Learning with multi-site fMRI graph data

Neuroimaging data collection is very costly, and acquisition is commonly distributed across multiple sites. However, factors such as different noise characteristics or inhomogeneities make it difficult to successfully combine multi-site functional imaging data. Here, we show that the distribution of signal quality measures across scanners can be significantly different, and that this will have an impact on correlation estimators necessary for computing functional connectivity graphs as well as topological features extracted from the graphs. We propose to find a stable subspace by using a discriminative projection that does not only minimise site differences, but also preserves discriminative class information. We compare our method with the 'regressing-out' approach in a cross-validation setting and show that regressing out can yield very poor results.