Morphological and volumetric brain changes that occur with aging are well-documented (see Galluzzi et al., 2008 for a recent review). Despite these changes, many studies aimed at identifying volumetric changes use reference data for tissue segmentation solely derived from young adults. The issues associated with using young adult reference data have been documented and studies have shown that study-specific reference data can improve segmentation results (Ashburner and Friston, 2000; Good et al., 2001; Thompson et al., 2001; Huang et al., 2010). In the current study, we suggest the use of age-specific reference data as an alternative to young adult or study-specific templates. We created and tested a database of age-specific reference templates. Findings suggest that age-specific reference templates can be used to increase segmentation accuracy in developmental neuroscience, neuropsychology and neurology research and clinical practice.

Volumetric brain changes occur with change in age during adulthood. Fotenos et al. (2005) measured whole brain volume differences from age 30 and reported a mean decline in total volume of −0.45% per year after age 65. A common strategy is to distinguish the changes in partial volume estimates (PVEs) of gray matter (GM), white matter (WM), and less often, cerebrospinal fluid (CSF). The most consistent partial volume change is a reduction in GM volume. Ge et al. (2002) reported % GM decline in subjects beginning at age 20 with a constant linear reduction across the span of early to late adulthood. Good et al. (2001), Sato et al. (2003), Sullivan et al. (2004), Taki et al. (2004), Smith et al. (2007) and Lemaître et al. (2005) reported similar negative linear relationships between cortical GM volume and age. Most of these studies show decline in GM beginning about age 20. However, recently Richards and Xie (2015) found that the decline in GM actually begins in the early adolescent years. Changes in WM volume have been found but are not as consistent in the literature. Ge et al. reported % WM changes in a quadratic pattern, with slight increases until age 40, and decreases thereafter. Salat et al. (2009) also reported a quadratic relationship between WM volume and age, with relative preservation or rise in volume until the late 50 s, followed by a steep decline. Lemaitre et al. reported significant WM decline (1.7 cm³/year) in their subjects aged 63–75 years. In contrast, Good et al., Sato et al., Sullivan et al., and Taki et al. reported no global WM loss in their subjects. An associated linear increase in CSF volume has also been reported. This finding is consistent across volumetric studies that report CSF results (Good et al., 2001; Lemaître et al., 2005; Smith et al., 2007).

Tissue segmentation is a crucial step for these volumetric studies. These studies classify brain tissue into three primary types: GM, WM, and CSF (or “other matter”, OM). Typically these studies used automated or semi-automated procedures for guiding the segmentation (e.g., FSL’s FAST, Zhang et al., 2001; SPM’s USN, Ashburner and Friston, 2005; Freesurfer, Dale et al., 1999; Atropos, Avants et al., 2011b). These automated procedures usually require a priori reference data (Zijdenbos et al., 2002) to guide tissue classification. These spatial priors generally come from a common anatomical reference template. This requires that individual structural brain images be first registered to the common anatomical space, a process known as spatial transformation or normalization. Accurate spatial normalization to a common template confers many advantages, in that it facilitates inter-subject, inter-sample and inter-population comparisons (Evans, 2005). Thus, an important piece in the accuracy of this process is the choice of both templates and normalization methods which can provide a good match between the population of interest and the participants under study. The most often-used MRI template was created by the Montreal Neurological Institute (MNI; Evans and Collins, 1993; Mazziotta et al., 2001a,b). The original MNI template was developed from 305 MRIs, with another one developed later from 152 MRIs. Each used MR images of participants with an average age of about 23 years. Some of the volumetric studies mentioned (Sato et al., 2003; Taki et al., 2004; Mandal et al., 2012) relied on this reference data for their analyses.

Using a template derived from the brains of young adults for analyses with the brains of older adults can be problematic. As Huang et al. (2010) point out, despite the widespread use of the MNI and other such normalizing templates, inter-subject variation in brain structures for special populations such as older adults can result in less than ideal compensation for brain region discrepancies after spatial normalization with these templates. An alternative to using the MNI template is to create a study-specific template, derived from structural images of the participants in the study. To demonstrate this, Huang et al. used a study-specific template rather than one developed from young adult images for spatial normalization within a functional MRI (fMRI) data analysis. Huang et al. found that more voxels were identified as significant in older adults when a study-specific template was used. The main advantage conferred by the construction of study-specific templates in this case is a very precise mapping to the common space, since it is derived from the actual participants of the study.

Study-specific templates have also been implemented within voxel-based morphometry (VBM; e.g., Peelle et al., 2012). VBM is an automated method used for voxel-wise statistical investigation of focal differences in brain anatomy as measured by MRI (Ashburner and Friston, 2000). In this context, creating a study-specific template may involve segmenting a group of images from the sample with MNI tissue priors, averaging those partial volumes to come up with study-specific tissue priors, and then using those priors to re-segment all of the images in the sample (see Lemaître et al., 2005 for an example). Studies comparing a study-specific template to the MNI template for VBM demonstrated that using a study-specific template reduced anatomical biases in the analysis (Ashburner and Friston, 2000; Good et al., 2001; Thompson et al., 2001). Again, it is likely that this is due to the precision in mapping afforded by a study-specific template. Recent work has shown that the use of high-resolution nonlinear spatial normalization tools further reduces these biases (Callaert et al., 2014).

A study-specific template is an improvement over the MNI template, but a second, possibly better, solution is the development of age-appropriate a priori reference data. This solution confers several advantages over study-specific templates. First, age-appropriate templates can be used over multiple studies allowing for easier study-to-study comparisons. Results using study-specific templates cannot easily be compared to the results of another study using a different study-specific template. Age-appropriate templates alleviate this issue. Second, a study-specific template may be created by averaging images from a broad age range, depending on the study goal. Age-appropriate templates can be created in smaller increments and may be more similar to any one individual image than a broad study-specific template (c.f. Rohlfing et al., 2009). Several age-appropriate templates have been done for pediatric ages (Wilke et al., 2003; Altaye et al., 2008; Yoon et al., 2009; Fonov et al., 2011; Sanchez et al., 2012a). Age-appropriate templates have been used with pediatric neuroimaging data and improve classification of tissue types in pediatric samples (Wilke et al., 2003; Sanchez et al., 2012a,b).

There are a few reported age-specific templates for adults. Lemaître et al. (2005), created a T1-weighted probabilistic brain atlas representing an average of 662 subjects aged 63–75 years. These subjects were a sub-sample of the Epidemiology of Vascular Aging cohort, a longitudinal study on vascular aging and cognitive function in healthy older adults (see Dufouil et al., 2001 for details of study). MR images for these subjects were acquired between 1995 and 1997 on a 1.0T MR scanner. Farrell et al. (2009) created four age-specific templates, T1- and T2-weighted templates representing 54 subjects aged 65–70 years and 25 subjects aged 75–80 years. These subjects were sub-samples of the “National Aging Brain” Study cohort (MacLullich et al., 2002) and the “Simpson’s Study” cohort (Shenkin et al., 2003), both studies of normal aging. The “National Aging Brain” subsample MRIs were collected on a 1.9T scanner. Sato et al. (2003) created reference templates for a very wide age range that included older adults: 10 age-specific templates created from 1,600 Japanese subjects aged 20–80 years. Reference templates were created with the T1-weighted images, acquired on a 0.5T MR scanner. More recently, Rorden et al. (2012) created matching CT and MRI templates for use with stroke-aged populations. Fifty MRI scans were used (mean age of participants 72.9 years) all acquired on a 3T scanner. Perhaps most relevant to the current work, Rohlfing et al. (2009) also proposed the use of age-specific, as well as sex-specific, templates. Based on 64 subjects (30 M/34 F) across various ages, they used a combination of regression models and non-rigid registration methods to show that decade-based age-specific templates provided increased accuracy when compared with a study-wide template. The current study follows the strategy of these studies to create age-appropriate templates. The current investigation constructed age-specific brain templates for adults 20–89 years of age. The procedure for template construction involved gathering images from open sources of adult MRIs and using a state-of-the-art non-linear iterative averaging method. The MR images were compiled from our own acquired MRI images, the Pediatric MRI Data Repository created by the NIH MRI Study of Normal Brain Development (NIHPD; Evans and BDCG, 2006; Almli et al., 2007; Waber et al., 2007), the Information Extracted from Medical Images (IXI) database (Heckemann et al., 2003; Ericsson et al., 2008), and the Open Access Series of Imaging Studies (OASIS; Marcus et al., 2007, 2010) cross-sectional and longitudinal image sets. We compared the use of age-specific templates to young adult and age-inappropriate templates for brain tissue segmentation, with the hypothesis that age-specific templates will provide less biased tissue segmentation. We did this by using the manually segmented images from the Internet Brain Segmentation Repository (IBSR; Filipek et al., 1994; Rohlfing, 2012). Manual segmentation is often considered a “gold standard” for tissue segmentation, though this process is both time-consuming and subjective (Dale et al., 1999). However, results from automated procedures may be compared with the manual segmentation to compare the relative accuracy of the automated procedures (e.g., Atkins et al., 2002; Valverde et al., 2015). We also demonstrated the potential usefulness of the age-specific templates by highlighting their ability to confirm characteristics of brain aging. We have previously constructed age-specific templates for 2 weeks through 4 years of age (Sanchez et al., 2012a) and 4.5 through 20–24 years of age (Sanchez et al., 2012b). Thus together with the current study, we have constructed a database of normative age-appropriate average MRI templates across the lifespan (Richards and Xie, 2015). These templates (head, brain, and segmented priors) are publicly available through our website.²

The database consists of age-specific templates divided into five-year increments, 10-year increments and multi-year age groups (25–39, 40–59, 60–89). Templates for T1W head and T1W brain exist for each age group. Additionally, PVEs and binary-segmented images were created for each template from the two different segmentation methods: Image (no priors) and MNI priors. The T1W average templates represent gray matter values normalized to 100, so that variations in voxel intensity between MRI volumes did not affect the averages. All averaged templates result from an initial linear registration and transformation with the MNI-152 template. Thus, the templates are loosely oriented to the MNI-152 volume. Subsequent steps utilized nonlinear transformations with age-specific templates created from prior steps, which minimized the influence of the MNI-152 template. The resulting age specific head and brain templates represent the average size for the participants at that specific age.

Axial (A) and mid-sagittal (B) slice for each five-year age-specific T1W brain (A) and head (B) template are pictured in Figure 3. The sagittal view of the averaged templates exhibits small morphological structures in fine detail, as can be seen with cortical and subcortical anatomy. Gradual brain atrophy is noticeable especially when comparing the templates across age. All templates appear to be relatively consistent in regards to level of detail and clarity.

The purpose of the current work was to create a set of age-specific adult brain templates that could be used for anatomical image processing. Our rationale for creating age-specific templates as an alternative to study-specific templates or young-adult templates (e.g., MNI) was two-fold. First, we wanted to create age-appropriate templates that could be used for multiple studies and second, we wanted to create templates in increments small enough to meaningfully represent neuroanatomical changes over age. We were successful in creating a set of templates that meaningfully represent neuroanatomical changes that occur throughout adulthood. The templates are based on a large number of images (>1000) and a wide range of ages (20–89), and provide five-year, 10-year, and multi-year increments. The age-specific templates created for this study should provide increased accuracy when used for the tissue classification steps often used with structural MRI analysis pipelines. The templates used in the current study are available online.⁴

The issues with using brain templates derived from young adult data for studies with special populations have been well-documented (Huang et al., 2010). These limitations are especially apparent when using young adult reference data as a tool for segmenting images from a range of ages into the different tissue types. As a potential solution, studies investigating differences in brain volume with VBM began using study-specific templates to improve tissue classification results. Early investigations comparing a study-specific template to the MNI template for tissue segmentation found that the study-specific template provided less biased estimates (Ashburner and Friston, 2000; Good et al., 2001; Thompson et al., 2001). Similar to these studies, we found that age-specific tissue priors provided less biased tissue estimates than did MNI or other young-adult tissue priors, especially for older adults. Age-specific five-year or 10-year tissue priors performed similarly and may prove equally useful in practical application. Additionally, we found that age-specific tissue average priors created with MNI priors as a base reference, while better than simply using the MNI, provided more biased tissue estimates than age-specific tissue priors that did not start with MNI base priors, especially for older adults. Aging investigations using study-specific priors that start with MNI base priors may offer a large increase in accuracy over simply using the MNI (e.g., Lemaître et al., 2005), however our results suggest age-specific tissue priors that do not start with MNI priors as a base reference may provide even better results.

The database from which we constructed the templates had approximately equal numbers of males and females. Generally template construction used for young adults does not distinguish gender. Our templates come primarily from Caucasian participants, though the exact distribution of racial and ethnic groups is unknown. It has been shown by others that templates based on Asian young adult samples have markedly differing characteristics to those based on primarily Caucasian samples. For example, Lee et al. (2005) found that a standard Korean brain template created based on Korean adults was shorter, wider, and less in height compared to the ICBM-152 template. Tang et al. (2010) created a Chinese MRI brain template and found similar results. This suggests that age (Huang et al., 2010) and race/ethnicity (Lee et al., 2005; Tang et al., 2010) may both be important factors when considering MRI template usage.

We also evaluated the templates with a small external validation dataset not used in the creation of the templates. We did this by using the manually segmented images from the IBSR (Filipek et al., 1994; Rohlfing, 2012) and comparing the automated segmentation with our segmenting priors applied on those MRI volumes with the manually segmented GM and WM. The age-appropriate average priors created with no base reference (see Figure 2) were the best fitting method for the external validation data set. This method outperformed MNI priors, age-appropriate priors with the MNI as the base reference set, and young-adult template priors (see Figure 4). These results suggest our age-specific templates will provide more accurate tissue segmentation in future datasets. The IBSR data volumes might be considered a “gold standard” manual segmentation against which to compare the classification methods. These comparisons show how well the age-appropriate average priors performed relative to a manual segmentation.

The quality and usefulness of these age-specific templates is validated through their ability to replicate age-related volumetric change trends in the literature. We found changes in GM and WM over adulthood with the segmented PVE volumes based on the participants’ age-appropriate five-year template. Similar to prior studies, the results showed a linear decrease in GM volume over the course of adulthood (Good et al., 2001; Ge et al., 2002; Sato et al., 2003; Sullivan et al., 2004; Taki et al., 2004; Lemaître et al., 2005; Smith et al., 2007), a quadratic change in WM, with slight increases up to age 50, followed by decreases thereafter (Ge et al., 2002; Lemaître et al., 2005; Salat et al., 2009), and a slight associated increase in OM (Good et al., 2001; Lemaître et al., 2005; Smith et al., 2007). There was a small apparent increase in the amount of GM at our oldest two ages (Figure 8), which seems different than the regular decline in GM at these ages reported in other study. However, this may be due to sampling biases or other exclusion factors in these data. It is worth noting that a similar analysis of these adult data along with child and adolescent data from our “Neurodevelopmental MRI Database” (Richards and Xie, 2015) showed that the GM volume peak occurs somewhat earlier, in late childhood or early adolescence (e.g., 10–12 years). This implies that the age trends noted in several studies reflect only the changes from the youngest age of the participants in the study, generally young adults.

The method of template creation that we used here should prove useful to the research and clinical communities, in particular for researchers focused on neuroimaging and aging. We utilized publicly available image post-processing software programs (FSL, ANTS) to produce our age-specific templates. Similar to others (Fonov et al., 2011; Sanchez et al., 2012a,b), the age-specific templates were created through an iterative approach that minimized the influence of MNI a priori data and maximized the preservation of the different sizes, shapes and tissue distribution of the adult data. The processing pipeline refined the images recursively, such that the optimization procedure was applied to the data at different resolutions, with successively higher resolution during the nonlinear registration. Our pipeline procedure should be useful to others who wish to create adult templates based on different parameters using state-of-the-art averaging programs.

Discrepancies between aging brains and the current reference data available continue to be an issue and the need for age-specific reference data is apparent. Several large-scale normative samples have provided the opportunity to address the need for age-specific brain templates. We implemented an automatic image-processing pipeline for template creation that has worked well with pediatric and infant images (Sanchez et al., 2012a,b) and created age-specific adult templates using state-of-the-art averaging techniques. The age-specific templates provide more precise a priori information to guide segmentation. Similar to study-specific templates (Ashburner and Friston, 2000; Good et al., 2001; Thompson et al., 2001), use of age-specific reference data will hopefully facilitate the generation of more reliable conclusions about the volumetric changes that occur with aging. The images from the current study complement our earlier studies and result in a database of normative age-appropriate average MRI templates across the lifespan (Richards and Xie, 2015).

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.