Poster supplementary animations

Poster title: Leveraging the Adaptive Particle Representation for efficient large-scale neurohistology

Authors: Jules Scholler$^1$, Joel Jonsson$^{2,3,4,5}$, Tomàs Jordà-Siquier$^6$, Jorge Barros$^1$, Laura Batti$^1$, Bevan L. Cheeseman$^{2,3,4,*}$, Stephane Pagès$^1$, Christophe M. Lamy$^6$ and Ivo F. Sbalzarini$^{2,3,4,5}$

$^1$The Wyss Center for Bio and Neuroengineering, Geneva, Switzerland
$^2$Technische Universität Dresden, Faculty of Computer Science, 01069 Dresden, Germany
$^3$Max Planck Institute of Molecular Cell Biology and Genetics, 01307 Dresden, Germany
$^4$Center for Systems Biology Dresden, 01307 Dresden, Germany
$^5$Center for Scalable Data Analytics and Artificial Intelligence ScaDS.AI, Dresden/Leipzig, Germany
$^6$Division of Anatomy, Faculty of Medicine, University of Geneva, Geneva, Switzerland
$^*$Now at: ONI Inc., Linacre House, Banbury Road, Oxford, OX2 8TA, UK

Below are animations to support the poster presented at ELMI 2021, where we present an end-to-end pipeline for analyzing large 3D cleared tissue samples such as whole mouse brains or large human brain sections, achieving 100+ times faster computation. In addition to faster processing, the adaptive particle representation (APR) [1] yields memory and storage compression ratios ranging from dozens to thousands depending on the labeling sparsity.

APR conversion

For APR conversion, parameters were set automatically using the Li thresholding method on the intensity thresholded gradient and local intensity fluctuations. We used a mouse whole brain acquired on a mesoSPIM. Eventhough the labelling is not sparse, the computational ratio (total number of pixels divided by total number of particles) was 17 and the memory compression ratio was 23 (which could be further compressed). At the end of the video we increase the gamma display in order to display larger particles in the background.

Left: original data - Right: APR converted data

APR stitching

After conversion, stitching can be done directly on APR by evaluating the pairwise registration between each neighboring tile and globally optimizing for a reliability measure. Each tile is only read once leading to a very efficient implementation, from 20 to 1000 times faster than TeraStitcher [2].

APR segmentation

We use Napari as a backend to efficiently display APR and created a tool to manually annotate dataset efficiently (labels are stored in a sparse array to keep the memory requirement low) and automatically sampled on the underlying APR data. We compute per-particle features by adaptive convolutions and any particle classification method (e.g. random forests like in iLastik [3]) can then be used to segment the data efficiently.

APR registration to atlas

Merged APR data can be reconstructed efficiently at a lower resolution and registered to the Allen Brain Atlas (25 µm here) for more in-depth analysis. Here we used Brainreg [4] to perform the non-rigid registration.

References

[1] B. L. Cheeseman, U. Günther, K. Gonciarz, M. Susik, and I. F. Sbalzarini, “Adaptive particle representation of fluorescence microscopy images,” Nature Communications, vol. 9, no. 1, Dec. 2018.
[2] A tool for fast automatic 3D-stitching of teravoxel-sized microscopy images (BMC Bioinformatics 2012, 13:316)
[3] S. Berg et al., “ilastik: interactive machine learning for (bio)image analysis,” Nature Methods, vol. 16, no. 12, pp. 1226–1232, Dec. 2019.
[4] Tyson, A.L., Rousseau, C.V., and Margrie, T.W. (2021). brainreg: automated 3D brain registration with support for multiple species and atlases.