MRI provides excellent spatial resolution and tissue contrast and is thus ideally suited for morphological analyses of the brain. To allow for the comparison of imaging studies across individuals, imaging data has to be spatially normalized. The most widely used approach to spatial frameworks is the one described by Talairach and Tournoux (1), based on a single (elderly) female brain, which is different from larger collections of normal brains used in later collections (2). Spatial normalization is generally based on adult data and thus poses special problems in children, since pediatric brains differ in size, composition, and shape from adult brains (3–5).
With regard to morphological studies of the human brain, several methods utilize a priori brain information to classify brain tissue, including a widespread method for the systematic and automated analysis of structural brain data (voxel based morphometry, VBM) put forward by Ashburner and Friston (6). This method utilizes tissue a priori information from an adult reference population, and the applicability of this data to a pediatric population will depend on the differences between these populations.
The question of how substantial the differences are between reference data derived from a normal pediatric population and adult data is the subject of this study. We set out to construct normative pediatric brain data and this dataset was compared to standard adult data, which is available in a widely used VBM software solution, SPM99 (Wellcome Department, University College, London, UK).
MATERIALS AND METHODS
MR images of 200 healthy children were used in this study. Subjects were recruited as part of an ongoing study on normal language development (7). Institutional review board approval and informed consent were obtained for all subjects. Rigid exclusion criteria were applied to ensure a normal pediatric population (5). All MRI scans were read for structural abnormalities by a qualified pediatric neuroradiologist and excluded from further processing if abnormal findings were present.
Data Acquisition and Preparation
Children were imaged with a Bruker Biospec 30/60 3 Tesla MRI scanner equipped with a head gradient insert (Bruker SK330). A whole-brain, 3D T1-weighted modified driven-equilibrium Fourier transform (MDEFT (8)) image was acquired (TR = 15 ms, TE = 4.3 ms, τ-time = 550 ms, flip angle = 20°, FOV = 19.2 × 25.6 × 14.4 cm, matrix = 128 × 256 × 96, resolution = 1.5 × 1 × 1.5 mm). All of the automated image processing was done using statistical parametrical mapping software (SPM99) or stand-alone scripts running in MATLAB (MathWorks, Natick, MA) unless stated otherwise.
As the single manual step in image preparation and analysis, manual determination of the anterior commissure was performed by a single investigator for all images. During this process, images were also carefully realigned along the main axes to correct for different head positions in the scanner, providing optimal starting estimates for the following normalization procedure. Images were rated (regarding the presence of arterial blood flow artifacts and motion artifacts), as described before (5), and low-quality images were excluded. The remaining high-quality images (n = 148) were then resliced to 1 × 1 × 1 mm isotropic voxels in axial orientation to reduce partial volume effects in further processing and to achieve a better fit with the axially oriented templates. As in all of the other processing steps within SPM99, a sinc-interpolation algorithm (9 × 9 × 9 neighbors) was used when possible. In order to minimize partial volume effects during image processing, all images in all steps were written out to 1 × 1 × 1 mm isotropic voxels.
Our aim was to construct normative pediatric brain data to allow for the comparison with adult data. These data comprise whole-brain templates and data derived from the whole-brain images (derivative data), i.e., tissue probability maps for gray matter (GM), white matter (WM), and cerebro-spinal fluid (CSF). In order to visualize the effects of data processing, we employed four different processing strategies, using the fully automated spatial normalization and segmentation routines implemented in SPM99. This segmentation utilizes a combined pixel intensity and a priori knowledge approach, using (adult) prior probability maps for gray matter, white matter, and CSF to make an initial probability estimate as to which tissue type a voxel most likely belongs to and then proceeds to do a cluster analysis with a modified mixture model (9). There is as yet no pediatric data available to be used instead.
We used the “classical” way of data processing (6) by normalizing the whole brain dataset into stereotaxic space, followed by its segmentation into GM, WM, and CSF. An “optimized” processing protocol was recently put forward (10), which we also implemented. Here, images are initially segmented in native space. These are then spatially normalized, using the corresponding tissue probability maps as the target image (e.g., a native-space gray matter image is normalized to the GM probability map), which bases the normalization parameters heavily on the tissue class under investigation and minimizes the influence of nonbrain and other tissue. These “optimal” normalization parameters are reapplied to the original whole brain image, which is segmented to yield an “optimally segmented,” normalized GM image. The procedure is repeated for WM and CSF.
Furthermore, in order to minimize the influence of adult a priori information on the resulting pediatric template, we applied a “two pass”-approach to both the classical and the optimized processing strategy: in the first pass, processing was done based on the adult data. In the second pass, all processing was based on the pediatric a priori data obtained during the first pass. Our analyses thus resulted in four datasets, which were named after our institution (Cincinnati Children's Hospital Medical Center): 1. CCHMC_fp (“first pass”: whole-brain template and a priori data); 2. CCHMC_sp (“second pass”: whole-brain template and a priori data); 3. CCHMC_oa (“optimized adult”: a priori data); 4. CCHMC_op (“optimized pediatric”: a priori data). These four pediatric datasets were then compared to the standard adult data (SPM). Figure 1 summarizes our data processing strategies (see legend and text below for full details).
All images were automatically transformed into stereotaxic space within SPM99 by a 12-parameter affine-only transformation (six-parameter rigid body, three zoom, three shear (11)). All resulting images were then averaged and, by modeling smoothly varying intensity changes, residual image inhomogeneities were removed (6). This image was (according to SPM-template specifications) smoothed with an 8 mm Gaussian kernel to facilitate fitting. This first-pass template (CCHMC_fp, and the first-pass brain mask, see below) was used as the target in the second-pass normalization, resulting in the second-pass template (CCHMC_sp).
Since the optimized processing protocols require the creation of not one, but three normalized whole-brain images (based on three sets of “optimal” normalization parameters for GM, WM, and CSF), there is no continuity constraint between these images requiring that the tissue classes smoothly overlap. Therefore, no whole-brain template could be constructed from these images.
In order to obtain pediatric a priori data, we set out to create derivative data, i.e., probability maps of GM, WM, and CSF, from the original images. Segmentation in all four processing strategies was done using the image inhomogeneity-correction (6), which was recently shown to markedly increase the reproducibility of SPM99-segmentation results (12). Again, processing was based on adult (CCHMC_fp, CCHMC_oa) or pediatric a priori data (CCHMC_sp, CCHMC_op). In order to reduce misclassified voxels and to minimize nonbrain tissue, each individual gray and white matter image was modulated with its “extracted brain” (determining if gray or white matter is present with a sufficient likelihood). The resulting, “masked” images were averaged and processed as mentioned above.
The brain mask, used in SPM99 to mask out nonbrain tissue, was derived as mentioned above. Missing data from the ventricles was added using ROIs drawn in MRIcro (v. 1.32 (13)), which was also used for cortical renderings.
Assessment of Differences
The pediatric a priori data obtained above was compared to the corresponding adult data (included in the SPM99 distribution) using the “image calculation” feature within SPM99. For the purpose of comparison, we prepared our pediatric probability maps according to the adult data (2 × 2 × 2 mm and smoothed by 8 mm). Size differences (see below) were accounted for by scaling the adult data to the pediatric data by an affine-only procedure. In order to preserve the original pixel intensities, this individual scaling was followed by a modulation with the Jacobian determinant (5). This ensures that all observed differences are due to tissue probability differences, not differences in size.
To further assess developmental influences on our results, we divided our pediatric sample into three equally sized subgroups (termed “young,” “medium,” and “old”). The derivative data obtained in the first-pass classical processing (CCHMC_fp) was averaged separately for each of these groups and was then compared to the adult data as described above. In order to detect global developmental trends in these tissue probability maps, we investigated the global image intensity in the native-space images that were segmented using pediatric a priori data (i.e., the first images in the CCHMC_op protocol [#4 in Fig. 1]).
Two calculations were performed on all images: adult > pediatric, and adult < pediatric. To account for different image contrasts due to different scanning procedures, field strengths, field inhomogeneities, etc., results were only considered valid if the tissue probability difference was found to be larger than 20%. All results are in neurological orientation.
Imaging data from 52 children (26%: 28 girls [54%], 24 boys [46%]) was rejected due to a quality factor of three or four (40 scans), technical failures (10 scans), or pathological findings (two scans). This left data from 148 children (79 girls [53.4%], 69 boys [46.6%]). The age and gender breakdown of this sample is given in Fig. 2. Average age was 135.87 ± 41.9 months (11.32 ± 3.49 years), median = 131 months (10.92 years), range 60–226.5 months (5–18.87 years) at the date of the MR exam. Ethnic origin was Caucasian in 132 (89.19%), African American in six (4.05%), Asian in five (3.38%), Multiethnic in two (1.35%), Hispanic in two (1.35%), and Native American in one (0.68%). Handedness was right in 133 children (89.86%) and left in 15 children (10.14%).
Breaking down the sample into three equally sized subgroups yielded “young” (49 children, 7.59 ± 1.20 years), “medium” (49 children, 10.91 ± 0.86 years), and “old” subgroups (50 children, 15.38 ± 1.88 years). The approximate group delineation is denoted in Fig. 2 by gray brackets.
Differences Between Pediatric and Adult Templates and A Priori Data
Comparing the pediatric whole-brain templates with the adult data showed that the pediatric brains underwent a stronger linear scaling: the pediatric template brains fill a larger space within the template's bounding box (Fig. 3).
High-quality a priori probability maps could be constructed in all processing approaches, with the resulting (smoothed) images already displaying only minimal, visually almost inapparent differences (Fig. 3). The result of the two calculations is shown in Fig. 4, with the color-coding representing the direction of changes: blue denotes a higher tissue probability in the adult data (adult > pediatric), while red signifies a higher tissue concentration in our data (adult < pediatric). The differences shown represent an at least 20% difference in tissue probability.
Developmental Trends in the Pediatric Data
The image intensity of the native-space tissue probability maps (mean and SD, in arbitrary institutional units [IU]) for the three age groups is plotted in Fig. 6. A strong overall negative correlation with age could be shown for GM (r = –0.38, P < 0.00001), while an overall positive correlation was found for CSF (r = 0.22, P = 0.006). The correlation of WM and age did not reach significance (r = 0.12, P = 0.14). GM was also significantly different between the medium and old (P = 0.0004) and the young and old subgroups (P < 0.00001). CSF was significantly higher in the old when compared to the young subgroup (P = 0.03).
Figure 7 shows an overlay of the differences between the young, medium, and old first-pass pediatric data and the adult data for GM (top) and WM (bottom), overlaid on the corresponding (smoothed) tissue maps. Color-coding is as in Fig. 4. The differences shown represent an at least 20% difference in tissue probability.
Our sample covers healthy children and adolescents from 5–18 years, which are most likely to actively participate in pediatric neuroimaging studies. “Apparently normal” imaging data from clinically referred younger children was used in the past; this, however, will introduce an uncontrollable bias into the study population (14, 15). We thus avoided the risk of using a “pseudo-normal” collective, allowing for the interpolation of our findings to the population as a whole (15).
The (ethically not justifiable) necessity of sedation makes it difficult to include data from normal younger children. Also, strong developmental changes make it problematic to incorporate children on the extreme ends of our spectrum into the same template. However, since our aim was to compare “adult” to “pediatric” data, the incorporation of the whole spectrum of pediatric brains available to us seemed to be the appropriate procedure. In order to further assess developmental trends, we divided our data into three subgroups. The results of these analyses are discussed below.
The adult template is based on 152 adults (mean age = 25.0 ± 4.9 years, range = 18–44 years), while our template (n = 148) covers 14 years of life. Handedness in our sample (right = 89.86%, left = 10.14%) is about equal to the adult data (right = 84.87%, left = 9.21%, unknown = 5.92%), and (with about 90% right-handers) in good agreement with the literature (16). Our sample (girls = 53.4%, boys = 46.6%) is even more balanced than the adult group (female = 43.4%, male = 56.6%) regarding gender composition. Neither the difference in gender composition nor in handedness is statistically significant.
We employed four processing protocols, in a way that reflects an evolution in image data processing towards both maximizing accuracy and minimizing the influence of adult a priori data. Nonlinear spatial normalization (17) was not used, as a close fit to the (adult) template on a regional level was not our goal: instead, we specifically wanted to preserve the possibly different shapes and tissue distributions of our pediatric brains. We then proceeded to obtain derivative brain tissue data using a standard and an optimized protocol, based on both adult and our pediatric data, thus resulting in four datasets (Fig. 1). The optimized approach to segmentation was recently suggested by Good et al. (10); it consists of determining tissue-specific normalization parameters, aimed at optimizing normalization and segmentation. The images resulting from this procedure are not necessarily complementary to each other since they are based on three sets of “optimal” normalization parameters. They thus might overlap (or not overlap), especially in areas susceptible to segmentation errors, precluding the construction of a whole-brain template.
Our pediatric brains result in a minimally larger pediatric template (compared with the adult template), a typical effect seen before in the construction of average templates (18). This does not seem to exert a negative influence on the results, with the resulting averaged image showing excellent delineation of even small morphological structures (Fig. 3). Using pediatric or adult data does not influence this effect, since it is apparent by the same magnitude (1:1.09 vs. 1:1.1), regardless of what data were used in the processing. By using a brain mask in both procedures, the normalization should be based heavily on brain tissue and thus brain size, which has reached 90–95% of adult values by age 5 (3, 14). This difference in size (9–10%) was compensated before the probability maps were compared to each other and does thus not influence the differences in the a priori data shown below.
It should be noted that the templates within SPM99 (with a bounding box of 181 × 217 × 181 mm) conform to the Brainweb-specifications (19), while the SPM99-default image size (156 × 188 × 135 mm) is considerably (and not proportionally) smaller. Therefore, the proportions in the displayed templates might seem slightly distorted. Since our aim was to create a template image, we did not write out images into standard SPM99-space and thus do not know if using our template would have any implications towards reporting locations in Talairach-coordinates (136 × 172 × 118 mm, after doing a conversion to account for the differing dimensions (20)).
As to spatial normalization, surprisingly few studies have been published with regard to the effect of using adult data to normalize pediatric brains. One small study (4), suggesting that normalization based on adult data is feasible in children older than 6 years, was flawed by using an old software version (SPM96, which has since undergone major overhauls (9, 17)) in a small number of neurologically sick children. We could recently show (5) that using such a pediatric template during spatial normalization has a profound, regionally specific, and age-dependent effect on the amount of nonlinear deformation (17) the original image undergoes. With the data obtained in this study, we hope to allow for the further exploration of such effects and the usage of potentially more appropriate reference data.
Differences in A Priori Data
Pediatric vs. Adult Data
Profound differences in tissue distribution are apparent between our pediatric and standard adult data (Fig. 4). The results are most striking in GM: our data indicates that, in general, pediatric a priori information will classify brain tissue as GM with a higher probability when compared to adult data, at the cost, however, of WM and CSF, both of which are present with higher probabilities in the adult data. This trend is slightly abated by the use of the optimized segmentation protocol, but strongly enhanced by using pediatric a priori information in both the classical and the optimized protocol. The differences are most apparent in central brain structures like the basal ganglia, in frontal regions, and in the cingulate gyrus.
Comparing the results from the four processing strategies with each other allows assessment of the effect of processing strategy (Fig. 1, strategies 1 and 2 [standard] and 3 and 4 [optimized processing]) and of using pediatric reference data (Fig. 1, strategies 1 and 3 [adult] and 2 and 4 [pediatric reference data]). This comparison is made for GM in Fig. 5: it shows that only minor differences exist between the results obtained by the standard (top row) and the optimized protocol (bottom row), with both showing comparable patterns. In contrast to this, the effect of using adult (left column) vs. pediatric (right column) a priori data is substantial, leading to a considerable increase of tissue being coded as GM in both the standard and the optimized protocol. Again, the effect is regionally specific, as evidenced by the sparing of distinct tissue areas in all results, and seems most pronounced in central and frontal brain regions.
Normal development heavily influences overall tissue probability in our three groups (young, medium, and old). As from the medium group on, GM significantly declines, while CSF increases significantly in the old group only when compared to the young group. WM changes do not reach significance (Fig. 6). Comparing the three age groups with the adult data (Fig. 7) reveals striking differences: in GM, probability differences are widespread in the young group and become gradually and consistently less as the children get older. The inverse is true for WM: here, a number of distinct areas show a higher probability to be WM in the adult data and become, again gradually and consistently, less prominent with the diminishing age difference between the datasets.
Our results are not surprising, since pediatric brains differ in shape and tissue composition from adult brains (3, 4, 21). The trends apparent in our data are in very good agreement with data on normal brain development obtained in in vivo (3, 14, 22–24) and in vitro studies (25, 26). These show that while GM volume peaks in late childhood / early adolescence and thereafter declines (with a focus on basal ganglia decreases (14)), WM volume continues to increase well into adulthood (22, 23). CSF volume increases in a linear fashion from early childhood on (3). These developmental trends (more GM, less WM and CSF in childhood) are mirrored in our imaging data and account for the profound and age-dependent morphological differences between the two populations. Our findings confirm our hypothesis on developmental trends underlying the tissue probability differences and show our data in line with previous studies on pediatric brain development. It also shows that even our “pediatric data” in itself displays considerable changes with age, implying that an even narrower age-matching may allow for more appropriate data processing. The fact that our data is in such good agreement with existing data on brain development is also reassuring, in that it demonstrates that age, and not technical issues (like image contrast, image processing, or scanning sequence), is the decisive factor that determines the differences between our pediatric and the adult datasets.
Implications for Pediatric Neuroimaging Studies
Using pediatric reference data has a substantial influence on the resulting images, with the effect being most pronounced in GM (Figs. 4, 7). Consequently, considerable differences in GM probability must be expected in the regions showing the strongest differences between the first- and second-pass data, i.e., central and frontal regions, including the cingulate gyrus. Figure 5 shows the dramatic effect of using adult (right column) vs. pediatric data (left column). Therefore, our results suggest that, when analyzing a pediatric sample with adult a priori information, significant tissue misclassification is likely to occur in the regions highlighted in Figs. 4 and 7. This effect is especially relevant for all studies of normal or abnormal brain development, since a strong bias will be inserted into the data already by processing them on the basis of possibly nonappropriate a priori information. Based on our results, we thus recommend using a priori data based on a comparable reference population (i.e., a normal pediatric population) when doing structural analyses involving children.
A point of interest in this context is: Which of our four dataset represents the “true” pediatric tissue distribution best? It could be argued that the first-pass data incorporates most of the characteristics of the pediatric data while retaining the anatomical relation between the probability maps, and might thus be best suited. Then again, the optimized protocol might have preserved the characteristics of the tissue best, since the whole processing is optimized to derive one class only. For both procedures, the second-pass might have further reduced residual “adult” influences, and this influence of using pediatric a priori data is substantial. Possible disadvantages of using the optimized data is the lack of a whole-brain template, while the further increase in the size of the data during the second pass might be held against these datasets. Also, the effect that using the first-pass data had in our analyses seems to support the notion that using this data allows one to uncover the decisive characteristics of the pediatric data. At present, we cannot confirm one or the other hypothesis, and further research on this issue might be necessary to definitively answer these questions.
Possible Limitations of This Study
The optimized protocol was only recently put forward (10), and the adult a priori data was not obtained using this algorithm; since we used this different algorithm in two of our four datasets, it cannot be ruled out that differences between them and the adult data are influenced by the methodological differences, thus not fully reflecting true differences in morphology. However, the changes are very consistent within the four datasets and the differences between the results using the standard and the optimized protocol are not substantial (Fig. 5), so we do not see this as a major confound.
Regarding image processing, it should be kept in mind that our figures show differences between smoothed images, since the unsmoothed adult data is not available. Due to the matched filter theorem, this will lead to the most sensitive detection of differences in the range of the width of the smoothing filter (8 mm (27)). Changes on a smaller or larger scale might not have been detected in our analyses.
Also, age was of course not the only factor differing between our sample and the adult sample contributing to the adult a priori data: there are very likely other confounds (e.g., alcohol consumption, a history of head trauma, etc.) that are present in the adult sample more than in our pediatric subjects. However, from Fig. 7 it seems that age and thus developmental status is the decisive factor contributing to the changes we see, and the differences between our data and the adult data should thus be attributed primarily to the age factor. Moreover, potential pediatric samples will very likely be much closer to our sample with regard to these nuisance variables, which is another point in favor of using pediatric data.
It is important to note that no examination of the effect of using pediatric a priori data in research studies was made here, as we felt this was beyond the scope of the current study. Ideally, this question should be answered by applying the newly derived pediatric template and a priori data in an independently obtained set of pediatric brain images. This is also important since we processed our pediatric data based on its own average, and an influence of this additional similarity between images and a priori data on our results cannot wholly be excluded. To allow for the exploration of these effects in pediatric functional or structural neuroimaging studies, the data obtained in this study is freely available to the scientific community from our website (www.irc.cchmc.org).
We have shown considerable differences in both whole-brain templates and tissue probability maps between our pediatric sample and adult data. Using pediatric a priori data during image processing strongly influences the resulting tissue classification. Since these observations indicate that the use of adult reference data in pediatric neuroimaging studies will very likely introduce a severe bias into the ensuing results, caution should be exercised when interpreting pediatric imaging data obtained on the basis of such adult data, especially when examining developmental changes.
We thank Anna M. Weber Byars, Ph.D., and Richard H. Strawsburg, M.D., for performing the neuropsychological and neurological examinations, and William S. Ball, Jr., M.D., for reading the structural scans. This study would not have been possible without the outstanding cooperation of all the children and parents who participated.