INTRODUCTION
The magnetic resonance imaging (MRI) brain template is a standardized anatomical reference image that captures average neuroanatomical features of a particular population or an age group to align different images to a standard space by accounting for various shapes, sizes, and orientations [1]. To conduct a comparative analysis, it is important to standardize and reorient brain images to a brain template. Standardization, which is part of the preprocessing step, involves the spatial normalization of individual images to a specific brain template. Various brain templates are available based on age and race, with additional templates accounting for regular updates due to methodological advancements [2,3]. A suitable brain template is selected based on the similarity to the population under study, with preferences given to templates with minimum deformational changes from the original images. This ensures that the template accurately reflects the neuroanatomy of the specific population, leading to more precise analysis and interpretation of brain imaging data.
The Montreal Neurological Institute (MNI) brain template is commonly adapted internationally due to the wide availability of the MNI152 template in various analysis tools and software [4-6]. However, the MNI152 template was constructed for the Caucasian population, which has morphological differences from other ethnic groups, especially those in East Asian countries [7-10]. For example, the Korean template revealed features that were shorter in length and height than did the MNI152 template [11]. Additionally, a task fMRI study demonstrated discrepancies in registering to the Chinese and MNI templates, with different brain activation percentages (72.8% and 80.7%, respectively) and spatial accuracy (the correct location being deeper in the interhemispheric fissure) [12]. These findings imply that misregistration to anatomically different Caucasian brain templates could lead to inaccurate analysis of non-Caucasian populations. Thus, attempts have been made to produce brain templates specific to the East Asian population [6,11].
Furthermore, recent studies have elucidated the need for population-specific brain template construction. Differences in brain morphology, such as shorter and narrower features, within the East Asian population have been found in the Indian_157 template when compared to the Chinese_56 template [10,13]. In addition, a brain microstructural study in Chinese, Malay and Indian neonate populations suggested differences in spinal-cerebellar and cortico-striatal-thalamic circuits among different populations [14]. Further studies have revealed that these differences are due to environmental, genetic, and cultural differences, such as perceptualizing languages [12,15,16]. Together, these studies suggest differences within East Asian countries and ethnicities, emphasizing the need for a population-specific Korean brain template.
For Koreans, two templates are available: the Korean template utilizing 1.5T MR images and the KNE96 template representing the elderly population [11,17]. However, recent advancements in MR techniques and methodology have highlighted the need to update the Korean brain template to represent the Korean population more accurately. For Korean template construction, a conventional method was used wherein one subject was chosen as a reference for affinely registering other brain images, introducing bias toward the selected individual. In contrast, the new methodology utilizes multiple iterations of normalization with a larger sample to generate an unbiased minimally deformed template, which is needed for the Korean population [2]. Moreover, data from individuals ranging in age from 18 to 55 years were used for Korean template construction. Pertinently, research has indicated that brain volume changes over the course of a lifespan [18]. Brain volumes tend to increase rapidly during childhood, reaching a plateau in early adulthood; then, a gradual volumetric loss begins in the late 30s. Hence, the KNE96 template was constructed to better represent the elderly population [17].
Consequently, this study aimed to construct a young adult age-specific Korean brain template that better represents the structural features of the Korean population utilizing the most recent template construction method.
METHODS
Participants
Images from a total of 224 healthy native Koreans were acquired from two separate brain MRI T1 datasets for this study (Table 1). The first dataset (KOR152 dataset) included 152 healthy individuals scanned at Seoul National University Hospital (SNUH). The validation dataset acquired from the Seoul National University (SNU) MRI Center consisted of 72 healthy individuals. All subjects were screened by certified psychiatrists using the SCID-I Nonpatient Edition to rule out any psychiatric illnesses. Individuals with a potential family history (first- to third-degree biological relatives) of psychotic disorders were excluded. Further exclusion criteria included individuals showing signs of substance abuse/dependence, neurological disease, or significant head injury.
All subjects provided written informed consent after receiving a comprehensive explanation of the study procedure (IRB no. H-1201-008-392 and 1905/001-010). For minors, consent was formally obtained from both the participants and their parents. This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of SNUH (IRB no. H-2312-013-1488).
Image acquisition and preprocessing
Structural images of all subjects were obtained with Siemens 3T Magnetom Trio MRI scanners using a T1-weighted 3D magnetization-prepared rapid gradient echo sequence (MPRAGE) on both the SNUH (first dataset, 152 subjects) and SNU (validation dataset, 72 subjects) campuses. The T1-weighted MPRAGE images were acquired with the following sequence parameters: for the first dataset, repetition time (TR)=1.67 s, echo time (TE)=1.89 ms, inversion time (TI)=0.9 s, field of view (FOV)=250 mm, flip angle=9°, voxel size=1 mm3, and slices=208; for the validation dataset, TR=2.4 s, TE=2.19 ms, TI=1 s, FOV=272 mm, flip angle=8°, voxel size=0.8 mm3, and slices=224.
The T1 images of all individuals were preprocessed using Advanced Normalization Tools (ANTs) [19,20]. Brain images were extracted using antsBrainExtraction, and to cross-check the accuracy of the process, two neuroscientists conducted a visual inspection. Next, the brain images were processed to remove inhomogeneities created during scanning. This step involved tissue segmentation into three classes: grey matter, white matter, and cerebrospinal fluid, followed by N4 bias field correction [21].
Brain template construction
We constructed a Korean age-specific brain template based on a previously published high-performance T1-weighted brain template constructed from MIITRA [2] images. Korean age-specific brain template construction involved an iterative process [22] based on the symmetric groupwise normalization method using the buildtemplateparallel (BTP) script [23].
First, preprocessed T1 images of each subject were rigidly aligned to the MNI template (1 mm×1 mm×1 mm) to standardize the images in Talairach space. The initial template (A0 image) was constructed by averaging rigidly registered images. Next, affine registration was conducted with rigidly registered individual T1 images as the input and A0 images as the reference. The mutual information and cross-correlation were selected as the cost function for optimization, followed by minimum deformation construction using the BTP script. The transformation matrix and deformations derived via affine registration were concatenated and applied to preprocessed T1 images to obtain an initial template for affine registration (A1n image). The affine registration process was iterated until convergence was reached, with a Pearson cross-correlation similarity index of 0.9995 between the initial template (A1n-1) and successive iterated templates (A1n).
The final step involved nonlinear registration with the aid of the final A1n image template and rigidly registered T1 images as input. The same method was applied for nonlinear registration, in which all transformations and deformations were concatenated to create an initial template for nonlinear registration (A2n image), ultimately yielding a final template (KOR152) with a Pearson cross-correlation similarity index of 0.9995 after the uttermost iteration.
Additionally, a gender-specific template KOR152 was created using the method described above: KOR152_M for males and KOR152_F for females. These gender-specific templates were constructed to account for potential morphological differences between males and females in the Korean population.
Global brain feature measurements
After the KOR152 template was constructed, the global brain features were measured between different templates using Medical Image Processing Analysis and Visualization software [http://mipav.cit.nih.gov]. Four global features (anterior commissure [AC]-posterior commissure [PC] distance, length, width, and height) were measured as described by Bhalerao [13] and Tang [8] (Figure 1A). Using the sagittal plane, the distance from the AC to the PC was initially measured. The midpoint of the AC-PC distance was used to measure the length of the brain in the axial plane. The height of the brain was measured at the midpoint of the AC-PC line in the coronal plane, and the width of the brain was measured when the midpoint was aligned in both the coronal and the axial planes. Additionally, the ratios of width to length (W/L), height to length (H/L), and height to width (H/W) were calculated for comparative analysis and evaluation of the templates.
Validation
To validate the precision of the KOR152 template, a distinct validation dataset was subjected to registration to different templates (KOR152, MNI152, Indian_157, and CN200) [15,24]. The preprocessing step detailed in the above section (‘Image acquisition and preprocessing’ section) was performed on the validation dataset. The displacement magnitude (dis-Mag) was used to assess the change in distance before and after registration for each template utilizing Euclidean distance [18]. Additionally, the log of the Jacobian determinant (log-Jacob) was calculated to evaluate the volume changes during registration to each template. Differences in dis-Mag and log-Jacob between different templates were statistically significant according to multivariate analysis of variance and analysis of variance results.
To assess the sample size of the KOR152 template, standard templates with various sample sizes were constructed using the KOR152 dataset. A random bootstrapping method was applied to create five different sample-sized templates, with five different samples for each size [2]. A total of twenty-five standard templates were constructed using the procedure described in the aforementioned template construction section, with sample sizes of 20, 40, 60, 80, and 100. The accuracy of each standard template was measured by registering the template to the validation dataset and calculating dis-Mag and log-Jacob before and after registration to each standard template.
RESULTS
Global brain feature measurements
Compared to features of the MNI152 template, the KOR152 template was slightly shorter in length and height but uniform in width (Table 2 and Figure 1B). Compared to those of other Asian brain templates, the AC-PC length and height were largely in line, albeit relatively shorter in length. Despite these variations, the W/L, H/L, and H/W ratios showed consistency across all templates. The KOR152 was more circular in axial view and more elongated in coronal view compared to Indian_157 and MNI152 templates, as indicated in W/L and H/W ratios. In contrast, the KOR152 exhibited a less circular axial view and appeared less elongated in the coronal view compared to the CN200. The KOR152 template was generally smaller than the other brain templates, but the morphological characteristics of the KOR152 template seemed to be positioned between the MNI152 and CN200 templates.
Validation
The T1 images from the validation dataset were normalized to four distinct population templates (Indian_157, MNI152, CN200, and KOR152) to evaluate template variability between different populations. The changes in distance before and after normalization to each template are depicted as dis-Mag in Figure 2A and B, along with the changes in volume determined using log-Jacob (Figure 2C and D). When the data were normalized to the KOR152 template, the least distance change (dis-Mag) was observed compared to that of the other three templates: Indian_157, MNI152, and CN200. Dis-Mag significantly differed between the KOR152 template and Indian_157 (p<0.001), MNI152 (p<0.001), and CN200 (p<0.001) templates after Bonferroni correction. A significant difference was also observed between the Indian_157 and MNI152 templates (p<0.001), the Indian_157 and CN200 templates (p<0.001), and the MNI152 and CN200 templates (p<0.001). Differences were most pronounced in the cerebellum and the inferior temporal cortex (Figure 2A). In the CN200 template, distance changes were generally minimal; however, a major change was noted at the edge of the cerebellum in both hemispheres and in the inferior temporal cortex. When registered to the Indian_157 template, a pattern similar to that of the CN200 emerged but generally showed dispersed changes throughout the cerebellum. For the MNI152 template, there was a general increase in distance change throughout the brain area, with greater attention given to the frontal lobe and cerebellum. Finally, the KOR152 template showed minimal changes in distance, with some alterations at the edges of the inferior temporal lobe.
For the volumetric change (log-Jacob), the KOR152 template exhibited the least volumetric change, followed by the changes in the Indian_157, CN200 and MNI152 templates (Figure 2D). The statistical analysis results significantly differed between the KOR152 and Indian_157 (p<0.01) templates, and an even greater difference, after Bonferroni correction, was obtained when comparing the KOR152 to the MNI152 and CN200 templates (p<0.001). Like in dis-Mag, major areas of volumetric change were visible in the frontal lobe, cerebellum and inferior temporal lobe (Figure 2C), validating the variability when normalizing to different brain templates.
To validate the sample size used for KOR152 template construction, we created templates with different sample sizes, which revealed a decrease in average dis-Mag and average log-Jacob values as the sample size increased (Figure 3). Templates constructed with a smaller sample size showed increased variability and greater divergence from the original data in both dis-Mag and log-Jacob. The KOR152 template showed the smallest changes in both measurements, confirming that an appropriate sample size was used for template construction.
DISCUSSION
In this study, the KOR152 template was constructed using a state-of-the-art method for creating age-specific brain templates in Korean populations. Global brain feature analysis revealed that the KOR152 template generally had a shorter length than the other population templates evaluated (MNI152, Indian_157, and CN200). Further investigation revealed significant differences in volumetric and distance changes before and after registration to different population templates. The KOR152 template showed the smallest volumetric and distance changes compared to those changes in the MNI152, Indian_157, and CN200 templates. Additionally, the current iterative method for nonlinear and affine registration has shown improved precision and adequacy by requiring a suitable number of samples in the range of 120–150, which is lower than previously mentioned by Yang et al. [24] suggesting an adequate sample size. Overall, these findings highlight the importance of creating population-specific templates as well as appropriate sample sizes for template construction to ensure accurate and reliable results in neuroimaging studies.
All the global brain feature measurements except the width were smaller in the KOR152 compared to the MNI152 template. Variations in global brain characteristics influence the overall shape of the brain, leading to a more rounded appearance in the axial view and a more oval shape in the coronal view in the KOR152 template. Such morphological differences between the KOR152 and MNI152 templates indicate that the MNI152 template may not be appropriate for spatial normalization in the Korean population. Similarly, the Indian_157 and CN200 templates had uniform AC-PC lengths and heights of the KOR15 template but differed in overall length and width, suggesting modest morphological variations among Asian populations. In axial and coronal views, the overall shape of the CN200 template was most rounded and oval-shaped, respectively, followed by the KOR152 and Indian_157 templates. This finding illustrated subtle morphological differences among Asian populations, with the KOR152 template positioned between the Indian_157 and CN200 templates. These disparities may contribute to discrepancies, underscoring the morphological distinctions within East Asian populations. Thus, further validation analyses were performed to assess the accuracy and applicability of different templates in the Korean population.
The mean distance change (dis-Mag) before and after registering different templates to the validation dataset revealed distinct differences, with the foremost difference observed in the CN200 template followed by the Indian_157, MNI152, and KOR152 templates. The inferior temporal lobe and cerebellum were mostly affected by distance changes throughout the different population templates. The CN200 template showed minimal differences, specifically at the edge of the cerebellum and inferior temporal cortex. This finding is consistent with the wider width of global brain features, indicating substantial changes in the width compared to the original data when registered to the CN200 template. The Indian_157 template, on the other hand, exhibited slightly greater variation across the brain, particularly in the cerebellum. These findings are in parallel with previous studies that have demonstrated distinctive differences within Asian populations [10,18,24]. Additionally, registration to the MNI152 template revealed widespread changes in distance across the brain, indicating significant deviations from the original data. Finally, the KOR152 template with minimal overall changes in distance highlights the necessity of using population-specific brain templates for accurate analysis.
Volumetric alterations (obtained via log-Jacob analysis) also demonstrated similar patterns. The frontal lobe, inferior temporal lobe, and cerebellum emerged as the primary brain regions that underwent significant changes in volume. Unlike the distance change, the use of the MNI152 template showed the most volumetric changes among the different templates, followed by the CN200, Indian_157, and KOR152 templates; thus, the MNI152 template may not be suitable for accurate analysis of brain volume in the Korean population. Moreover, significant differences within Asian templates in distance and volume changes underscore the importance of utilizing population-specific brain templates for precise and reliable analysis in neuroimaging studies. Therefore, it is crucial to use population-specific brain templates when analyzing neuroimaging data, as demonstrated by the distinct differences in distance and volume changes observed.
In addition, sample size effects were analyzed to validate the sample size used to construct the KOR152 template. A random bootstrapping method was applied to create templates with sample sizes of 20, 40, 60, 80, and 100. The distance and volume changes (determined by the dis-Mag and log-Jacob analyses, respectively) from each sample template were calculated to compare the values with those of the KOR152 template. The analysis of sample size effects demonstrated that as the sample size increased, the distance and volume changes in the brain templates decreased. This illustrated that with the current template construction method, smaller changes can be observed with a sample size of 152 compared to previous studies that demonstrated a sample size of 200 [24].
The limitations of this study included the use of a registration algorithm for registering the validation datasets into different templates. The accuracy of spatial normalization is known to vary depending on the template used as well as the default settings of registration algorithms provided by different programs [25]. Our study adapted the ANTs SyN method for the registration algorithm [17,26], which is widely available and has outstanding registration performance. Second, the number of samples used for template construction should be noted. Although a previous study of the state-of-the-art method showed that the sample size plateaued at approximately 120 and 150, further studies could be beneficial for template construction if an even larger sample size were used to validate the suitability of the sample size effect on the construction. Finally, the effect of sex should be interpreted with caution due to the disproportionate female-to-male ratio used for the construction of the KOR152 template. To ensure that minimal confounding factors affected template construction, we validated the results with a proportionate dataset of males and females, which revealed the appropriateness of our KOR152 template to the Korean population. However, further study could be beneficial for validating the effect of sex.
In conclusion, this study has important implications. The MNI152 template and the Korean template [11] were constructed with a 1.5T scanner with bias due to manual segmentation techniques. 3T scanner images and the state-of-the-art method for template construction were used to capture Korean-specific morphological features with greater accuracy. Furthermore, in alignment with prior studies, it is evident that differences exist within the Asian population. This further emphasizes the necessity for a population-specific template to ensure accuracy in neuroimaging studies. The KOR152 T1-weighted template is available for download at https://identifiers.org/neurovault.collection:16289.
