Part of Advances in Neural Information Processing Systems 17 (NIPS 2004)
Finn Nielsen
We provide a method for mass meta-analysis in a neuroinformatics database containing stereotaxic Talairach coordinates from neu- roimaging experiments. Database labels are used to group the in- dividual experiments, e.g., according to cognitive function, and the consistent pattern of the experiments within the groups are de- termined. The method voxelizes each group of experiments via a kernel density estimation, forming probability density volumes. The values in the probability density volumes are compared to null-hypothesis distributions generated by resamplings from the entire unlabeled set of experiments, and the distances to the null- hypotheses are used to sort the voxels across groups of experi- ments. This allows for mass meta-analysis, with the construction of a list with the most prominent associations between brain ar- eas and group labels. Furthermore, the method can be used for functional labeling of voxels.
1 Introduction
Neuroimaging experimenters usually report their results in the form of 3- dimensional coordinates in the standardized stereotaxic Talairach system [1]. Auto- mated meta-analytic and information retrieval methods are enabled when such data are represented in databases such as the BrainMap DBJ ([2], www.brainmapdbj.org) or the Brede database [3]. Example methods include outlier detection [4] and iden- tification of similar volumes [5].
Apart from the stereotaxic coordinates, the databases record details of the exper- imental situation, e.g., the behavioral domain and the scanning modality. In the Brede database the main annotation is the so-called "external components"1 which are heuristically organized in a simple ontology: A directed graph (more specifically, a causal network) with the most general components as the roots of the graph, e.g.,
1External components might be thought of as "cognitive components" or simply "brain functions", but they are more general, e.g., they also incorporate neuroreceptors. The components are called "external" since they are external variables to the brain image.
WOEXT: 41 Cold pain
WOEXT: 40 WOEXT: 261 Pain Thermal pain
WOEXT: 69 Hot pain
Figure 1: The external components around "thermal pain" with "pain" as the parent of "thermal pain" and "cold pain" and "hot pain" as children.
"hot pain" is a child of "thermal pain" that in turn is a child of "pain" (see Figure 1). The simple ontology is setup from information typically found in the introduction
section of scientific articles, and it is compared with the Medical Subject Headings ontology of the National Library of Medicine. The ontology is stored in a simple XML file.
The Brede database is organized, like the BrainMap DBJ, on different levels with scientific papers on the highest level. Each scientific paper contains one or more "experiments", which each in turn contains one or more locations. The individual experiments are typically labeled with an external component. The experiments that are labeled with the same external component form a group, and the distribu- tion of locations within the group become relevant: If a specific external component is localized to a specific brain region, then the locations associated with the external component should cluster in Talairach space.
We will describe a meta-analytic method that identifies important associations be- tween external components and clustered Talairach coordinates. We have previously modeled the relation between Talairach coordinates and neuroanatomical terms [4, 6] and the method that we propose here can be seen as an extension describing the relationship between Talairach coordinates and, e.g., cognitive components.