Intensity normalisation for large-scale fMRI brain decoding

Among the long-term goals of the fairly new area of brain decoding is the exploitation of the results for the creation of advanced brain-computer interfaces, which can potentially establish a solid communication channel with people in vegetative state. Recent attempts for large-scale brain decoding form a both powerful and promising foundation towards that goal, since they aim to extract accurate representations of certain stimuli within the human brain, given a large number of different studies. An inherent problem with across-study brain decoding is that the classification algorithms end up discriminating among studies instead among stimuli. This is due to study-specific nuisance effects, which cannot be removed by standard preprocessing methodologies, and which may cause two volumes representing different stimuli within a single study to be closer to one another than two volumes representing similar stimuli across different studies. Considering that a large number of previous studies suggest that across-subject and across-session decoding works, we have come to believe that the problem of degraded across-study accuracy is introduced by differing stimuli activation values across studies, originating from study-specific and not subject-specific idiosyncrasies. Therefore, the problem of correct stimuli classification across studies is reduced to the one of consistent intensity normalisation across studies, in order to provide persistent representations of stimuli in the brain. In this work, we provide a thorough discussion on the performance of four different intensity normalisation techniques, in order to evaluate their applicability as a pre-processing step for large-scale brain decoding.