignment. Although computationally impressive, such tools often require a great deal of programming expertise or access to proprietary software. As a result, the current large-scale image processing pipelines may be inaccessible to non-experts, and there is a need for large-scale image processing tools for researchers focused on biological questions rather than computational challenges.The protocols presented here are designed to be easy to setup and applicable to users without much experience in setting up complex development environments. Since some users may only want to use part of our pipeline, the protocols are partitioned into three computational modules: first, our image destriping for removing streaks and performing flat-field correction in raw LSFM images; second, stitching for creating a single 3D image from the individual 2D images; and third,

Cellular Anatomy of the Mouse Primary Motor Cortex

“…Imaging was done using 3.6x objective (custom Lifecanvas design; 0.2NA, 12mm working distance, 1.8um lateral resolution) and using three lasers (488um, 561um, and 642um wavelengths). Acquired data was post-processed and region-segmented as described in Swaney et al 60 Acquired 3D images were aligned to the Allen brain reference atlas (CCF V3) based on tissue autofluorescence from 488nm laser illumination. Delineation of cortical layers in MOp was refined using anti-NeuN and anti-Neurofilament-M signals.…”

Ultrafast immunostaining of organ-scale tissues for scalable proteomic phenotyping

“…To demonstrate the value of holistic labeling with eFLASH, we established an image analysis pipeline for atlas alignment, brain region segmentation, and cell detection for generating a quantitative map of various proteins. Volumetric images were automatically aligned to an atlas 4 by linear and non-linear transformations based on Elastix 17 then manually refined 18 . Each aligned 3D image volume was indexed to approximately 580 brain regions with 7 hierarchies.…”

“…All light-sheet imaging was done with either of the following objective lenses: 3.6x objective (custom Lifecanvas design, 0.2NA 12mm WD lateral resolution 1.8um in XY), 10x objective (Olympus XLPLN10XSVMP, 0.6NA, 8mm WD, lateral resolution 0.66um in XY). Acquired data was post-processed with algorithms described in Swaney et al 18 . A complete table of imaging modalities and conditions for every data included in this paper can be found in Supplementary table 2.…”

“…Atlas Alignment. Atlas alignments of mouse brain hemispheres labeled with eFLASH to the Allen brain reference atlas, CCF V3 63 , were carried out using the hybrid automated atlas alignment method described in Swaney et al 18 , which combines Elastix 17 and manual refinement tools to improve alignment accuracy.…”

“…These counts were then used to color the regions in the Allen Brain Mouse Atlas coronal SVG image files. Calculations and visualizations were done using the Nuggt python package 18 .…”