Abstracts

Rationale: Quantitative neuroimaging analyses have demonstrated characteristic gray and white matter abnormalities in group comparisons of different types of non-lesional partial epilepsy. It is unknown to what degree these type-specific patterns exist in individual patients and if they could be exploited for diagnostic purposes. In this study, a two-level multi-modality imaging Bayesian network approach is proposed that uses information about individual gray matter volume loss and white matter integrity to differentiate between patients suffering from non-lesional temporal lobe epilepsy with (TLE-MTS) or without (TLE-no) mesial-temporal sclerosis and frontal lobe epilepsy (FLE). Methods: 25 controls, 19 TLE-MTS, 22 TLE-no and 14 FLE were studied on a 4T MRI and T1 weighted structural and DTI images acquired. Spatially normalized GM and FA abnormality maps (binary maps with voxels 1 SD below control mean) were calculated for each subject. At the first level, each group's abnormality maps were compared with those from all the other groups using Graphical-Model-based Morphometric Analysis (GAMMA). GAMMA uses a Bayesian network and a Markov random field based contextual clustering method to identify cluster of voxels that provide the maximal distinction between two groups. The result is a label map (binary image of the voxel subset associated with the function variable) a belief map (label map weighted by the confidence in the voxel/function variable association) and a probability distribution. GAMMA also determines each subject's group membership using a regional state inference algorithm. The belief maps can be used to determine the group membership of other subjects who have not been part of the initial comparison. The second level Bayesian network used a priori expert knowledge about gray and white matter abnormalities in these three epilepsy types to combine the information obtained for each individual at the first level and to calculate a probability for each subject a. to be a patient, a patient with normal imaging or a healthy control. b. if a patient, the probability to suffer from TLE or FLE, and c. if TLE, the probability to be TLE-MTS or TLE-no. Results: The two-level Bayesian network classified of 48 the 55 patients. 75% of those who were classified were correctly identified 7 could not be classified because they had structural abnormalities not exceeding those observed in controls. The sensitivity of the two level Bayesian network to distinguish between the three patient groups was 0.84 for TLE-MTS, 0.72 for TLE-no and 0.64 for FLE, the corresponding specificity was 0.87 for TLE-MTS and TLE-no and 0.86 for FLE. Conclusions: The two-level multi-modality Bayesian network approach was able to distinguish between the three epilepsy types with a reasonable high accuracy even though the majority of the images were completely normal on visual inspection. It is likely that this accuracy can be further increased, e.g. by adding additional imaging information, e.g. PET or clinical and electrophysiological information.