Meta-analysis of neuroimaging data: a comparison of image-based and coordinate-based pooling of studies.
Citation: Salimi-Khorshidi, G., Smith, S. M., Keltner, J. R., Wager, T. D., Nicholos, T. E.. (2009). Meta-analysis of neuroimaging data: a comparison of image-based and coordinate-based pooling of studies. Neuroimage, 45, 810-23.
With the rapid growth of neuroimaging research and accumulation of neuroinformatic databases the synthesis of consensus findings using meta-analysis is becoming increasingly important. Meta-analyses pool data across many studies to identify reliable experimental effects and characterize the degree of agreement across studies. Coordinate-based meta-analysis (CBMA) methods are the standard approach, where each study entered into the meta-analysis has been summarized using only the (x, y, z) locations of peak activations (with or without activation magnitude) reported in published reports. Image-based meta-analysis (IBMA) methods use the full statistic images, and allow the use of hierarchical mixed effects models that account for differing intra-study variance and modeling of random inter-study variation. The purpose of this work is to compare image-based and coordinate-based meta-analysis methods applied to the same dataset, a group of 15 fMRI studies of pain, and to quantify the information lost by working only with the coordinates of peak activations instead of the full statistic images. We apply a 3-level IBMA mixed model for a "mega-analysis", and highlight important considerations in the specification of each model and contrast. We compare the IBMA result to three CBMA methods: ALE (activation likelihood estimation), KDA (kernel density analysis) and MKDA (multi-level kernel density analysis), for various CBMA smoothing parameters. For the datasets considered, we find that ALE at sigma=15 mm, KDA at rho=25-30 mm and MKDA at rho=15 mm give the greatest similarity to the IBMA result, and that ALE was the most similar for this particular dataset, though only with a Dice similarity coefficient of 0.45 (Dice measure ranges from 0 to 1). Based on this poor similarity, and the greater modeling flexibility afforded by hierarchical mixed models, we suggest that IBMA is preferred over CBMA. To make IBMA analyses practical, however, the neuroimaging field needs to develop an effective mechanism for sharing image data, including whole-brain images of both effect estimates and their standard errors.