Abstracts

Rationale: EEG-fMRI localizes epileptic foci by detecting hemodynamic changes in the brain that are correlated to epileptic events visible in EEG. However, scalp EEG is insensitive to activity restricted to deep structures and recording the EEG in the scanner is complex and results in major artefacts that are difficult to remove. This study presents a new framework for identifying the BOLD manifestations of epileptic discharges without having to record the EEG.Methods: The method is a two-stage approach. The first is based on the detection of epileptic events for each voxel. Recently, Khalidov et al. (Signal Processing 2011; in press) developed a new wavelet basis called activelets , which is able to model the HRF by using a linear approximation of the balloon model. Activelets form a dictionary in which the hemodynamic response is sparsely represented. This is why a sparse representation algorithm (IEEE Information Theory 2008; 54: 4789-4812) was used to select the activelet coefficients giving the sparsest solution for each voxel. The second stage is to gather voxels according to proximity in time and space of detected activities. This spatio-temporal clustering was performed by graph-cut algorithm (with 4 clusters). One cluster corresponding to the sparsest mean component was selected. 46 runs obtained from 15 epileptic patients were selected. Patients were selected if they showed clear activation in EEG-fMRI. These runs were divided into 3 groups: one of 28 runs with 1 to 5 events per run (adapted to our method due to the sparsity constraint), one of 13 runs with more than 10 events per run and one of 5 runs without any event. The number of events was read by a neurologist on scalp EEG.Results: First, the method detected at least one concordant event in 26/28 and 11/13 runs for the first and second group, respectively. Although the method did not detect all events read on EEG, the region formed by the voxels of the selected cluster was close to the region obtained by EEG-fMRI (table 1). The delay (known as time-to-peak ) between the event read on EEG and the event detected in fMRI was similar to the traditional value (6s) used in the fMRI analyses, but with important inter-patient variability. Then, we showed that the maps obtained from our method were more similar to EEG-fMRI (with 26/28 and 9/13 significant positive correlations) than 2D-TCA method (Hum Brain Mapp 2009; 30: 3393-405) (11/28 and 6/13 significant positive correlations). Finally, for the third group (without event), 3/5 runs showed an activation concordant with the patient s diagnostic. One event detected by our method in fMRI corresponded to an EEG event that was not initially marked by the neurologist but that was a posteriori assessed to be a spike.Conclusions: The method provided similar results to EEG-fMRI analysis, especially for runs with few epileptic events. Studies similar to EEG-fMRI could be possible without recording the EEG and also for patients with infrequent epileptic events, therefore considerably extending the possibility of localizing epileptic foci with fMRI.