Combining multi-parametric MR images for the detection of epileptogenic lesions

Detecting epileptogenic brain lesions in patients with intractable epilepsy is important as patients may then be eligible for surgery. We compared two methods of anomaly detection via combining multi-parametric MR images. The early fusion approach consists in building a single global model; in the late fusion approach three local models are constructed based each on a single MR sequence, and their decisions combined. We proposed an original approach for decision combination based on majority voting and Stouffer's method coupled with a univariate statistical analysis. Cross validation with depth EEG electrodes findings in three patients including patients with normal MR on visual examination showed that the late fusion approach achieves the best performance.