In the last few decades, conventional histology has benefited from the expansion of light microscopy (Wilt et al., 2009; Ghaznavi et al., 2013; Milligan et al., 2019), in conjunction with the development of a wide range of biological staining techniques (Kuan et al., 2015; Kim et al., 2017; Erö et al., 2018; Tward et al., 2020; Wang et al., 2020). Cutting and acquisition protocols have become more and more sophisticated over time, providing a broad variety of procedures. This made it possible to observe the brain in an unprecedented way (Vandenberghe et al., 2016; Erö et al., 2018; Milligan et al., 2019; Tward et al., 2020). However, the resulting data remain massive and difficult to analyze for most of the labs. This is the case for the mouse brain in preclinical studies (Milligan et al., 2019).

Automated tools for analyzing these tissues, allowing the detection of biological objects and identification of the anatomical regions of interest (ROIs) to which they belong, are essential. Object segmentation has seen a tremendous upturn with the expansion of deep neural networks (Ronneberger et al., 2015; Falk et al., 2019). However, accurately identifying ROIs is still challenging and usually requires a brain atlas or expert knowledge of neuroanatomy.

As a result, many histological protocols are focused on specific anatomical regions, lesion areas, or pathological biomarkers, only on several well-chosen slices of interest within the brain (Lebenberg et al., 2010; Mesejo et al., 2012; Kim et al., 2015, 2017; Niedworok et al., 2016; Pagani et al., 2016; Renier et al., 2016; Ye et al., 2016; Dudeffant et al., 2017; Stolp et al., 2018; Zeng, 2018; Chen et al., 2019; Eastwood et al., 2019; Pallast et al., 2019; Bayraktar et al., 2020; Hérard et al., 2020; Sen et al., 2020; Song et al., 2020; Lam et al., 2022; Yee et al., 2022). It is prone to many drawbacks: this tedious work often yields non-reproducible operator-dependant results, suffers from inter- and intra-individual variability, and requires special attention in the statistical analysis design.

Digital mouse brain atlases aimed both to establish a rigorous, precise, and common reference of delineation for anatomical ROIs and, more importantly, to use them as a segmentation tool (Dauguet et al., 2007; Lein et al., 2007; Lau et al., 2008; Dubois et al., 2010; Johnson et al., 2010; Papp et al., 2014; Kuan et al., 2015; Tward et al., 2020; Wang et al., 2020).

A digital atlas being tree-dimensional (3D), the experimental volume needs to be reconstructed so that their respective dimensionality matches. But it is possible to reconstruct the organ in 3D using registration techniques when all or enough serial slices are cut and digitized (Ourselin et al., 2001; Modat et al., 2014; Agarwal et al., 2016; Niedworok et al., 2016; Fürth et al., 2018; Eastwood et al., 2019). This is the main issue to tackle, which cannot be achieved in most of the studies since protocols are not designed to yield 3D histology. One solution to overcome the lack of histological material is to use blockface photography (Toga et al., 1994) as a whole-brain template to achieve 3D reconstruction of several histological modalities of the same sample (Dauguet et al., 2007; Dubois et al., 2010; Vandenberghe et al., 2016). Indeed, 3D histology protocols are time-consuming, expensive, and neurobiologists often acquire only a limited number of slices. Therefore, the delineation of anatomical regions is mostly performed manually on the experimental data and/or the identification of their corresponding atlas slice is based on prior anatomical knowledge (Lebenberg et al., 2010; Ye et al., 2016; Iglesias et al., 2018; Pichat et al., 2018; Balakrishnan et al., 2019; Chen et al., 2019; Chon et al., 2019; Henderson et al., 2019; Pallast et al., 2019; Wu et al., 2019; Yates et al., 2019; Bayraktar et al., 2020; Hérard et al., 2020; Lam et al., 2022; Rodarie et al., 2022).

Furthermore, with the expansion of artificial intelligence techniques used to automatically segment brain slices, the need for reliable annotated database creation has dramatically increased in the last 5 years (de Vos et al., 2017, 2019; Krebs et al., 2017; Li and Fan, 2017; Rohé et al., 2017; Sokooti et al., 2017; Yang et al., 2017; Balakrishnan et al., 2018, 2019; Krepl et al., 2021; Sadeghi et al., 2022; Carey et al., 2023). Hence, automated, rapid, and adaptable atlas segmentation tools are still lacking but mandatory, for instance, when dealing with the segmentation of so-called single brain slices (devoid of 3D reference) needing to locate the 2D plane of each slice within a 3D atlas template volume. As the mouse brain has an elongated shape, most of the studies observe mouse brains in the coronal incidence (Bohland et al., 2010; Berlanga et al., 2011; Renier et al., 2016; Vandenberghe et al., 2016; Stæger et al., 2020), and we therefore focused on this incidence. Three parameters enable the exact location of a single slice plane within the atlas volume: (1) the z-position of the slice along the rostro-caudal (antero-posterior, AP) axis orthogonal to the coronal plane; (2) the tilting angle φ around the dorso-ventral (infero-superior, IS) axis; and (3) the tilting angle β around the transversal (left-right, LR) axis (Figure 1). Some tools, such as cutting matrices, can be used to obtain a quasi-perfect coronal cutting incidence, i.e., with φ and β tilting angles close to zero and therefore negligible, but usually, φ and β tilting angles lead to discrepancies when comparing “real life” slices and atlas ones.

Some studies focused on identifying possible tilting angles φ and β to refine 2D-plan location within the 3D atlas space (Xiong et al., 2018), while others proposed automated methods or user-friendly softwares to handle 2D slices within a 3D space toward z-position-oriented estimation (Puchades et al., 2019; Tappan et al., 2019). These strategies present a more or less accurate estimation of both tilting angles and are not fully automated since they all include manual processing to estimate the z-position. Basically, manual processing limits the use of such methods on a large scale for the study of mouse cohorts, in particular. More recently, a feature-based method called AMaSiNe was proposed to automatically estimate z-position, φ, and β (Song et al., 2020). Authors evaluate them with precision (<100 μm), and segmentation results have been validated on two specific small regions only (primary visual area and dorsal lateral geniculate complex). However, the method is non-reproducible for the analysis of a single slice. In addition, the method is only robust from a minimum of three slices. Finally, a completely different approach has been proposed, using deep neural networks (Sadeghi et al., 2022; Carey et al., 2023). These methods require a large number of slices to train the network and rely on manual ground truth definition. Such estimates are prone to inter- and intra-individual error; their result is subjective and usually performed only on a relatively small part of the dataset. Moreover, the large variety of histological staining, along with the different imaging modalities, makes it very difficult to build up an exhaustive database to train a fairly generic neural network. Most of the existing methods are either very complex and not user-friendly (codes without interface) to be implemented by neurobiologists or require knowledge in neuroanatomy to be used appropriately, both greatly reducing their scope of application.

The method we propose is intended to be generic enough to be used by anyone and benefits from a user-friendly interface. The fully automatic mode we propose gives reliable results, and the user can still adjust parameters. We focused on the estimation of the z-position of single coronal slices. Our automated method is reproducible and can align and segment any number of single slices within a digital 3D atlas. Moreover, we developed a dedicated multi-slices extension to meet ROI-driven histological protocols, resulting in a set of slices from the same brain. Our method is based on a linear registration algorithm as well as an independent and multimodal similarity criterion. The Block Matching (BM) algorithm (Ourselin et al., 2001) was chosen as a robust strategy to register data from different modalities. This method was later included in the NiftyReg library (Modat et al., 2014) and is still well used in many applications (Niedworok et al., 2016; Iglesias et al., 2018; Balakrishnan et al., 2019; Borovec et al., 2020; Mancini et al., 2020). Normalized Mutual Information (NMI) (Studholme et al., 1998) was chosen as a robust similarity metric adapted to multimodality. This metric has proven its efficiency in many biomedical image processing applications (Jefferis et al., 2007; Geha et al., 2008; Dorocic et al., 2014; Costa et al., 2016). The idea of the method is to explore registrations between the experimental single slice and the ones from the atlas template, with increasing degrees of freedom. The NMI criterion is used to propose a generic evaluation framework of the relative similarity between slices after each step of registration. Basically, the method combines similarity information coming from Rigid and Aff ine registration, which explains the acronym we defined for this method: giRAff. We chose to refer to the Allen mouse Brain Atlas (ABA), a digital atlas widely used in neurobiology (Lein et al., 2007; Lau et al., 2008; Kuan et al., 2015). Also, we focused on histological slices covering the cortex, excluding the main olfactory bulb and the cerebellum. Most of the biological samples come from healthy subjects, but we also present some preliminary results on a pathological subject (Alzheimer's disease mouse model).

In addition, high-performance computing strategies were used to reach our goal of segmenting a large number of histological slices. Indeed, as registrations have a relatively high computational cost, calculations were distributed on hundreds of CPU cores through the dedicated tool SomaWorkflow (Laguitton et al., 2011). Finally, to make the method easy-to-use, it was implemented within the user-friendly open-source software interface BrainVISA (Cointepas et al., 2001; Lebenberg et al., 2010).

The giRAff method has the advantage of being exhaustive in exploring all the possible correspondences after linear registration between a single slice under study and the slices from the average template. This exploration is performed in a minimum of time thanks to a distributed implementation. The choice of the registration algorithm as well as the similarity metric was made to suit multimodal studies, and their independence provides robustness in the identification of the right z-position for a given single slice.

The giRAff_m extension has been specially designed for multi-slices studies, where the RSF is taken into account for an accurate and realistic estimation of the common z-position for a given dataset.

All these developments are gathered in a generic pipeline able to automatically segment any number of slices by atlas. The method presents the advantage of being embedded in an easy-to-use software for simple utilization (see Supplementary material S6).

We used two complementary metrics to evaluate the efficiency of the method in its two different aims: its ability to identify the right z-position of single histological slices, whatever their number, and its ability to present relatively good atlas segmentation scores after registration.

In this study, we proposed a method to automatically segment one or a set of single slices using a 3D digital atlas. The giRAff method, based on linear registration tools and on the NMI as a similarity metric, showed its ability to deal with any number of slices, adapting to very different standard histological protocols (3D fluorescence and 2D brightfield imaging). We demonstrated the robustness and the efficiency of the method by applying it on two different datasets: autofluorescence data, which was not affected by cutting artifacts, and histological slices from routine experimental protocols. It was indeed able to identify, depending on the protocol considered, the z-position of one or more single slice(s) with an accuracy of the order of one slice within the atlas template. This amounted to an identification deviation of less than about 100 μm on average, with dice scores comparable to those obtained by an expert. The method also showed its ability to deal with slices suffering from histological artifacts using the multi-slices approach.

The method was based on a balanced use of the similarity information evaluated after rigid and affine registration in an exploratory approach. In this context, the rigid-affine weighting w was of crucial importance as it allowed to adjust the use of NMI information to take advantage of the benefits from each type of registration. Indeed, in the exploratory approach we proposed, the two types of registration can be complementary. Rigid registration is often rough and avoids the identification of a particular slice that is the closest to the single slice considered, whereas affine registration makes the difference in improving tissue registration thanks to a greater number of degrees of freedom (shearing and scaling). On the contrary, affine registration could make slices correspond to each other with an inappropriate superposition of tissues forced by large deformations, whereas rigid registration does not allow such modifications, limits the deformations, and permits the differentiation of these slices. The use of a weighted proportion of the similarity information created a robust study framework for their comparison in an exploratory context. This represents a useful parameter to tune according to the amplitudes of the deformations considered or according to the biological protocol used. For the two modalities tested in this study, the trend was toward either 0.5 or 1. What we would suggest for users is to consider one or the other of the rigid-affine weighting given in the manuscript by default for their own data according to the imaging modality chosen. In the case of another specific imaging modality or for any doubt on the rigid-affine weighing chosen, the operator could easily test adjusting it from 0.5 to 1 or from 1 to 0.5. If this improves the result in their opinion for their own dataset, they should obviously reuse it by default for the next iterations with other data produced in the same modality. Initialization of the registration by centering the slices on each other was therefore a mandatory step in this gradual pipeline. Even if this centering process was presented as being manually performed in this study, it would be possible to readily add a simple algorithm to perform this task in an automated manner. Maximizing the overlap of the binarized tissue surface could be used, for example, to improve the method in the future.

The giRAff method was inspired by the operating mode an expert uses when manually identifying the position of a single slice: neuroanatomists flip atlas pages and try to match the shapes of certain anatomical regions in an exploratory way, as well as qualitatively estimate the similarity in a visual manner. Our pipeline does the same using linear registration and NMI. Although NMI has shown its robustness in various multimodal brain applications, its efficiency remains discussed within the scientific community (Zheng, 2006; Xiong et al., 2018; Song et al., 2020). This similarity metric is known to have non-significative values in absolute: comparing two objects whose nature does not have anything in common can even result in a significantly high NMI score (Rohlfing, 2011). The use of NMI was solely relative in our pipeline, considering its score on any template slice in comparison to each other. This information was never used in an absolute manner, and the nature of the objects being compared was the same, thus avoiding this limitation. The NMI was not used as a similarity metric to estimate registration but only to objectively evaluate the quality of the slice-to-slice correspondence after registration.

In the dataset we used, we purposely selected uncut 3D coherent histological brain volumes (autofluorescence of a cleared brain acquired with a light-sheet microscope), which was considered as a succession of 2D virtual slices. In such a way, it was possible to test different data processing approaches with artifact-free tissue. This could be one of the reasons why the precision of the z-position detection was better for the autofluorescence (lower Δ_sn) than for the cresyl violet. We first opted for this favorable context to make a proof of concept, taking autofluorescence as a kind of ideal case (Piluso et al., 2021a). Then, we confronted with “real life” histological preclinical routine protocols (digitized Nissl/cresyl violet-stained brain sections), based on our robust and adjustable pipeline.

When using the giRAff method for a given individual slice, its z-position is estimated only once. This estimate may suffer from deviations that could be due to the presence of artifacts in the slices, by poor quality registration, or by a relative similarity value that is not significant enough. In the giRAff_m approach, the joint estimation of the position z_m from a set of slices provided a statistical quantity of estimates sufficient to significantly reduce the deviation Δ_sn and its dispersion in general. This improvement was based on the assumption that a large majority of the slices had little or no artifact, and that the registration and similarity metric were robust enough to accurately estimate the z-position of such slices in the dataset. As a result, for one or several slice(s) suffering from artifacts, representing a minor proportion of a given dataset, giRAff_m provided better z-position identification results than giRAff.

On average, for a multi-slices dataset, the z-position of single slices was detected with a precision of one slice in the atlas (~ 100 μm). This deviation is comparable to the one that experts could make on such a dataset, as long as one single slice considered does not perfectly match one given slice in the template. Indeed, because of its slice thickness and its exact location on the AP axis, as well as possible tilting angles, experts sometimes hesitate between two adjacent slices from the template to identify the right z-position of an experimental single slice. Therefore, they are constrained to make an arbitrary choice, assuming that the position they have identified is only accurate within one slice (100 μm).

Some regions with little pixel support, such as substantia nigra, presented poor dice score results compared to other bigger regions with larger pixel supports. In the linear registration algorithm used, few degrees of liberty were allowed to try to optimize a global transformation at a whole-image scale. This obviously tended to maximize the overlap between regions, including larger pixel support at the expense of other smaller regions including significantly fewer voxels. In such regions, a difference of one single voxel was far more significant than in other regions. Using non-linear registration after estimating the z-position could significantly increase the overlap between such small regions and then significantly increase their dice score.

Concerning the cresyl violet data, the M₇ brain showed higher Δ_sn scores (larger deviation) than the other brains without using the multi-slices extension giRAff_m, confirmed by the presence of artifacts due to the histological and digitization protocols (bubbles, added tissue fragments, and external noise). This is a typical example of artifacts that can occur during a conventional histological protocol. Automatically segmenting histological slices with significant artifacts has always been a challenging task for the scientific community (Agarwal et al., 2016). Most of the time, automatic atlas segmentation of these slices is basically impossible. Our proposed giRAff_m extension has the advantage of optimizing z-position detection on a set E of multiple single slices, and thus could be able to identify and segment such slices including artifacts. Results were pooled to obtain the best z_m position estimated for all the slices. Thus, for a set of slices from the same brain, including slices with important artifacts, it was then possible to decrease their rate of contribution π_s (until 0) in the global estimation of the z_m position, but yet achieve their automatic segmentation reliably. Considering a majority of good quality slices selected from E and a robust regression (significantly high coefficient of determination, typically above 0.97), the giRAff_m extension can propose an automated atlas segmentation corresponding for any other slice suffering from those artifacts from the same brain in a robust way, especially without taking them into account in the global estimation of the position z_m. If the rate of contribution π_s was presented as a subjective parameter to add manually as input information within the multi-slices pipeline, further improvements could lead to the use of image processing algorithms able to automatically detect artifacts within histological slices (Agarwal et al., 2016). This would lead to an automated setup of the rate of contribution π_s for each single slice as a consequence. Moreover, using the multi-slices giRAff_m extension allowed for automated estimation of the RSF γ between the data considered. This reinforced the fact the method we proposed is versatile, robust, and adaptable to many types of protocols or histological brain data.

Considering a multi-slices dataset, we focused on a constant inter-slice distance between single slices under study in this article. But in practice, this distance could be heterogeneous. The principle of the multi-slices extension giRAff_m for the analysis of such slices would be exactly the same; the different inter-slice distances can be given as input information within the giRAff_m pipeline.

In conventional histological protocols, tilting angles may occur when slicing the 3D organ. A non-zero φ angle around the IS axis can generate anatomical differences between the left and right side of the slice, which are easy for an expert to identify due to the brain symmetry with respect to the interhemispheric plane. However, it is more challenging to identify a non-zero β angle around the LR axis that will generate differences between the top and bottom of the slice. This angle is most often observed as non-zero, and neurobiologists then have to deal with neighboring slices to perform the segmentation manually. Thanks to a rigid 3D-3D registration between each considered brain and the template volume, it was possible to estimate these tilting angles around the IS and LR axis, and they are of low amplitude (<5°, see Supplementary material S3), hence our focus on the z-position determination. Considering those realistic tilting angles of low amplitude, the accuracy of the giRAff method nevertheless made it possible to preserve automatic segmentations for which dice scores are still comparable to those of an expert. Indeed, as protocols for acquiring those brains may be representative of standard protocols in conventional brain histology performed in the coronal incidence, we assume that tilting angles rarely exceed an amplitude of 5° with modern equipment and in a similar study framework. If this angulation generates genuine anatomical differences compared to data without angulation, the method we proposed made it possible to compensate for this drawback. Indeed, we chose to process data produced in routine histological protocols in this article, i.e., including real tilting angles caused by the cutting process. Histological data presented in this article included their native tilting angles. As the giRAff method detected the z-position of the single slice with high accuracy, its anatomical environment was well identified (basically in the thickness range of about 200 μm). Following this location, registration ensured the best matching of the tissue between the single slice considered and the template slice identified, as it would have been done in the case of considering the respective adjacent slices of its direct neighborhood. In the coronal incidence and with a slice thickness of about a hundred micrometers, anatomical variations are small from one slice to the next adjacent one. The template data are smooth, and very few discontinuities appear when examining the slices one after the other along the AP axis. More specifically, a tilting angle would generate small anatomical differences between the right and left of the slice for an angle φ around the IS axis and between the top and bottom of the slice for an angle β around the LR axis compared to the template data. In practice, using linear registration would basically correct most of those segmentation errors because the presence and location of anatomical regions are almost the same from one slice n to its n-1 and n+1 (or more) neighbors. Indeed, anatomical differences generated by a tilting angle cause linear deformations along one, two, or both axes (IS and LR in the coronal incidence), which affine registration can compensate with shearing. This was confirmed by dice scores calculated, which were widely comparable to those of an expert in the end. The only necessary condition is that the z-position of the single slice considered must be accurately estimated, typically with a deviation less or equal to one slice in the template, to avoid too large anatomical difference between slices considered. It is just a matter of comparing data which are comparable, i.e., extracted, cut, and digitized within a rigorous, consistent, and realistic study framework. If neurobiologists are asked to cut coronal mouse brain slices using a microtome, it is reasonable to believe that their skills will enable them to obtain tilting angles below 5° as observed in the data presented in this article. Visual quality control and steel matrices could also be used for this purpose.

The method we proposed was based on linear registration in a pipeline with an increasing number of degrees of freedom. The use of non-linear registration could compromise the identification of the correct z-position of a given single slice. Indeed, too many degrees of freedom would excessively distort all template slices to match, in an inappropriate way, the single slice under consideration. It would then be challenging to distinguish which was the most similar. In contrast, the use of non-linear 2D-2D registration between the single slice and the template slice identified at the z-position at the end of the giRAff pipeline would certainly enable the segmentation results to be refined. This could be useful for the analysis of small regions, for example. This constitutes one of the further improvements the method could benefit from. Moreover, the lack of ground truth will make the task even harder.

A benchmark between the different methods of segmenting single slices should be carried out to identify which could give the best results according to the experimental data under study. Such a benchmark should accurately compare all the methods using a dedicated common dataset as well as an appropriate metric to evaluate their respective performance. This comparison is too vast to be presented exhaustively and precisely in this paper and could be the scope of another study. Indeed, each method has its own particular way of working, and its results may be of a different nature, making them difficult to benchmark. Nevertheless, we wanted to briefly test whether our method offered competitive results compared with those provided by the most recent state-of-the-art method. A quick comparison was led on two independent single coronal Nissl-stained slices between the latest fully automated method from the state-of-the-art (DeepSlice from Carey et al., 2023) and our giRAff method. We estimated NMI similarity metric after applying both methods in the same conditions. Those unitary tests showed that similarity between the resulting slices from our method outperformed DeepSlice by about 20%, while requiring a longer processing time (<30 s for DeepSlice and about 1 min for giRAff, estimated per slice). Looking at the anatomy in the identified template slices, the z-position determined for both methods was very close, if not equal. Only some slight registration differences were observed, where the registration algorithm used in giRAff provided the best results according to the NMI criterion. These were very preliminary unitary tests, hence the need for this benchmark to be fully explored in future.

The giRAff method was developed to be fully automatic and embedded in an easy-to-use interface with very few input parameters so that it can be easily used by a non-expert. Optional parameters can be adjusted if the user wants to contribute with their own knowledge, such as the selection of the region(s) of interest studied. This information will reduce the number of adjacent template slices to consider in estimating the z-position of a single slice. Only template slices including this or those anatomical region(s) will be pre-selected, thus decreasing the computation time.

In automatic mode, the method segments a single slice in 1 min on a high-performance computing infrastructure. The result benefits not only from the six regions we focused on but from all the subregions defined in the ABA reference. This is comparable to the time it may take an expert to identify the correct z-position of a single slice within the atlas template. For the same processing time, the giRAff method additionally provided direct atlas segmentation of the single slice. Moreover, no knowledge of brain anatomy or even in coding was required to use the method. Its interface and the few input parameters required by our pipeline make it usable by anyone with full autonomy. Even without supercomputer infrastructure, using about 20 computing cores from a workstation, for example, the method for one single slice worked in a reasonable time of about 15 min.

First, preliminary results as well as complementary studies on a brain suffering from pathological lesions showed encouraging results for the method to be able to handle such data in the context of dedicated protocols. This opens the door for automated segmentation of slices from pathological mouse models, whether neurodegenerative or other diseases, as long as data did not suffer from too large anatomical alterations. Similarly, the use of this pipeline can be extended to other rodents, such as rat for instance, or even in other modalities, such as magnetic resonance imaging. Promising results have been obtained on this modality (Piluso et al., 2021b), and future work aims at validating the use of the method in such cases. In addition, the use of this method will indirectly allow better targeting of conventional histology protocols to reduce the amount of brain data to be used in a study.

The wide variety of existing histological protocols as well as the great numbers of anatomical structures in the mouse brain makes the analysis of histological slices quite tedious and complex. In conventional preclinical histology for the analysis of the mouse brain, it is rare to have enough slices to reconstruct the brain in 3D and, sometimes, working on 3D data is not a prerequisite. It is possible to study only one single slice within the brain, but this is also unusual. In contrast, many protocols are based on a fairly large number of slices to perform quantitative studies on particular anatomical regions or around a specific pathological lesion, for example, still precluding 3D reconstruction. Whatever the case, the generic giRAffMapper pipeline was optimized to accommodate most protocols involving any number of single slices. We showed that our method was able to automatically identify the position of single slices within a mouse brain atlas with less than one slice deviation on average and in 1 min for one slice. Atlas segmentations were comparable to those of an expert. The giRAff method does not need any 3D brain volume reconstruction; it is versatile, generic, user-friendly, and requires few input parameters. In future, we aim to take into account real slice angles and use non-linear registration tools to further refine the segmentation of anatomical regions from increasingly precise atlases. This study paves the way for automated atlas segmentation through a simplified interface of any histological mouse slice, half- or whole-brain slice, for pathological models, for different modalities and possibly for different species. This is done in a fully automated way and does not require any particular knowledge of the study involved, nor in neuroanatomy in general, nor even in coding, to be able to use it. This significantly widens the scope of use of such anatomical detailed atlases within the scientific community for a complex task that usually had to be performed only by experts.

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.

The animal study was approved by dedicated Institutional Animal Care and Ethics Committees, where the experimental procedures involving animal models described in this paper come from already published papers (Renier et al., 2016; Vandenberghe et al., 2016). The study was conducted in accordance with the local legislation and institutional requirements.

All authors listed have made a substantial, direct, and intellectual contribution to the work and approved it for publication.