Pattern recognition approach to the detection of single-trial event-related functional magnetic resonance images

Functional magnetic resonance imaging (FMRI) is an imaging technique for determining which regions of the brain are activated in response to a stimulus or event. Early FMRI experiment paradigms were based upon those used in positron emission tomography (PET), i.e. employing a block design consisting of extended periods of 'on' against 'off' activations. More recent experiments were based on event-related FMRI, harnessing the fact that very short stimuli trains or single events can generate robust responses. FMRI data suffer from low signal-to-noise ratios, and typical event-related experiment paradigms employ selective averaging over many trials before using statistical methods for determining active brain regions. The paper reports a pattern recognition approach to the detection of single-trial FMRI responses without recourse to averaging and at modest field strengths (1.5T). Linear discriminant analysis (LDA) was applied in conjunction with different feature extraction techniques. Use of the unprocessed data samples as features resulted in single-trial events being classified with an accuracy of 61.0 +/- 9.5% over five subjects. To improve classification accuracy, knowledge of the ideal template haemodynamic response was used in the feature extraction stage. A novel application of parametric modelling yielded an accuracy of 69.8 +/- 6.3%, and a matched filtering approach yielded an accuracy of 71.9 +/- 5.4%. Single-trial detection of event-related FMRI may yield new ways of examining the brain by facilitating new adaptive experiment designs and enabling tight integration with other single-trial electrophysiological methods.