Cross-modal coherent registration of whole mouse brains

Abstract

: Recent whole brain mapping projects are collecting large-scale 3D images using powerful and informative modalities, such as STPT, fMOST, VISoR, or MRI. Registration of these multi-dimensional whole-brain images onto a standard atlas is essential for characterizing neuron types and constructing brain wiring diagrams. However, cross-modality image registration is challenging due to intrinsic variations of brain anatomy and artifacts resulted from different sample preparation methods and imaging modalities. We introduced a cross-modality registration method, called mBrainAligner, which uses coherent landmark mapping as well as deep neural networks to align whole mouse brain images to the standard Allen Common Coordinate Framework atlas. We also built a single cell resolution atlas using the fMOST modality, and used our method to generate whole brain map of 3D full single neuron morphology and neuron cell types. INTRODUCTIONRecent advances in high-resolution light microscopy 1-6 , tissue clearing 7-13 , sparse labeling techniques 2,14-17 and full morphology neuron tracing methods [18][19][20][21] now make it feasible to map the mammalian whole-brain at single cell resolution [22][23][24] . Large international efforts such as the BRAIN Initiative Cell Census Network (BICCN) 25 , MouseLight project 19 , Allen Mouse Brain Connectivity Atlas 26 and Mouse Connectome project [27][28][29] , are engaged in cell typing, mapping long-range axonal projection and microcircuit connection analyses. Massive image datasets are acquired through a variety of high-resolution, high-throughput imaging techniques, such as serial two-photon tomography (STPT), fluorescence micro-optical sectioning tomography (fMOST), light-sheet fluorescence microscopy (LSFM), volumetric imaging with synchronous on-the-flyscan and readout (VISoR). A critical technique required to use these data is registration (alignment) of brains, their compartments, and various neuronal structures including somas, dendrites, and axon arbors. Data that are acquired across subjects, modalities, time or even species must be mapped onto a canonical coordinate space. Its importance is seen in multiple ways. First, registration of datasets of different modalities generated in different projects provides a valuable resource for everexpanding studies of neuron wiring and functions. Second, the registration to a common standard 'atlas' enables digital delineation and quantification of brain anatomy and neurons' morphology and other attributes such as transcriptomics. Third, registration allows quantitative comparison of features of neurons across samples under different conditions, thus enabling analysis of distinct aspects of neurobiology in a coordinated manner. Established brain atlases such as the Allen mouse brain atlas (using STPT imaging) 30,31 have helped neurobiologists tremendously to study gene expression patterns, brain anatomy, and neural circuits. With the newly produced whole brain single neuron morphology database (e.g.,

The BRAIN Initiative Cell Census Network Data Ecosystem: A User’s Guide

“…(Wang et al 2020) ( Figure 3C,D ). A recent fMOST atlas was derived from CCFv3 (Qu et al 2022)) extending registration accuracy for this modality ( Figure 3E ). The CCFv3 is being refined and improved through the BICCN and currently provides a definitive mouse brain reference framework.…”

Smart imaging to empower brain-wide neuroscience at single-cell levels

“…As simple as it sounds, however, challenges exist, especially for the landmark detection, considering the variations in brain anatomy and intensity diversity caused by different sample preparation and imaging procedures. For this reason, a coherent landmark mapping (CLM) method was adopted to coherently deform the landmark points in the target image to find their best matches in the reference image [ 120 ]. The robustness of the registration is enhanced taking into consideration the brain regions segmented by a deep neural network.…”

“…Unsupervised networks such as N2V [ 207 ], PN2V [ 208 ], Noise2noise [ 209 , 210 ], Noise2Self [ 211 ] and Cyclegan [ 212 ] have shown to compete the supervised networks in denoising tasks. Promising results of deep learning are demonstrated for registration [ 120 , 213 , 214 ]. 3D segmentation is well achieved via 3D neural networks, such as CDEEP3M [ 215 ] and Cellpose [ 216 ].…”

Fluorescent transgenic mouse models for whole-brain imaging in health and disease

“…Simultaneous with the development of imaging technologies, the need for image processing tool sets for terabyte-sized datasets has emerged. The development of a common atlas of the mouse brain (Common Coordinate Framework v3; Wang et al, 2020), combined with image stitching (Bria and Iannello, 2012;Wang et al, 2020), transformation to multi-resolution image formats (Bria et al, 2016), and spatial registration (Tward et al, 2020;Chandrashekhar et al, 2021;Jin et al, 2022;Qu et al, 2022), allows the quantification of cell densities (Renier et al, 2016), axonal projections (Ye et al, 2016), vasculature (Kirst et al, 2020), and the reconstruction of full single neurons (Winnubst et al, 2019;Peng et al, 2021;Gao et al, 2022).…”