The SIGMA rat brain templates and atlases for multimodal MRI data analysis and visualization

ARTICLE The SIGMA rat brain templates and atlases for multimodal MRI data analysis and visualization D.A. Barrière 1,2,3,8, R. Magalhães1,2,4,5,8, A. Novais4,5, P. Marques4,5, E. Selingue1,2, F. Geffroy1,2, F. Marques4,5, J. Cerqueira 4,5, J.C. Sousa4,5, F. Boumezbeur1,2, M. Bottlaender 1,2, T.M. Jay3, A. Cachia 3,6,7, N. Sousa 4,5* & S. Mériaux 1,2* Preclinical imaging studies offer a unique access to the rat brain, allowing investigations that go beyond what is possible in human studies. Unfortunately, these techniques still suffer from a lack of dedicated and standardized neuroimaging tools, namely brain templates and descriptive atlases. Here, we present two rat brain MRI templates and their associated gray matter, white matter and cerebrospinal fluid probability maps, generated from ex vivo T 2-weighted images (90 µm isotropic resolution) and in vivo T 2 -weighted images (150 µm isotropic resolution). In association with these templates, we also provide both anatomical and functional 3D brain atlases, respectively derived from the merging of the Waxholm and Tohoku atlases, and analysis of resting-state functional MRI data. Finally, we propose a complete set of preclinical MRI reference resources, compatible with common neuroimaging software, for the investigation of rat brain structures and functions. https://doi.org/10.1038/s41467-019-13575-7 OPEN 1NeuroSpin, Institut des Sciences du Vivant Frédéric Joliot, Commissariat à l’Énergie Atomique et aux Énergies Alternatives, 91191 Gif-Sur-Yvette, France. 2Université Paris-Saclay, 91191 Gif-Sur-Yvette, France. 3Université de Paris, Institut de Psychiatrie et Neurosciences de Paris, INSERM, 75005 Paris, France. 4Life And Health Sciences Research Institute (ICVS), School of Medicine, University of Minho, 4710-057 Braga, Portugal. 5ICVS/3B’s, PT Government Associate Laboratory, Braga, Guimarães, Portugal. 6Université de Paris, Laboratoire de Psychologie du Développement et de l’Éducation de l’Enfant, CNRS, 75005 Paris, France. 7Institut Universitaire de France, 75005 Paris, France. 8 These authors contributed equally: D.A. Barrière, R. Magalhães. *email: [email protected];[email protected] NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications 1 1234567890():,; Magnetic Resonance Imaging (MRI) has assumed a central role in the investigation of the brain anatomy and function. This role can be attributed to, amongst others, its multimodal capabilities and the ability to study different tissues and their properties by acquiring different contrasts exhibiting high spatial resolution in a single scanning session. MRI studies typically involve the analysis of several subjects divided into experimental groups. In addition, each subject may undergo acquisitions using a variety of modalities which may span several scanning sessions. Because each brain has a unique volume and shape, and is uniquely positioned within the scanner, comprehensive analysis requires standardized anatomical templates and spaces, which enable the spatial normalization of the MRI data and the common mapping of experimental effects, allowing the aggregation of multiple studies. Furthermore, there is a need for standardized atlases that allow the labeling of findings with anatomically and functionally relevant Regions-Of-Interest (ROIs). In humans, the Talairach space and template1and the Montréal Neurological Institute (MNI) template2,3are the most frequently used. That said, many different atlases and brain segmentation schemes have been created for specific purposes, emphasizing sub-cortical4–8and cortical4,5,9,10 parcellations, white matter tracts11–13, as well as maps created using functional information14,15 (for review see refs. 16,17). However, while this plethora of tools, methodologies, templates and atlases is available for human studies, a comparable array is either non-existent or is still incomplete for rat studies. Some neuroimaging resources are available for the rat brain, but they have some technical limitations leading to usability constraints. The Waxholm Space atlas (http://www.nitrc.org/ projects/whs-sd-atlas), built from ultra-high spatial resolution ex vivo acquisitions of a single Sprague-Dawley rat brain provides both T 2-weighted anatomical and Fractional Anisotropy templates, as well as an atlas which delineates 76 anatomical structures. There is a high emphasis on sub-cortical regions, especially the hippocampus18,19. On the other hand, the Tohoku University atlas (http://www.idac.tohoku.ac.jp/bir/en/ db/rb)20, based on lower spatial resolution data acquired in vivo from thirty Wistar rats, provides a T 2 -weighted template, Gray and White Matter (GM and WM) and CerebroSpinal Fluid (CSF) tissue priors, and an atlas with 96 lateralized anatomical areas. There the focus is on cortical areas taken from the Paxinos-Watson rat brain atlas. An older alternative integrated onto the DPABI (http://rfmri.org/dpabi)analysisplatformhas been built from a dataset of ninety-seven Sprague Dawley rats acquired in vivo. It provides an anatomical T 2 -weighted brain template associated with a set of brain tissue priors, and a coregistered version of the Paxinos-Watson rat brain atlas21.Each of these resources is characterized by significant limitations. TheuseoftheWaxholmresourcesmaybehinderedbylargefile sizes (resulting from the very high spatial resolution) when conducting analyses with 4D datasets such as functional MRI, as well as from having been created from a single subject, thus leading to the incorporation of specific-subject/acquisition related anatomical details and artefacts. Both these factors create difficulties when performing normalization procedures with widely used software. Finally, the Waxholm template does not provide its own set of brain tissue priors. The Tohoku and DPABI resources do not provide full brain coverage and the lower spatial resolution may limit their use with high-resolution acquisitions. Additional resources such as rat brain atlases published as PDF images, web-based navigators or image stacks do exist, but are not compatible with the common neuroinformatic formats required by standard brain imaging software22–24 (Table 1). Table 1 Comparison of rat brain templates currently available in the literature (OB Olfactory Bulb, HB HindBrain). Name Modality Magnetic field intensity Anatomical contrast Spatial resolution of acquired images Acquisition type Number of animals Rat strain Template Atlas GM, WM and CSF priors Brain coverage MR image format References Waxholm Anatomy/ Diffusion 7 Tesla T 2 * (GRE) 39 × 39 × 39 µm3Ex vivo 1 SpragueDawley Yes Yes No Full NIFTI Papp et al.18, Kjonigsen et al.19 Tohoku Anatomy 7 Tesla T 2 (RARE) 125 × 125 × 300 µm3In vivo 30 Wistar Yes Yes Yes OB missing NIFTI ValdesHernandez et al.20 Schwarz (DPABI) Anatomy 4.7 Tesla T 2 (RARE) 150 × 150 × 100 µm3In vivo 97 SpragueDawley Yes Yes Yes OB missing NIFTI Schwarz et al.21 Ratat1 Anatomy 14.1 Tesla T 2 (SEMS) 50 × 50 × 200 µm3Ex vivo 9 LongEvans Yes Yes No Full No (JPEG, PDF) Wisner et al.22 Karolinska Anatomy 4.7 Tesla T 2 (RARE) 110 × 110 × 50 µm3In vivo 5 SpragueDawley Yes No No OB and HB missing NIFTI Schweinhardt et al. (2003)48 Figini Anatomy 7 Tesla T 2 (SEMS) 170 × 170 × 580 µm3In vivo 10 SpragueDawley No Yes No OB and HB missing NIFTI Figini et al. (2015)49 Paxinos & Watson Anatomy/ Diffusion 7 Tesla T 2 * (GRE) 25 × 25 × 50 µm3Ex vivo 5 Wistar Yes Yes No Full No (JPEG, TIF, PDF) Johnson et al.23 Development Atlas Anatomy/ Diffusion 7 Tesla T 2 * (GRE) 25 × 25 × 50 µm3Ex vivo 45 Wistar Yes Yes No Full No (PDF) Calabrese et al.24 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 2NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications Herein, we present a new framework for rat brain imaging data analysis and visualization, one which combines a variety of MRI imaging types acquired in vivo and ex vivo, with resolutions suitable for both morphological and functional analyses. Combining the Waxholm and Tohoku atlases, we built a new anatomical atlas with a full brain coverage, which is complemented by a functional atlas derived from resting-state acquisitions data acquired from a large population of subjects. It was our goal to develop a comprehensive set of MRI references and resources, optimized for the rat brain, which would allow investigators to perform unified analyses of both structural and functional data. Use of these tools has been illustrated in the morphometric and connectivity analysis performed in this study. Results SIGMA anatomical template and tissue probability maps. Using six ex vivo MGE acquisitions, we built a high-resolution (90 μm isometric voxel) anatomical template of the rat brain, covering the entirety of the organ, from cerebellum to olfactory bulb (Fig. 1a). When compared to other available resources (DPABI, Tohoku and Waxholm), only the Waxholm template provides similar coverage of the brain, although the contrast-tonoise ratio (CNR) is not optimal between gray (GM) and white (WM) matter structures. To further investigate this parameter for all resources, we computed the CNR between GM and WM for the Waxholm, Tohoku and SIGMA templates as proposed by Tapiovaara et al.25 (Table 2). We were able to confirm that our MGE imaging strategy, especially the parametrization of echoes and acquisition matrix, results in a high CNR between the two main tissue classes (WM and GM) and has a spatial resolution simultaneously capable of capturing anatomical detail while limiting the influence of magnetic susceptibility artifacts present in T 2-weighted images (Supplementary Fig. 1). Tissue probability maps for both the SIGMA and Tohoku templates are shown in Fig. 1b. It is noteworthy that the Waxholm template corresponding to a single anatomical image does not provide any TPM, and the DPABI template defines only two of the three tissue classes (parenchyma and CSF) and as such is unable to distinguish GM and WM. Comparing the SIGMA and Tohoku templates and TPMs, several differences are evident through visual inspection (Fig. 2). They are: (i) more detailed definition of the WM fibbers of the corpus callosum (Fig. 2a); (ii) enhanced definition of the ventral hippocampus with finer delineation of GM and WM (Fig. 2b, c); (iii) more accurate delineation of the central and lateral ventricles in the CSF TPM (Fig. 2d). Template comparison for VBM analysis. When we characterized the performance of the SIGMA and Tohoku templates during the VBM analysis of stress-induced change in the DPABI ab Tohoku Waxholm Brain DPABI CSF DPABI Brain SIGMA SIGMA Brain Tohoku Gray matter Tohoku Gray matter SIGMA White matter SIGMA CSF SIGMA White matter Tohoku CSF Tohoku Brain parenchyma DPABI Fig. 1 Comparison of anatomical templates of rat brain. a Axial, sagittal and coronal views of the DPABI, Tohoku, Waxholm and SIGMA templates. bCoronal views of the tissue probability maps (Gray Matter, White Matter and CSF) associated with the DPABI, Tohoku and SIGMA templates. All the images have been linearly co-registered within the same space for comparison. For the DPABI template, only two tissue classes are available (brain parenchyma and CSF). Tissue probability maps are not provided with the Waxholm template. Table 2 Comparison of the Contrast-to-Noise Ratio (CNR) between the Gray Matter (GM) and the White Matter (WM) for the Waxholm, Tohoku and SIGMA anatomical templates of rat brain computed as proposed by Tapiovaara et al.25: CNRGM=WM ¼SgraySwhite jj ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi σ2 grayþσ2 white p. Name Number of animals Spatial resolution of anatomical template Number of voxels in gray matter Mean signal in gray matter (S gray ) Standard deviation in gray matter (σ gray ) Number of voxels in white matter Mean signal in white matter (S white ) Standard deviation in white matter (σ white ) Contrast-tonoise ratio between gray and white matter (CNR GM/WM ) Waxholm 1 39 × 39 × 39 µm325,902,501 22,141 3265 12,165,815 18,981 3016 0.71 Tohoku 30 125 × 125 × 125 µm3642,257 24,299 3779 291,962 15,146 2605 1.99 SIGMA 6 90 × 90 × 90 µm31,851,454 36,151 1627 979,977 26,019 4127 2.28 For the Tohoku and SIGMA templates, the GM and WM are defined from the corresponding tissue probability maps. For the Waxholm template, the GM and WM are defined from the labeled regions in the corresponding atlas. Source data are provided as a Source Data file NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 ARTICLE NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications 3 hippocampus we found similar results. With both templates stressed animals demonstrated a global decrease of gray matter concentration (GMC) within the hippocampus in accordance with previously reported observations26. The observation was not dependent on the priors set used for the analysis, whether SIGMA (Fig. 3a, t=1.812, df =10,p< 0.05, 50 voxels, two sample Student t-test) or Tohoku (Fig. 3b, t=1.812, df =10,p< 0.05, 50 voxels, two sample Student t-test). We further explored this comparison between templates by building the distribution of tvalues computed within the hippocampus. The histogram of these values reveals very similar distributions (Fig. 3c), with a higher kurtosis when performing the VBM analysis with the SIGMA template (k SIGMA =3.29, k Tohoku =3.05) and an accentuated skewness when using the Tohoku template (sk SIGMA =−0.34, sk Tohoku =−0.52). Furthermore, the VBM analysis with the SIGMA template reveals a larger number of voxels corresponding to p< 0.005 for the two sample Student t-test (nvox SIGMA =878, nvox Tohoku =297), as well as a wider range of t-values (t SIGMA = [−4.67–4.34], t Tohoku =[−4.06–3.08]). Since the exposure to stress is expected to induce reduction in hippocampal volume26,it would appear that the VBM analysis using the Tohoku priors gives a more conservative estimate of the effect, while a similar analysis using the SIGMA priors is more sensitive. SIGMA anatomical atlas of the rat brain. Using the SIGMA anatomical template as a reference, we registered two MRI compatible atlases (Tohoku and Waxholm) of the rat brain into a common space. The process allowed us to overlap and merge the two atlases, creating a more comprehensive coverage of the rat brain (Fig. 4). From the Tohoku atlas, 124 labels (62 per hemisphere, cortical emphasis) were extracted, while from the Waxholm atlas, 122 labels (61 per hemisphere, subcortical emphasis) were used (Supplementary Table 1). The product of this synthesis was a new rat brain atlas comprised of 246 structures (123 per hemisphere). Comparing the volumes of the ROIs before and after registration, we observed that the normalization procedure slightly increased the average volume of the regions of interest (+3.59%, Supplementary Table 2). For some structures, the volume variations could be accentuated due to the Voronoi diagram approach used to dilate competitively each ROI, and therefore fill the space within the brain mask to occupy the brain volume entirely. On the other hand, some white matter structures were found to be shrunk (i.e. fasciculus retroflexus), likely because the mask used to delineate the white matter tracts was calculated from the average of 6 animals, instead of just one as in the Waxholm atlas. We visually inspected and carefully checked the results of the combination, and then reclassified and aggregated the labeled structures according to the systems to which they belonged (auditory, insular, temporal cortex, etc.), as well as with respect to anatomical topography (cortex, basal ganglia, etc.), tissue type (GM, WM and CSF) and hemisphere (left or right). Cortical structures were subdivided into functional (i.e. primary and secondary motor cortices) or structural (agranular, dysgranular, agranular/dysgranular, granular and posterior agranular insular cortices) areas according to the Paxions–Watson atlas (Fig. 5a, b). Sub-cortical structures such as the hippocampus (CA1, CA2, CA3, DG) and white matter tracts (optical tract, anterior commissure, SP5, etc.) were fully segmented, thus completing the labeling (Fig. 5c, d). Processing of fMRI data and functional atlas creation. In order to identify noise and artifacts not accounted for during the standard processing of fMRI data, an Independent Component Analysis (ICA) was run for each scan, allowing the decomposition of the acquired signals into a multitude of temporal patterns. Each component of every scan was visually inspected and those identified as noise or artifact removed through a regression, with the residuals kept as signals of interest. The number of artifact components identified per scan ranged from 2 to 35, with an Tohoku ab cd Tohoku GM Tohoku WM Tohoku CSF SIGMA SIGMA GM SIGMA CSF SIGMA WM Fig. 2 Comparison between Tohoku and SIGMA templates and tissue probability maps. a Coronal views of the Tohoku (left) and SIGMA (right) anatomical template of rat brain at the same coordinates (Bregma: –3.8 mm). Blue arrows indicate differences of corpus callosum shape between both templates. bCoronal views of the Tohoku (left) and SIGMA (right) White Matter (WM) probability maps at the same coordinates. Yellow arrows demonstrate differing classification of WM within the fimbria of the hippocampus. cCoronal views of the Tohoku (left) and SIGMA (right) Gray Matter (GM) probability maps at the same coordinates. Orange arrows show the differences in GM classification within the fimbria of the hippocampus. dCoronal views of the Tohoku (left) and SIGMA (right) Cerebrospinal Fluid (CSF) probability maps at the same coordinates. Green arrows show the differing CSF classification within the lateral ventricle. ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 4NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications average and standard deviation of 15.2 ± 5.7, respectively. A number of sources of noise and artifacts were identified. One common artifact result from head movement, typically seen in voxel correlations all along the edge of the brain (Supplementary Fig. 2a). Another characteristic, which allows artifact components to be identified, is that they appear over a single coronal slice, while physiologically relevant signals would not be expected to be spatially restrained to one slice, acquisition-related events certainly could be. Artifacts of this type were found to exhibit very high z-scores in the equivalent space of one specific brain slice (or the equivalent after normalization and smoothing), resulting either in a partial saturation of the slice that resembles the shape of the brain (ghosting artefacts, Supplementary Fig. 2b, c), or in a complete saturation of the slice (antenna channel crosstalk artefact, Supplementary Fig. 2d). Visual inspection of the 54 different components from the original group ICA resulted in none being identified as artifact. To test the reproducibility of the group ICA maps, we used a total number of permutations Nbetween 5 and 400 in steps of five. From the variance analysis (Supplementary Fig. 3a), we observed stabilization after 280 permutations. When we ranked the components by reproducibility, we found stabilization with increasing N, with acceptable values after 100 permutations (Supplementary Fig. 3b, c). It was a significant finding that the ranking of the least reproducible components was very stable across all values, suggesting the choice of components excluded was stable also. At the conclusion of the analysis, we chose 300 permutations, as a safe stable point. In the reproducibility analysis, we identified a logarithmic bimodal distribution similar to the one described in the original RAICAR methods27, characterized by a local minimum at 0.63 (Supplementary Fig. 3d). All associated components were thresholded by this value and the reproducibility level was determined as the number of components above it. Fifty-three components were found to exhibit reproducibility above the half of maximum value of the reproducibility level (Supplementary Fig. 3e). Among the 53 reproducible components, 6 corresponded to bilateral symmetric ROIs. The result was the identification of 4.74 a b c –4.82 4.74 –4.82 SIGMA template Tohoku template 7000 5000 Cumulative frequency 3000 1000 –4 –3 –2 –1 t-values 0123 –26 –32 –30 –24 –22 –28 –26 –32 –30 –24 –22 –28 4 2 0 –2 –4 4 2 0 –2 –4 Fig. 3 Comparison of VBM analyses performed using the SIGMA or Tohoku resources. a Hippocampus mesh plot representing the surface map of Gray Matter Concentration changes between control and stress animals at the same threshold obtained from VBM analysis using the SIGMA priors and template. bHippocampus mesh plot representing the surface map of Gray Matter Concentration changes between control and stress animals at the same threshold obtained from VBM analysis using the Tohoku priors and template. cComparison of t-values distribution within the hippocampus when performing the VBM analysis with the SIGMA or the Tohoku template. Control and Stress animals (n=6 for each) were compared using the two sample Student t-test implemented in SPM8 (t=1.812, df =10, p< 0.05, 50 voxels). Source data are provided as a Source Data file. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 ARTICLE NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications 5 59 distinct functional ROIs (labeled using the Paxinos–Watson rat brain atlas, Supplementary Table 3) that comprises the SIGMA functional atlas of the rat brain. The regions defined for the most part overlapped with anatomical cortical or subcortical structures, such as the hippocampus and the thalamus (Fig. 6). Comparison of structural and functional atlas networks.We wished to assess the consequences of using our functional, as opposed to anatomical, atlas for the determination of functional connectivity. To do so, we compared the two during the analysis of rs-fMRI data. First, we tested the voxel-wise variance from each voxel matching a ROI, in order to evaluate how much each voxel diverges from the group mean. Comparing the voxel variance maps obtained with the functional and structural atlases, using non-parametric permutation tests, revealed significant effects in both directions for 91% of the voxels (both increased and reduced variance Fig. 7a). Increased variance was generally found using the anatomical atlas (59% of the voxels versus 31% for the functional atlas). The volumes of the ROIs of the reduced anatomical and functional atlases did not significantly differ (t= 0.41, df =139, p=0.68, hedge’sg=0.074, x anatomical =16358 ± 16210 μm3and x functional =17345 ± 9984 μm3, confidence interval =[14455 19088], two sample Student t-test). The graph theoretical properties of both anatomical and functional connectomes were found to display a power-law distribution at the lower density values (between 0.025 and 0.1) and a transition to Gaussian distributions for higher values (Fig. 7b, c). Both networks were also found to have small-world like properties: average σ anatomical =3.57 (c a /c a_rand =3.53, e a /e a_rand =0.98) and average σ functional =4.11 (c f /c f_rand =4.18, e f /e f_rand =1.02) calculated at a scarcity of 0.1. When applying the Newman modularity algorithm to the connectomes, we found modularity scores of 0.86 and 0.84 (anatomical, functional), which resulted in partitioning the connectomes into 5 and 4 sub-modules (color-coded in Fig. 7b, c). For both connectomes, a grouping of neighboring anatomical regions emerged which was composed by a ventral network (colored in yellow), an anterior dorsal cortical network (red), a posterior cortical network in the functional atlas (teal) and a brainstem and posterior sub-cortical network (blue). Discussion The use of preclinical MRI is one target of growing interest in the study of the brain structure and function in both healthy and pathological conditions. The use of advanced MRI techniques, coupled with the development of specific animal models, is a powerful way to push new breakthroughs in the understanding of brain functioning and pathology. Herein, our goal was to develop a comprehensive set of MRI compatible templates and atlases for the rat brain, to allow the unified and standardized analysis of multimodal rat brain MRI data and to pave the way for the development of multicentric preclinical studies (Supplementary Fig. 4). We have described a new set of resources for the analysis of anatomical and functional MRI data acquired from the brain of the Wistar rat. The resources function in a common reference space and are: (i) a high-resolution ex vivo T 2-weighted anatomical template and its associated GM, WM and CSF tissue probability maps (TPMs) for voxel-based morphometry analyses, (ii) an in vivo T 2 -weighted anatomical template and its associated GM, WM and CSF TPMs, (iii) an anatomical whole-brain atlas composed of 246 structures, derived from the merger of the Waxholm and Tohoku atlases and (iv) a functional brain atlas composed of 59 structures. In addition, we describe a brain mesh suitable for data visualization with common neuroimaging software (such as SPM, FSL, BrainVisa/Anatomist). Using an ex vivo imaging approach based upon high spatial resolution MGE acquisitions, we were able to achieve high contrast between the GM, WM, and CSF in the rat brain samples and a high level of anatomical detail. Our protocol avoided the deformation which would have accompanied brain extraction from the skull. The parametrization of echoes ensured a high signal-to-noise ratio, while simultaneously restricting the influence of macroscopic magnetic susceptibility artefacts present in T 2-weighted images that are known to reduce the MRI signal within the entorhinal cortices and the amygdala. To these images, we applied the unified segmentation approach of SPM, which relies on a generative model combining bias correction with deformable tissue priors, and a Gaussian mixture model, to generate an anatomical template of the rat brain and corresponding GM, WM and CSF TPMs. The segmentation process –52–66 –10–24 –80 –38 –52 –66–80 –38 +18+4 SIGMA anatomical template SIGMA anatomical atlas 3D view of the SIGMA anatomical atlas –10–24 +18+4 ab Fig. 4 The SIGMA anatomical template and atlas of rat brain. a Coronal slices of the ex vivo SIGMA anatomical template of the rat brain and the corresponding slices of the SIGMA anatomical atlas and (b) 3D representation of the SIGMA anatomical atlas. ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 6NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications started with the existing Tohoku priors of GM, WM, and CSF, and the output TPMs were then used in an iterative workflow as new input priors. This process allowed significant improvement in the quality of the generated TPMs, as well as guaranteeing a complete coverage of the brain. Compared to other templates available in the literature (DPABI, Waxholm and Tohoku), the SIGMA anatomical template exhibits a superior CNR between the two main tissue classes (GM and WM), accompanied by one of the highest spatial resolutions (90 µm isotropic). Only the Waxholm template has been acquired with a higher spatial resolution (39 µm isotropic). Therefore, we believe our template offers a good compromise between resolution and CNR and further, that our template with associated TPMs will be of great utility for neuroimaging studies requiring both spatial registration and segmentation of MRI anatomical data. The creation of a second anatomical template, in the same space as the ex vivo SIGMA template but built using in vivo T 2 -weighted acquisitions provides another resource for the registration of in vivo anatomical and functional MRI data such as is required with BOLD sensitive EPI acquisitions. We recognize and point out a peculiarity of our TPMs, which is that certain regions of sub-cortical gray matter (thalamus, PAG) are included in the white matter prior. We believe this is due to a high fibber density within these structures which results in intensity range similar to that of white matter. This is a common phenomenon, one also found in other standard resources (Fig. 1). a d Temporal (T) Olfactory (O) Limbic (L) Somatosensory (S) Auditory (A) Insular (In) Au1 dAu2 vAu2 DI AI GI AIP ADI Fr OFr PrL DLO Ect TeA Ent Prh L-Ent-Int L-Ent L-Ent-Ext M-Ent Prh-35 Prh-36 S1J S1BF S1DZ S1FL S2 S1DZ0 S1ULp S1HL S1SH S1Tk Retrosplenial (Rs) Motor (M) Cingulate (Cg) Parietal (P) Visual (V) V1 V1M V1b V2L V2MM V2ML lPA mPA pcP pdP prP RSD RGa RGb M1 M2 Cg2 Cg1 b c S A In O T V P Cg L Cg S P Rs V A T O L M ac opt ml sp5 sp5 ac cc cc cc cc cc cc Th Th Hb cb ic ac pc ic Pons pc Pons ac cb cc ac Th Th ic ic DG DG DG DG CA1 CA1 CA1 CA1 CA2 CA2 CA3 CA3 CA3 CA2 CA1 CA2 CA3 DG Fig. 5 Cortical and sub-cortical details of the SIGMA anatomical atlas. a,bLateral and dorsal views of the cortical areas after normalization of the Tohoku atlas to the SIGMA anatomical template. The cortex has been segmented into cortical areas such as auditory (A), insular (In), temporal (T), limbic (L), olfactory (O), somatosensory (S), motor (M), cingulate (Cg), retrosplenial (Rs), parietal (P) and visual (V). Each area has been subdivided (using the Paxinos-Watson atlas) into functional areas (i.e. primary and secondary motor cortices) or structural areas (i.e. agranular, dysgranular, agranular/ dysgranular, granular and posterior agranular insular cortices). c,dLateral and dorsal views of sub-cortical structures (hippocampus and white matter tracts) after normalization of the Waxholm atlas on the SIGMA anatomical template. Legend of labeled regions:−Auditory: Au1 =primary auditory cortex; dAu2, vAu2 =secondary auditory cortex, dorsal and ventral areas. −Insular: AI, ADI =agranular insular and dysgranular insular cortices; GI, DI =granular and dysgranular insular cortices; AIP =posterior agralunar insular cortex. −Temporal: TeA =temporal association cortex; Ect =ectorhinal cortex. − Limbic: Fr =frontal association cortex; DLO =dorsolateral orbital cortex; OFr =orbitofrontal region; PrL =prelimbic cortex. −Olfactory: Prh =perirhinal cortex; Prh-35, Prh-36 =perirhinal areas 35 and 36; Ent =entorhinal cortex; L-Ent, M-Ent =lateral and medial entorhinal cortices; L-Ent-Int, L-Ent-Ext = lateral entorhinal cortex, internal and external parts. −Somatosensory: S1J, S1FL, S1HL, S1SH, S1ULp, S1Tk, S1BF =primary somatosensory cortex, jaw, forelimb, hindlimb, shoulder, upper lip, trunk and barrel field regions; S1DZ, S1DZ0 =primary somatosensory cortex, dysgranular region and dysgranular zone 0; S2 =secondary somatosensory cortex. −Motor: M1, M2 =primary and secondary motor cortices. −Cingulate: Cg1, Cg2 =primary and secondary cingular cortices. −Retrosplenial: RGa, RGb =retrosplenial granular A and B cortices; RSD =retrosplenial dysgranular cortex. −Parietal: lPA, mPA = lateral and medial parietal associative cortices; pcP, pdP, prP =parietal cortex postero-caudal, dorsal and rostral parts. −Visual: V1 =primary visual cortex; V1b, V1M =primary visual cortex, binocular and monocular areas; V2L =secondary visual cortex, lateral area; V2ML, V2MM =secondary visual cortex, mediolateral and mediomedial areas. −Hippocampus: CA1, CA2, CA3 =cornu ammonis areas; DG =dentate gyrus. −White matter tracts: ac =anterior commissure; cc =corpus callosum; f =fornix; Hb =habenular commissure; ic =internal capsule; ml =medial lemniscus; opt =optic tract; pc =posterior commissure; sp5 =spinal trigeminal 5 tract; Th =thalamus. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 ARTICLE NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications 7 To confirm the potential of the SIGMA anatomical template of the rat brain and its corresponding tissue probability maps, we tested their use as a registration reference by analyzing the MRI data acquired on a cohort of animals previously characterized from a behavioral (elevated plus maze) and neuro-endocrine (corticosterone assay) points-of-view26. When we compared the results obtained with the Tohoku and SIGMA templates using the same anatomical MRI dataset and the same SPM data analysis pipeline, similar findings in the expected direction were obtained. It is of interest that the use of SIGMA resources resulted in more prominent effects. Recent studies suggest that interpolation introduced by nonlinear normalization procedures promotes over-regularization. To address these issues, alternative methodologies such as minimal deformation templates have been employed in both human and primates studies28–30. It would thus appear reasonable to investigate using these alternativeprocessing strategies in rats to improve spatial normalization, particularly in longitudinal studies, by expanding the available resources to include age and strain-specific templates. The creation of a complete MRI compatible atlas is a complex challenge considering the large number of structures and subdivisions identified within the Paxinos–Watson reference atlas. The absence of gyri in the rat brain makes the parcellation of the cortical regions very challenging. Only through accurate registration of a MRI anatomical template with the Paxinos–Watson atlas, is such a detailed delineation feasible18,20,21. Currently there exist two MRI templates with corresponding atlas for the rat brain: the Waxholm atlas18, which contains primarily sub-cortical structures, and the Tohoku atlas20, which emphasizes cortical anatomy. While the two are complementary in their coverage, their differing spatial orientation and resolution make their use in the same study impractical or difficult. To address this issue, we registered these two atlases into the space of our SIGMA anatomical template. By so doing we are able to provide a comprehensive and detailed atlas of the rat brain for MRI morphometry and segmentation. To enhance its accuracy, we complemented this Waxholm–Tohoku merged atlas with an improved more accurate delimitation of the white matter and ventricles. We recognize that the segmentation of some sub-cortical regions (e.g. thalamus, amygdala) will need to be improved to provide a more detailed rat brain atlas in stereotaxic coordinates dedicated to MRI analysis. To complement the anatomical atlas herein described, we developed a functional atlas for the rat brain, using a group ICA analysis validated through a RAICAR approach. From this analysis, we identified 59 bilateral ROIs covering cortical, sub-cortical and brainstem structures that are functionally distinct. Despite having been derived from purely functional data, this atlas broadly, if not precisely, correlates with the general anatomical boundaries and many are associated with specific anatomical structures. A primary motivation for the creation of this atlas is derived from the need to perform brain segmentations which is optimized for functional MRI analysis, since the signal sources do not necessarily match typical anatomical boundaries. A similar requirement has been identified by those performing human studies, resulting in efforts to generate more diverse, multi-modal atlases14,15. A significant result of our study is that by comparing voxel-ROI variance we have shown that the use of a functional atlas reduces variability, thus decreasing the loss of information resulting from the reduction of dimensionality. The present work was further motivated by the recognition of the need to develop resources having resolution parameters and ROI sizes appropriate for rat fMRI studies. Many anatomical atlases, such as the Waxholm or Tohoku atlases, are characterized by a large number of regions defined with a high spatial resolution. However, standard functional MRI acquisitions are unable to reach such high spatial resolutions while preserving an appropriate contrastto-noise ratio or repetition time. As a consequence, either an upsampling of the functional MRI datasets or a down-sampling of the atlas resolution is necessary when processing the data. Both these scenarios present drawbacks, so having appropriately designed and optimized resources can greatly improve and facilitate the analysis of functional MRI data. Using our rs-fMRI dataset, we validated the use of both our atlases for functional connectivity analysis by building whole brain connectomes. We demonstrated not only that the small-world metrics, degree distribution and nodal modularity for both networks fit the expected parameters31, but also are comparable for both approaches. Users –80 –66 –10 –66 –52 –38 –10 +4 +18 –24 –80 –24 –52 –38 +18+4 SIGMA anatomical template SIGMA functional atlas 3D view of the SIGMA functional atlas ab Fig. 6 The SIGMA functional atlas of rat brain. a Coronal slices of the ex vivo SIGMA anatomical template of the rat brain and the corresponding slices of the SIGMA functional atlas and b3D representation of the SIGMA functional atlas. ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 8NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications thus have the flexibility to choose and adapt the unique requirements of their investigations. Moreover, the availability of highly consistent anatomical and functional atlases within the same reference space should allow their concomitant use and encourage testing of more complex hypotheses. As a limitation, the SIGMA functional atlas has been built from resting-state fMRI acquisitions performed on anesthetized animals, using adapted concentrations of isoflurane permitting to keep animals in the same physiological conditions of breathing and temperature. Even if isoflurane is the most popular anesthetic product used for preclinical MRI studies32, it is also known to obscure the naturally occurring functional connectivity by inducing synchronous cortico-striatal fluctuations and silencing the subcortical activity33–36. However, a prolonged fMRI acquisition on awake animals is still technically challenging. Therefore, we chose to rely on the most commonly used protocol for restingstate fMRI acquisitions to emulate a rat brain functional atlas that could be used to study functional connectivity alterations in isoflurane anaesthetized rat. Animal models deliver crucial information for the understanding of brain structure and function both in healthy and pathological conditions. The SIGMA template and rat brain atlases were created to bridge the gap between the basic and clinical neurosciences by providing to the preclinical neuroimaging community specific resources built to be used in conjunction with the neuroinformatic tools and methodologies commonly used in human MRI studies. It is our hope that these resources will help basic neuroscientists to conduct their analyses of anatomical and functional datasets in a more standardized way, with the goal of reaching more reproducible conclusions. Methods Animals. In total 47 male Wistar rats (Janvier, Le Genest-St-Isle, France) 8 weeks of age were housed in pairs and maintained on a 12-h light/dark cycle, at 55% humidity with access to food and water ad libitum. All in vivo experiments were conducted in strict accordance with the recommendations of the European Community (2010/63/EU) and the French legislation (decree no. 2013–118) for use and care of laboratory animals. Experimental protocols (stress protocol and MRI scanning) were approved by the “Comité d’Éthique en Expérimentation Animale du Commissariat à l’Énergie Atomique et aux Énergies Alternatives—Direction des Sciences du Vivant Ile-de-France”(CETEA/CEA/DSV IdF, protocol number ID 13-023). All experiments and resources described in this work were executed and created as a part of the SIGMA project, a project co-financed by the Portuguese FCT and French ANR. The projects main goal was to study the impact of stress upon the rat brain revealing imaging biomarkers of resistance and susceptibility to stress. All the data herein presented were acquired on control animals. In the Results section, we describe two examples of the application of the SIGMA resources on the analysis of data resulting from this project. 1.0 2 mm –2 mm –8 mm 0.5 0.0 1.0 –0.8 5 4 1 1 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 Relative frequency Relative frequency 0.5 0.0 0102030 Nodal degree 40 0.025 0.05 0.075 0.1 0.125 0.15 0.175 0.2 0.025 0.05 0.075 0.1 0.125 0.15 0.175 0.2 0102030 Nodal degree a b c Fig. 7 Functional connectivity analysis performed with anatomical or functional atlas. a Significant differences (p< 0.05, non-parametric permutationbased group comparison with a Student t-test, corrected for multiple comparisons using Family Wise Error) in the voxel-mean ROI signal variability. Red represents reduced variability using the functional atlas while blue represents increased variability. Functional connectomes built with bthe SIGMA anatomical atlas and cthe SIGMA functional atlas. Each module within the network is coded by the node color. Shown here are the average networks visualized using BrainNet Viewer (left), the average connectivity matrices (center) and the plots of the degree distribution for densities between 0.025 and 0.2. (t=0.41, df =78, p=0.68, two sample Student t-test). Source data are provided as a Source Data file. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-13575-7 ARTICLE NATURE COMMUNICATIONS | (2019) 10:5699 | https://doi.org/10.1038/s41467-019-13575-7 | www.nature.com/naturecommunications 9