Introduction Functional MRI has been shown to be capable of decode human mind states ; for example, to classify brain responses to categorical visual stimuli . However, due to fMRI’s high cost and restrictive environment, it is not practical for routine use in brain-computer interface (BCI). In contrast, EEG is more practical for BCI. In this study, we examined the feasibility of using fMRI to assist EEG signal classification by transforming scalp EEG into corresponding source activation patterns. Our approach demonstrated a dramatic improvement in classification accuracy (than usual spatial filtering method ), which may be highly beneficial in BCI applications. Method Three subjects (2M1F, 23-25 y.o.) participated in the present study with scalp EEG (a 64-channel Brain Products system) and anatomical MRI data acquired. Functional MRI was acquired from only one of the subjects separately from the EEG. We presented (duration= 500ms) 4 categories of pictures (face, building, cat and vehicle) to the subjects in a simple visual perception task. In the EEG experiment, these pictures were presented in a random order with a mean ISI of 1150 ms. In the fMRI experiment, pictures of the same category were presented in blocks with random order. SPM8 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8) was used for MRI data analysis and EEG source localization. The locations of scalp electrodes were co-registered with anatomical MRI data using a 3D digitizer. With boundaryelement head models generated from anatomical images, a multiple sparse priors approach (MSP)  was used for single trial EEG source localization in each subject. This is a current density reconstruction process with dipole positions constrained by the activations in the functional MRI data from Sub A (Fig.1). Then, the source signals within the time range of 140-200 ms were averaged for each node in the head model and thus an activation pattern map was generated for each trial. These patterns were feed into a support vector machine (SVM) for brain state classification. To compare with the source localization approach, single trial data were also analyzed with a spatial filtering approach , in which feature enhancement were achieved by projecting scalp EEG onto a linear subspace with the largest spatial variances.