Skip to Main Content
It is a standard approach to consider that images encode some information such as face expression or biomarkers in medical images; decoding this information is particularly challenging in the case of medical imaging, because the whole image domain has to be considered a priori to avoid biasing image-based prediction and image interpretation. Feature selection is thus needed, but is often performed using mass-univariate procedures, that handle neither the spatial structure of the images, nor the multivariate nature of the signal. Here we propose a solution that computes a reduced set of high-level features which compress the image information while retaining its informative parts: first, we introduce a hierarchical clustering of the research domain that incorporates spatial connectivity constraints and reduces the complexity of the possible spatial configurations to a single tree of nested regions. Then we prune the tree in order to produce a parcellation (division of the image domain) such that parcel-based signal averages optimally predict the target information. We show the power of this approach with respect to reference techniques on simulated data and apply it to enhance the prediction of the subject's behaviour during functional Magnetic Resonance Imaging (fMRI) scanning sessions. Besides its superior performance, the method provides an interpretable weighting of the regions involved in the regression or classification task.