 Hi, I'm presenting the IBM tool box which is an extension for SPM, a well-known new imaging software, and first we'll have an overview of meta-analysis in your imaging, so why would we like to do that and what's the difference between the two main types of meta-analysis namely coordinate-based and image-based. We'll then focus on image-based meta-analysis and have a look at the different approach that are available to do that, and finally we'll quickly review some results investigating the validity of those different approaches in the context of new imaging. So first, meta-analysis in new imaging, so why will we do meta-analysis? In a typical new imaging study you start by designing an experiment, you acquire some data, you analyze it, and you end up with some results. If we look at the literature we have plenty of studies and meta-analysis is simply an analysis that will combine the results of previous studies to get new results. In the context of new imaging the main reason for doing that is to increase power because it is well known that most studies have a small sample size, so increasing the power might help us in discovering more subtle effects or getting more robust results. But conducting a meta-analysis is also a way to combine results in a novel fashion from previous studies. So if we have a closer look at what's the data available in a typical new imaging study at the different stage, so when we start our experiment we have well new imaging data. Then after the data acquisition phase we have about in a typical study 500 megabyte per subject leading to 20 gigabyte if we take a typical study with 40 subjects. By the time we have we got our results we have five times more data about 2.5 gigabyte per subject. But when we publish our results the data that we actually share is very very limited. So if we have a closer look at what are we sharing exactly so first we have a qualitative view of the detections like 2D figures but what's really the quantitative part of the results in a new imaging study is usually a table displaying the XYZ coordinates of the local maxima and the value of statistics at these different points. In fact these could be represented as a brain volume with really sparse value corresponding to the local maxima. So currently if I want to perform a meta-analysis based on published results the only thing I can do is to combine this really sparse map into a coordinate based meta-analysis. But there is an increasing interest towards data sharing and we could imagine that if in the near future we share more results when reporting a new imaging study we would be able to perform an image-based meta-analysis and this has been shown to be the optimal approach because you take into account all the value brain-wise instead of focusing on a few peaks. So when we look at image-based meta-analysis the gold standard in a statistical point of view. So first if we look at the typical study we start with one subject we have pre-processed data, we fit a model, estimate and we get contrast and standard map. We can threshold these maps or statistical map and get detections. We usually have several subjects and at the level of the study we combine the contrast and standard maps to get a new contrast and standard map at the study level. Again special dean give us the detection. Now when we have several studies and we want to perform a meta-analysis we do exactly the same approach. We go back to the contrast and standard map and combine them into a new model fitting and estimation. So in fact the gold standard approach is a three-level approach with a subject level, a study-level and a meta-analysis level. And to be able to do the meta-analysis level which is what we are interested in we need the contrast and standard maps for each study and we need to make sure that these are expressed in the same units. But that's where it's getting a bit difficult. The units of the contrast maps will depend on a lot of parameters including the scaling of the data at the subject level, the scaling of the predictor at the subject and the study level and the scaling of the contrast at the subject and the study level. On top of that contrast estimate and standard map are in fact rarely shared. So there are in fact all the statistical approach that can be used in the context of image-based meta-analysis. They are known to be suboptimal but they are based on limited data, for instance, on the Z statistic. So instead of sharing the contrast and standard error map we could only share the Z statistic map. So other approach are based on the Z statistic and the sample size and so on. But it's important to keep in mind that all of these approach are suboptimal. They are based on restrictive assumptions and in order to use them in your imaging we need first to check that these assumptions are verified in your imaging or that these approach are robust to violation to their assumptions. And that's where we introduced the IBM Toolbox which is a plug-in for SPM available on GitHub that provides an implementation of all the methods that I've been presenting in the previous slide. So just to give you a small view of some results that we got using the Toolbox. So we investigated 21 pain studies in healthy subjects and we combined them in a one-sample test. So what you have on the left of this picture is the gold standard approach and the approach is presenting on a gray background are the one that takes into account a possible study heterogeneity between studies while the one on the white background are approaches that assumes that there is no between study variance. Even qualitatively from this image you can see that the approach on the white background that ignore between study variance tends to overestimate the statistics leading to a higher rate of false positive. So in conclusion we have presented the IBM Toolbox, the idea is to support future image-based meta-analysis. In practice it's quite difficult to perform the gold standard approach which is a third level mixed effect in your model. The IBM Toolbox is here to provide different approach and we also have quantitative results that I'll be happy to discuss with you at my poster and we also plan on having in investigating other models including two sample analysis. So I would be happy to answer any questions. How much do you think, I mean this also requires some standardization of labeling what is a pain study and you know you need to know how to kind of represent to a registry of like this is an fMRI analysis of a pain study. This is an fMRI analysis of a working memory study and so on. Yeah that's very true. Is there any plan to sort of help prove that? I know you're working on festival things to do that. Do you see that as the main impediment for the meta-analysis or do you see... I mean we clearly need more tools to be able to define our studies in a more standard way in order to be able to do meta-analysis and this is only part of the story.