New approaches for coordinate-based meta-analysis of functional neuroimaging studies

Functional neuroimaging has enabled the detailed, in-vivo delineation of functional and dysfunctional activity in the human brain. The amount of inference, however, that can be drawn from each individual study is limited by inherent drawbacks such as the small sample sizes and low reliability. Additionally, all neuroimaging findings pertain to a specific contrast between experimental conditions, while the actual goal is to elucidate general organisational principles of brain (dys-) function. However, given the ever-increasing number of studies in human brain mapping, an abundance of data is now ready to be synthesized and modelled.

Here we present an overview on recent conceptual and methodological developments for quantitative meta-analyses of neuroimaging data.

An important aspect of meta-analyses on functional imaging results is that they have to accommodate the spatial uncertainty associated with the reported coordinates. Classically this was achieved by applying a user-defined kernel, which, however, introduces a subjective element to these quantitative approaches. This was now objectified by empirical estimates of local maxima variability, allowing the modelling of spatial uncertainty associated with reported foci in terms of between-subject and between-template variance.

Earlier algorithms for coordinate-based neuroimaging meta-analyses used fixed-effects inference, assessing an above chance convergence between all activation foci. Evidently, this made these analyses susceptible to false positives derived from clustered foci in a single experiment. This approach was hence revised in favour of a random-effects model that statistically tests for convergence of results between studies, regarding the distribution of foci reported within each study as a fixed property.

In a second step, the previously prevailing permutation methods for assessing the statistical significance in meta-analyses have been replaced by an analytical solution for the computation of the respective null-distributions. This innovation not only yields considerably shorted computation times but in particular allow the simulation of statistical fields in meta-analyses needed to correct the inference for multiple comparisons using family-wise error or cluster-level thresholding approaches.

Besides allowing inference about the convergence of activations reported in different studies on a particular topic, new methods for coordinate-based meta-analyses also enable the comparison between tasks using meta-analytical contrasts.

Finally, new approaches also allow modelling the influence of external variables on the activation probabilities across studies. That is, by inclusion of covariates in meta-analyses it becomes possible to assess the influence of, e.g., sample age on the reported activations. This algorithm is conceptually related to correlation analyses in neuroimaging but operates on the level of studies as the unit of observation. It hence allows inference on the characteristics of experimental results in imaging experiments.

Collectively these developments in quantitative meta-analyses enable the assessment of multiple different paradigms in samples of subjects or patients that far exceed the possibilities of any individual centre. Quantitative meta-analyses could hence represent an important tool for the integration of individual neuroimaging results into generalisable knowledge on the organisation of the human brain and pathomechanisms of neuro-psychiatric disorders.