Baboons (Papio anubis) are primates phylogenetically close to humans with 96% of their genomes in common (Rogers et al., 2000; Comuzzie et al., 2003; Cox et al., 2013). The baboon and human brains share similarities in structure, function, and responses to neurological diseases, which make the baboon a popular laboratory primate species. Many neurobiology studies using baboons utilize imaging techniques such as PET or MRI for primary outcome measures (Horti et al., 2006, 2014, 2019; Szabó et al., 2007; Salinas et al., 2011; Griffith et al., 2012; Wu et al., 2014; Kumar et al., 2018). MRI can generate high-resolution images for evaluating brain structures (Wu et al., 2014) and provide matrices to identify neurodevelopmental disability (Griffith et al., 2012). PET is also widely used in baboon studies providing functional data for neurobiological processes in vivo (DeLorenzo et al., 2011). PET was used to investigate the role of neurotransmitter receptors using radioligands, such as [¹¹C]JHU75528 for cannabinoid receptor system (Horti et al., 2006) and [¹⁸F]ASEM for α7-nicotinic cholinergic receptors in baboon brains (Horti et al., 2014). PET has been performed to study neuroinflammation (Horti et al., 2019), or to observe changes in cerebral blood flow with photosensitive epileptic baboons (Szabó et al., 2007), as well as to inspect disease models, including cancer (Kumar et al., 2018), atherosclerosis (Du et al., 2020), and Alzheimer’s (Tavitian et al., 1993) disease in baboons. Due to the fundamental limits of using scintillation detectors, PET demonstrates a clear drawback of low-resolution that makes it difficult to identify the borders between different brain areas. A brain atlas is an essential reference particularly important for PET to measure region-specific activity in the brain.

The Davis and Huffman histological atlas was the first atlas for baboon brains with sufficient details in the cerebral cortex areas (Davis et al., 1968). This photomicrographic baboon atlas did not provide a clear region definition and was, therefore, limited in its comprehensive application. The later Riche atlas provided anatomical labels on the histological brain sections in the orbito-meatal plane (Riche et al., 1988). This atlas was the first to be used in the analysis of PET images obtained from a benzodiazepine antagonist, [¹¹C]Ro 15-1788, and a dopamine D2 receptor antagonist, [⁷⁶Br]bromospiperone (Riche et al., 1988). The atlas referred to in Black et al. (2001) was a mapping of the Davis and Huffman atlas onto an MRI image template. The atlas’ regional information was created by a rigid alignment of the T1-weighted structural magnetic resonance template image and a PET blood flow template. These early baboon atlases approximated brain regions with labels in a photographic format, which makes them difficult to be utilized in a digital image analysis pipeline. In 2002, the Greer (Greer et al., 2002) atlas was developed based on baboon MRI images, and incorporated a similar focus as the current study to analyze PET images by regions. Like the Black atlas, the Greer atlas was developed on a 1.5-Tesla MRI image template and included a limited group of animals (n = 4–6) that confined its applicability and expandability to analyze current high-resolution image datasets. Compared to the extensive options available for the macaque brain atlas, existing baboon brain atlases are limited in their flexibility and usefulness in the modern baboon imaging study.

Recently, Love et al. (2016) performed a baboon brain template project to create a thorough and representative MRI image template from 89 individual baboons, including males, females, adolescents, and adults, to represent the brain anatomy of a large baboon population. This MRI template has 1.7 times greater resolution (0.6 mm³ voxel size) than the most recent published baboon brain template (0.49 × 0.49 × 1.5 mm³ voxel size) (Greer et al., 2002), allowing the development of a brain atlas with increased quality and accuracy (Black et al., 2001; Greer et al., 2002). The goal of this study was to create a baboon brain atlas based on the Love template to improve the analysis of PET and MRI data. Critical anatomical regions that are partially available in the previous atlases, including the corpus callosum, hippocampus, and amygdala, were delineated and incorporated into a cohesive single atlas for ease of unitary analysis. To assess the atlas as a viable tool for brain image analysis, atlas-based quantification results were compared with commonly-used PET and MRI measures, including PET time-activity curves (TAC), volumes of distribution (V_T), standardized uptake values (SUV), and regions of interest (ROIs). The comparison measures were generated from manual segmentations developed by neuroimaging experts on both MRI and PET. Two sub-pipelines for PET image analysis were developed to test the atlas, depending on the availability of same-animal structural MRI images. Results suggested that the current baboon atlas generated from the group-averaged brain template facilitates an accurate and high throughput analysis of baboon PET and MRI images.

Since there had been no fully segmented atlas available for the baboon brain, a new brain atlas was created to facilitate high throughput analysis of baboon imaging data. This atlas provides identification of 24 distinct brain regions based on the MRI template in a resolution that is at least 1.7 times higher than the existing atlases. The high-quality MRI template includes a large number of baboons of different ages, gender, and weight, which is more representative of a wider range of baboon brain structures (Love et al., 2016). The new atlas has been statistically evaluated not only with individual T1-weighted MRI but also with brain PET data generated using five known radiotracers. Results suggest that the current atlas is sufficiently well-representative of general baboon brain structure to benefit a high-throughput image analysis for PET and MRI.

An atlas is an important reference and tool for analyzing brain image data. It can facilitate analysis such as volumetric analysis, measuring regional uptake, and tracing dynamic responses with time (Gholipour et al., 2017). For example, in Parkinson’s disease volume changes of substantia nigra regions have been used to indicate disease progression and predict neurological disability (Ziegler et al., 2013). A well-defined brain atlas accompanied with automated registration, segmentation and quantification can replace arduous hand-drawing and expedite imaging analysis (Choi et al., 2016). This is particularly evident when dealing with large datasets that make manual segmentation impractical and human operator bias unavoidable. The atlas is especially helpful for analyzing PET data acquired without accompanying structural images to easily identify patterns of radiotracer distribution in large datasets (Bohm et al., 1986). The segmented standard atlas is also beneficial when analyzing images acquired by modern PET-CT or PET-MRI machines because it can be registered easily to anatomical CT or MRI images and provide detailed regional references (Judenhofer et al., 2007; Hofmann et al., 2008, 2009, 2011; Ruan et al., 2020).

Comparison of the manual and atlas-based image quantification showed consistently high agreements for both PET and MRI data (Table 5). High-resolution MRI was advantageous for performing registration in the atlas-based analysis to return a high DSC level (0.83 ± 0.07). The atlas was also tested on the analysis of PET images acquired with five different radioligands. Depending on the brain pharmacokinetics of the radiotracers, PET images can display variable contrast in different functional regions and usually do not present a clear anatomical structure of the brain. The BA plots showed that the atlas-based and the manual SUV values were comparable with a bias ranging from −5.29 to −8.17% depending on the tracer (Figure 7). In comparison, inter-judge agreements of the manual data appeared slightly tighter, ranging from −2.23 to −3.77% (Table 5). These results are similar to the data evaluated from the established atlases for other primates such as macaque (Gholipour et al., 2017).

Compared to the high DSC value (∼0.8) between the manual and atlas-based analysis in MRI, the DSCs were much lower in the PET images (0.58−0.64) (Table 3). This discrepancy may have been caused by insufficient PET image resolution that blurred the structural borders for a good registration and ROI delineation. Low-resolution PET images can also introduce a partial volume effect that spills image contrast over different regions of the brain (Gonzalez-Escamilla et al., 2021). Figure 6 shows the characteristic of different PET tracers that creates difficulty to identify the anatomy of activity-based geometry in the corresponding PET images. This issue may be resolved by acquiring simultaneous PET-MRI for low-uptake tracers or increasing the quality of registration between the MRI atlas and PET with future software packages.

High uptake tracers like [¹¹C]metergoline delivered larger differences in SUV values, particularly in smaller regions, negatively influencing ICC. In the regions with low radiotracer uptake, the ICC value artificially increased due to the limitations of image contrast from PET of radiotracer activity. When radiotracer uptake decreased, the dynamic range of the SUV data lowered, and the ability to discern small changes decreased due to low contrast. Then there may appear to be higher ICC, while DSC and correlation remain lower than is suggested solely by the ICC data. To fully discern how much tracer uptake is present, both ICC and DSC data must be considered.

Without clear landmarks on PET images, manually segmenting presented unique challenges and was dependent on each expert’s own judgment. Confronting this limitation, the atlas-based analysis may alleviate such rater bias with a standardized process of registration to structural PET or MRI, associating the contrast in multiple brain regions and landmarks throughout the entire brain. The overall quality of atlas-based analysis may be higher than one rater’s manual segmentation since manual ROIs are usually drawn on one slice of the image volume at a time. For example, in [¹¹C]rolipram PET, there was no clear border to identify the anatomical cerebral cortex region. As a result, the cerebral cortex on [¹¹C]rolipram had the lowest DSC rating. In the absence of clear anatomical markers, the atlas may provide a standard reference to compare uptake values from the blurry low contrast PET images, as is done in PMOD (PMOD Full Documentation, 2021). Further confounding the issue, disagreement between judges was not consistent while the atlas-based segmentation and expert drawings consistently disagreed on specific regions with all tracers, demonstrating the consistency of atlas-based analysis. As a result, the average bias between judges was lower (3.07 ± 0.78%) than that for judge-atlas (7.11 ± 1.58%).

Comprehensive labeling of every neuroanatomic structure in the Haiko89 MRI template has yet to be done. The current baboon atlas provides definitions of 24 distinct brain regions sufficient for most PET image analyses. To account for acute deformities in brains, secondary atlas may be required. For example, human brains that have been significantly affected by Alzheimer’s disease will have clear imaging biomarkers such as abnormalities in the hippocampus and amygdala, and ventricular enlargement in a U-shape (Coupé et al., 2019). Since these changes are very stark in comparison to a healthy brain, a secondary atlas could increase the quality of registration and subsequent quality of segmentation. The current atlas can also serve as a foundation for the future development of more detailed atlases. This primary baboon atlas may be transformed onto a particular experimental dataset, and manually modified as necessary to create a secondary atlas to be used in conjunction with the primary. The future secondary atlas would serve to address the limitations of a single atlas when used with particular research demands. Since the current atlas was created based on the Haiko89 template image that incorporated all of the sub-groups together, a secondary atlas is needed to provide substantial benefit for analysis accuracy when used with datasets consisting of only specific subgroups such as elderly baboons, young baboons, or datasets with only one sex.

MRI scans that had profound artifacts were less likely to be registered successfully, adversely impacting the accuracy of segmentation. Similarly, another component of difficulty was registering the PET and MRI images. This was countered in this study by utilizing well-documented automated registration for MRI images and manual landmark registration between PET and MRI images (Neelin et al., 1993; Shan et al., 2011; Keszei et al., 2017). At the time of writing, there is an active research and proprietary development of automated solutions for the issue of independent PET-MRI image registration (Robertson et al., 2016). Future atlas-based analysis may be incorporated with automated registration solutions, allowing rapid analysis of the images acquired between different imaging modalities.

This study describes the creation of a high-quality digital atlas for the baboon brain that will help combine PET functional/neurochemical data with MRI-based anatomy. The new atlas was demonstrated useful to acquire comparable results with the expert manual quantification of T1-weighted MRI and PET images acquired with five different radioligands. Since the baboon and human brain share large similarities in structure, function, and metabolism, an atlas-based approach in the baboon model may provide exceptional data needed to inform human neuroimaging studies. The atlas is made publicly available in electronic format to benefit the baboon research community that utilizes MRI and PET.

The datasets presented in this study can be found in online repositories. The name of the repository and accession number can be found below: Figshare, https://doi.org/10.6084/m9.figshare.16663339.

The animal study was reviewed and approved by Institutional Animal Care and Use Committee (IACUC).

AA, T-WT, PW, YK, and NI conceived of the study and provided project structure, data, and direction. JH and AR provided PET data. RS assisted with atlas creation, background research collection, and carried out validation data creation tasks. RK assisted with MRI validation data. C-HH assisted with atlas modification scripts in MATLAB. SL provided MRI data and project direction feedback. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.