Introduction

In neuroimaging, atlasing refers to registering acquired higher dimensional brain data, e.g., 4D time series or multimodal images, to a 3D reference space called “atlas” that is parcellated and labeled based on known functional and/or anatomical features. Anatomical parcellation of the brain and atlas creation is one of the oldest neuroscience modalities initiated in modern neuroscience with phrenology in the nineteenth century (Fig. 1) [2]. Numerous different atlases are currently available for a number of species, and this battery is constantly growing as applications in experimental neurological research progress to better understand structural and functional relationships of the healthy and diseased brain. During the last decades, with the advancement of non-invasive neuroimaging and computational technology, atlasing is rejuvenated and many ongoing large projects and initiatives such as the Human Connectome Project (https://www.humanconnectome.org), the Allen Brain Initiative (http://www.brain-map.org), and Brain/MINDS (http://brainminds.jp/en/) put concerted efforts in mapping the brains of different species to great detail. To this end, very recently, atlases started to serve more specific, tailored applications. For example, methodology to generate high-resolution mouse vascular atlases that are very important in imaging techniques such as magnetic resonance imaging (MRI), positron emission tomography (PET), X-ray computed tomography (CT), functional ultrasound (fUS), and others have impacted the neuroimaging field. To date, the vast majority of neuroimaging researchers use brain atlases since these provide possibilities for automated, more systematic and novel analyses of imaging data. However, there seems little consensus and guidance on the choice of atlas and/or software to align imaging data with it, especially in basic and preclinical research. In this review, we therefore would like to provide an overview of available atlas resources for both animal species and humans. However, due to the fact that many laboratories use own atlas adaptations, we cannot guarantee completeness. A brief insight into methodology to register imaging data to an atlas and examples of research applications will be given as well as a discussion of remaining methodological and conceptual challenges from the point of view of the neuroimaging community.

Fig. 1.
figure 1

Phrenological chart of the faculties from the late nineteenth century. Adapted from [1].

Full size image

Histological Atlases

Brain atlases are of utmost importance for any neuroimaging approach as soon as a certain structure, a region of interest (ROI), has to be anatomically identified. Historically, anatomical brain atlases have been derived from two-dimensional histological sections, which were labeled by experts, e.g., the very well-known and highly appreciated atlases for mouse and rat by Paxinos et al. [3, 4] or the human Talairach atlas [5]. Other frequently used databases exist for mice and rats [6, 7], dogs [8, 9], cats [10], or monkeys [11, 12]. These histological atlases provide microscopic, i.e., cellular resolution, which was and still is of importance to sufficiently delineate the different substructures in the brain. At this cellular level, histochemistry provided specific markers like Nissl staining or ACh-esterase staining serving as basic information for brain parcellation. This approach goes back to the times of Brodmann’s parcellation of the human brain [13]. Despite advances of modern neuroimaging like non-invasive 3D datasets, 2D histological atlases remain the gold standard for the broad community of neuroscience researchers. This results from the fact that immunohistochemical and histological staining techniques provide molecular insight, e.g., antibody as well as genetically defined (mRNA) markers at the (sub-) cellular level. Therefore, there is a need to register 2D images into a “real world” 3D reference system. However, this remains a challenge (see below). In the neuroscience community and the public, the most perceived initiative to encompass all of this information in a single, publicly available database, is the Allen Brain Institute, which provides 3D cellular, genetical, and connectivity information of the mouse but also human brain (http://atlas.brain-map.org/). Most importantly, compared to atlases based on single or few specimens, the Allen brain atlas (ABA) of the mouse brain used modern nonlinear image registration/warping techniques to build an anatomical atlas from > 1500 specimens resulting in a group template that takes into account the interindividual differences within a species.

Neuroimaging Atlases and the Resolution Gap

Despite technical advances, all histological atlases suffer to a different degree from cutting and preparation/staining artifacts, and most importantly, the full three-dimensional context is lost although several approaches have been worked out to cope with some of these constraints, e.g., z-stacking of 2D slices [14, 15], the block-face method [16,17,18,19], or serial two photon microscopy [20, 21]. Furthermore, the role of fast 3D light sheet microscopic imaging (LSM) of cleared tissue will certainly grow to improve the quality of histological atlas databases [22]. Despite these advances of ex vivo techniques, methodologically inherent in 3D imaging techniques like MRI, SPECT, PET, or CT is the 3D context, the preserved original in vivo geometry, and the high if not even artifact-free data quality. For example, the size and shape of ventricles might substantially change from the in vivo to the ex vivo condition and obviously even more during any further preparation step. Such critical issues must be taken more into account in the future. Moreover, the noninvasive methods allow for repetitive anatomical but also functional studies on the very same specimen. Unfortunately, all these imaging methods only have less spatial resolution than histological techniques. However, technical improvements particular in the MRI field lead to much higher resolutions in the order of tens of microns [23, 24] which coined the term MRI microscopy [25, 26]. Based on these resolutions and improved imaging contrasts, MRI-based full 3D brain atlases became available not only particularly for mice and rats [27,28,29,30,31,32,33,34,35,36] but also for other species like gerbils [37], dogs [38, 39], cats [40], and monkeys [41,42,43,44]. Of note, after 3D imaging, the brain undistorted in the skull, e.g., by MRI microscopy, classical histology, and staining procedures can be applied adding dimensions, i.e., new information. Therefore, multimodal whole brain atlases were generated combining the best from the two worlds like the Waxholm datasets of the rat [33] and the mouse [34] as only two examples and multiple efforts for human brain, reviewed in [45]. Very valuable information resources do exist on the internet providing various collections of brain atlases:

  1. 1.

    http://www.brain-map.org

  2. 2.

    https://scalablebrainatlas.incf.org/

  3. 3.

    http://brainmaps.org/

  4. 4.

    https://www.uantwerpen.be/en/research-groups/bio-imaging-lab/research/mri-atlases/

  5. 5.

    http://www.loni.usc.edu/atlases/

  6. 6.

    http://brainmuseum.org/

  7. 7.

    http://www.thelancet.com/journals/laneur/article/PIIS1474-4422(02)00199-0/fulltext

  8. 8.

    https://sites.google.com/site/chuanglab/atlas

Atlas Registration

At this point, another methodological progress is of great importance: non-affine registration or warping approaches for building probabilistic atlases that take the variance across a group of individual specimens from the same species into account. Probabilistic atlases, due to averaging across many individual subjects (after appropriate registration), provide much higher SNR in the datasets and thus more accurate delineation of structures, and it is possible to constantly add individual data to the cohort underlying the database [45]. This provides a mechanism to regularly improve the representativeness of the atlas for the species of interest [46]. Generating an atlas from a huge number of individuals (> 1500 in case of ABA) provides an advancement on its own render new anatomical details possible as the individual barrels of mouse barrel field cortex which cannot be seen on single subjects (extended Fig. 8 in [6]).

Even more importantly, registration of high quality, in particular the nonlinear transformation techniques, allows for a new branch of anatomical analyses, namely voxel-wise group statistics, which were introduced under the term “voxel-based morphometry” (VBM) [47]. In the meantime, this research area has diversified widely [48, 49]. These analyses led to fundamentally new insights into neurobiological research exemplified here in the context of learning and memory with the hallmark papers of Maguire et al. [50, 51]. The authors showed that London taxi drivers have an above average hippocampal volume for performing their task. Thereafter, VBM was used to assess different pathologies of the human brain like in the context of chronic pain, e.g., migraine or comparing across different chronic diseases [52,53,54,55]. This success story was translated to mice, e.g., paralleling the taxi driver finding in mice which learned different spatial tasks [56]. VBM also serves as an important tool in preclinical research, especially in rodent studies of brain pathologies that come along with morphometric changes, e.g., due to atrophy in Huntington’s disease [57], edema-induced tissue swelling [58], genic modification [59], and neurodegeneration [60, 61].

A typical workflow on how atlas registration is used for neuroimaging research is illustrated in Fig. 2. An accurate image registration technique is the main prerequisite to the validity of such a workflow. Nowadays, the most prevalent approach for medical imaging is the use of intrinsic image registration to an average template based on intensity-based similarity measures such as Mutual Information, followed by affine and an optional elastic transformation (reviewed in [62,63,64]). Fortunately, several image registration tools combining these steps for neuroimaging data are available nowadays such as FSL, SPM, FreeSurfer, ANTs, AFNI, and Vinci (for a comprehensive list and links for download see for example www.nitrc.org). However, most of these tools were developed and optimized for human neuroimaging data and often fail when processing animal data. Several projects have therefore generated dedicated databases and registration or segmentation pipelines, e.g., for the mouse [58, 65, 66] or ovine [67] brain.

Fig. 2.
figure 2

Typical workflow for the use of atlases in neuroimaging studies. Using image registration techniques, individual imaging data are aligned with a group average and an atlas in a standard coordinate space allowing region- and voxel-wise group analyses. The inverse transformation of the image registration allows mapping the results back in standard space on the original data. Example images are taken from MRI acquisitions in mouse models of neurological diseases and from the Allen mouse brain atlas. Adapted from [58].

Full size image

One caveat of atlas registration is that several preprocessing steps (e.g., brain-extraction) and tuning of a large set of registration, regularization, and interpolation parameters is necessary. This leaves many degrees of freedom to individual researchers. Although variability in atlases and protocols is still high between research groups [68], the human MRI field could be a good example on how to tackle this problem. To this end, the Montreal Neurological Institute (MNI) human atlas is generally considered a standard and detailed data processing pipelines for image segmentation, registration, and voxel-based statistics have been published for large cohort studies such as the Human Connectome Project in order to make data processing more coherent [69, 70]. For animal studies, large population imaging studies are in their infancy and common agreements are scarce. Efforts to improve this, e.g., publication of detailed parameter settings when using image registration software, are highly encouraged. A good practice example is given by the elastix wiki parameter database which allows users to upload their settings when using the software along the publication of their results (http://elastix.isi.uu.nl/).

Role of Atlases for Functional Imaging

In addition, and somewhat parallel to structural analyses, there is he usage of brain atlases in the context of functional imaging studies. Here, activation biomarkers assessed by 2-deoxy-d-[14C]glucose autoradiography [71,72,73], 2-deoxy-2-[18F]fluoro-d-glucose PET, SPECT [70], or functional MRI blood-oxygenation-level-dependent (fMRI BOLD) activation are intended to be analyzed in a brain structure-specific manner and compared between groups of subjects. To obtain a “match” between functional data, in general at a (much) lower resolution, and higher resolved brain atlases, different registration approaches are applied (for review see [74]). Here, the obtained quality of such a match between atlas and the single subject data space is of great impact for the final results. Automated functional analysis pipelines, which are used in a “black-box” manner, starting from the raw data as input and provide (significant) statistical differences at the output should be much more scrutinized or better falsified than is currently the case. In addition, the complexity and consequently the constraints of the statistical analysis have to be taken into account avoiding false positive results [75, 76].

Structural-Functional Relationship

The possibility to obtain atlas-based multiregional functional activation parameters, like BOLD, tracer accumulation in Manganese-enhanced MRI, PET, SPECT, etc. as well as anatomical data (connectivity, brain region volume), and correlation/regression with behavioral/functional testing, allows for more modern, higher dimensional data analyses. For example, voxel-wise analysis of lesion location over groups in atlas space allows mapping of regions whose damage best explain a functional deficit after stroke [77]. More and more, the dynamic and complex interaction mechanisms between multiple brain regions are taken into account using a combination of statistics used to analyze functional data and graph theory [78,79,80,81]. Vice versa, a better understanding of the relationship between brain structure and function has recently been used to feed multimodal neuroimaging data [82], or neuroimaging fused with genetic data [83] into personalized models of a simulated brain that can be used to predict the brain’s function in silico. Another new endeavor is to provide multimodal anatomical-functional parcellation schemes, e.g., incorporate functional data like resting state fMRI data to obtain parcellations of brain structures at a much higher level fusing anatomical/histological and functional information [84, 85].

Developing Brain and Exotic Species

Brain development is a highly complex process in terms of 3D relationships during development. In order to visualize these relationships, developmental atlases like emap (https://www.emouseatlas.org/emap/home.html) and ABA mouse and human were introduced. Further mention should be made of the fact that more and more high-resolution atlases are being made available for single brain structures/areas like the cerebral cortex [86] or the rat hippocampus [86]. Of note, brain atlases are becoming available for more and more species such as insects (drosophila, honey bee, desert locust) [87,88,89], birds [90,91,92,93,94], fishes [95,96,97,98,99,100], and opossum rat [101], using state of the art ex vivo and in vivo imaging technologies (see, e.g., https://scalablebrainatlas.incf.org/ for an overview for the mammalian species [102]). It is noteworthy that most of these atlases are freely accessible to the scientific community. Finally and very recently, additional data entities led to an enormous boost in dimensionality of digital brain atlases to name only a few topics: connectivity assessed by MR diffusion tensor imaging [103], connectivity assessed by viral trace injections in the ABA [6], physiological activity in response to visual stimuli in the Allen brain observatory project (http://observatory.brain-map.org/visualcoding/), or OMICS approaches like ViBrism [104].

Vasculature

Next to the general mouse whole brain anatomical atlases, more specific atlases have been developed focusing on the brain’s vasculature describing vascular spatial location, vascular distribution patterns, and regional vascular density [105,106,107]. The cerebral vasculature has a key supporting role by supplying the brain with oxygen and nutrients, and removing brain waste metabolites. By doing so, it is maintaining the homeostasis energy metabolism of the brain. Structural vascular abnormalities can have severe impact on normal brain functioning and have been observed in various neuropathological diseases such as stroke, Alzheimer’s disease, Huntington’s disease, multiple sclerosis, and brain tumors [108]. Hence, in-depth knowledge about the cerebral vascular architecture and its alterations in disease are quintessential. Furthermore, spatial information of cerebral vasculature is very useful for improving processing in techniques depending on blood flow or oxygenation such as (resting state) fMRI, fUS, and intrinsic optical signal imaging.

In recent years, advancements in preclinical imaging techniques increased the possibilities for visualization of the cerebral vasculature at global brain scale thus allowing the development of full brain vascular atlases. Ex vivo histology-based automatic sectioning imaging techniques such as knife-edge scanning microscopy (KESM), micro-optical sectioning tomography (MOST), and serial two photon tomography allow whole brain vascular imaging at submicron resolution [20, 109,110,111]. However, acquisition of these datasets is very time consuming due to the embedding procedures and extensive imaging time. Nonetheless, these techniques allow precise investigation of the whole brain cerebral vascular architecture at the capillary level (< 10 μm). More recently, other ex vivo histological techniques have been developed allowing whole brain vascular imaging based on brain clearing techniques such as Clarity and iDisco in combination with LSM for vascular imaging [112,113,114,115]. These techniques rely upon imaging an almost transparent whole brain combined with vasculature fluorescent staining. Examples are lectin-based staining or vascular casting [113, 116]. In contrast to automatic sectioning techniques, the fast imaging time of LSM allows high throughput vascular imaging up to a resolution of ~ 5 μm. Furthermore, LSM also provides additional auto-fluorescence background images which can later on be used to co-register the whole brain vasculature to anatomical atlases, e.g., the ABA [15, 117].

In contrast to ex vivo imaging techniques, non-invasive in vivo imaging techniques have the clear advantage to assess the development of whole brain macro-vasculature and venous sinuses within the same subject over time. CT and transcranial functional ultrasound can be used to visualize the whole brain macro-vasculature with resolutions down to 19 and 50 μm, respectively [118,119,120]. Although these techniques are able to acquire high-quality whole brain vasculature, they lack anatomical information that can help to spatially localize vessels. Thus, combining these techniques with MRI for soft tissues can provide additional information that can help area identification and alignment to other atlases. Furthermore, MRI can also acquire whole brain vasculature by either intrinsic vascular signal using time of flight imaging sequence (TOF) or contrast agents, allowing combined acquisition of anatomical and vascular information up to the resolution of 20 μm (Fig. 3) [121,122,123].

Fig. 3.
figure 3

Generation of a vascular atlas of the mouse brain. In vivo 3D structural MR image (a) and time of flight angiography (b) of large vessels with high blood flow. Vascular casting followed by light sheet microscopic imaging adds an angiogram of all vessels including small vessels and veins (c). Autofluorescence images acquired in the same orientation and the structural MR image can be registered to the Allen brain atlas (d) in order to build up a probabilistic vascular atlas database in standard space.

Full size image

Conclusion

Given the great success of 3D brain atlases and their analysis framework through the flagship project Allen brain atlas already allows in silico studies on the anatomy and function of the brain. This is very effective and cost-saving, reduces the number of laboratory animals (3R principle), and allows to address completely new research questions. Advances in ex vivo and in vivo imaging techniques will further close the gap between histological and neuroimaging atlases which will add more and more atlas dimensions and will help to better understand the complex interplay between the brain’s electrical activity, cellular/molecular composition, structural connectivity, vascular supply, and gene expression.