Template-O-Matic: A toolbox for creating customized pediatric templates

Abstract

Processing pediatric neuroimaging data is a challenge due to pervasive morphological changes that occur in the human brain during normal development. This is of special relevance when reference data is used as part of the processing approach, as in spatial normalization and tissue segmentation. Current approaches construct reference data (templates) by averaging brain images from a control group of subjects, or by creating custom templates from the group under study. In this technical note, we describe a new, and generalized method of constructing such appropriate reference data by statistically analyzing a large sample (n = 404) of healthy children, as acquired during the NIH MRI study of normal brain development.

After eliminating non-contributing demographic variables, we modeled the effects of age (first, second, and third-order terms) and gender, for each voxel in gray matter and white matter. By appropriate weighting with the parameter estimates from these analyses, complete tissue maps can be generated automatically from this database to match a pediatric population selected for study. The algorithm is implemented in the form of a toolbox for the SPM5 image data processing suite, which we term Template-O-Matic. We compare the performance of this approach with the current method of template generation and discuss the implications of our approach.

Introduction

Magnetic resonance imaging (MRI) has become the imaging method of choice for developmental neuroscience as it offers a non-invasive window into the development of the human brain (Schaer and Eliez, 2007, Wilke and Holland, in press). Despite the difficulties associated with scanning children (Byars et al., 2002), numerous studies have now used MRI to describe normal and abnormal brain development (Castellanos et al., 2002, Giedd et al., 1996, Gogtay et al., 2004, Gothelf et al., 2007, Lenroot et al., 2007, Peterson et al., 2003, Reiss et al., 1996, Schmithorst et al., 2005, Wilke and Holland, 2003, Wilke et al., 2003b).

However, processing of MR images from children poses distinct problems because several steps in image post-processing require implicit or explicit use of reference data derived from adults. For example, routine procedures like tissue segmentation or spatial normalization, if based on adult reference data, have the potential of introducing a severe bias into pediatric imaging data (Hoeksma et al., 2005, Machilsen et al., 2007, Muzik et al., 2000, Wilke et al., 2003a). Spatial normalization of pediatric imaging data based on adult prior probability maps has been suggested to be sufficiently accurate for coarse-resolution fMRI data, following smoothing (Burgund et al., 2002, Kang et al., 2003), but errors become increasingly important when using structural or functional imaging data with a higher resolution. This emphasizes the importance of using appropriate reference data when processing pediatric imaging data, ideally based on a large sample of subjects reflecting the characteristics of the population under study (Good et al., 2001).

An average template created from a small number of reference subjects may not capture enough variance in the template and may also introduce bias. Acquiring normal, age-appropriate, pediatric brain image reference data is difficult, costly and time consuming, so it is not feasible for every pediatric brain imaging study to construct its own template based on a large normative sample. Until recently, pediatric brain image reference data from a large control population was not readily available to those wishing to construct a template matched to a given study population. This has now changed with the completion of a large-scale MRI study of normal brain development, conducted over several sites in the USA (Evans et al., 2006). It would be straightforward to process this data as described before (Wilke et al., 2002, 2003a) and to generate appropriate average templates from this normal database. However, such an average template created from reference subjects will also bring about unwanted effects, such as not capturing enough variance especially when only few subjects contribute to the template.

As an alternative to using static averages of individual probability maps, here we propose a dynamic statistical approach. This is implemented by analyzing the normal database and then reconstructing appropriate reference data, not from the individual datasets, but from a statistical model that estimates the influence of all variables of interest on the tissue probabilities. For example, if the influence of age is appropriately modeled within each voxel, this information can then be used to construct a prototypical gray matter map for any given age (within the range of available reference data). This transcends previous approaches to provide pediatric (Wilke et al., 2002, Wilke and Holland, 2003) or adult (Hill et al., 2002, Mazziotta et al., 1995, Mazziotta et al., 2001) reference data as the influence of specific variables of interest can be isolated and unwanted sources of variance can be removed from the data. We believe this approach represents a significant step forward for brain image data analysis in general and for pediatric neuroimaging studies specifically, where it is imperative to consider explicitly dynamic morphology to avoid biasing results. With this technical note, we describe such an approach.