Pattern Recognition in NeuroImaging: What can machine learning classifiers bring to the analysis of functional brain imaging?

The study of the brain development and functioning raises many question that are tracked using neuroimaging techniques such as positron emission tomography or (functional) magnetic resonance imaging. During the last decades, various techniques have been developed to analyse neuroimaging data. These techniques brought valuable insight on neuroscientific questions, but encounter limitations which make them unsuitable to tackle more complex problems. More recently, machine learning based models, coming from the field of pattern recognition, have been promisingly applied to neuroimaging data.
In this work, the assets and limitations of machine learning based models were investigated and compared to previously developed techniques. To this end, two applications involving challenging datasets were defined
and the results from widespread methods were compared to the results obtained using machine learning based modelling.
More specifically, the first application addressed a research question: Is it possible to detect and characterize mnemonic traces? The fMRI experiment comprised a learning and a control tasks, both flanked by
rest sessions. From previous studies, patterns of brain activity generated during the learning task should be spontaneously repeated during the following rest session, while no difference should be observed between
the pre- and post-task rest session in the control condition. Using univariate and multivariate feature selection steps before a Gaussian Processes classification, mnemonic traces could be detected and their
spatio-temporal evolution characterized. On the contrary, an analysis of the rest sessions based on the detection of independent networks did not provide any results supporting the theory of memory consolidation.
The second application tackled a clinical issue: Can a pattern of brain activation characteristic to idiopathic Parkinson’s disease be detected and localized? The dataset considered to address this question comprised
the fMRI images of aged healthy subjects and Parkinsonian patients while they were performing a task of mental imagery of gait at three different paces. The signal comprised in a priori selected regions of interest allowed for the support vector machines classification of healthy and diseased volunteers with an accuracy of 86%. To localize the discriminating pattern, a methodology based on the weight in labelled regions (e.g. from the anatomical automatic labelling or Brodmann atlases) was developed, which enabled the comparison between univariate and multivariate results and showed a nice overlap between them. Furthermore, models could then be compared quantitatively in terms of pattern localization, using a specifically defined measure of distance. This measure could then be used to compare the patterns generated from different folds of a same model, from different feature sets, or from different modelling techniques.
The present study concluded that machine learning models can clearly and fruitfully complement other analysis techniques to tackle challenging questions in neuroscience. On the other hand, more work is needed in
order to render the methodology fully accessible to the neuroscientific community.