Localization of the Epileptic Focus using Multi-scale Deep Learning

PET imaging is considered as one of safest and most significant methods for localizing the epileptic focus in the field of brain neuroscience. With the advent of the era of Artificial Intelligence, it becomes one of hot research topic that neurologist and radiologist are assisted to locate the epileptic focus by using Computer-aided Diagnosis method. In this study, we propose a novel method for localization of epileptic focus in PET imaging using Multi-scale Deep Learning method. Firstly, three different Mask Region-based Convolutional Neural Network models were used to extract the candidate of epileptic focus. The three models were produced by transferring learning towards Mask R-CNN utilizing training images that made up of PET scanning from three different scales. Each training image set contained 375 brain PET axial slices of epilepsy. Then three Deep Learning models were combined using the ensemble learning to reduce false positive results and to localize the epileptic focus. 48 PET axial slices were utilized as test set for this research. The sensitivity and specificity of this new method were 0.80 and 0.8125 respectively. The experimental results show the effectiveness of this method in locating epileptic focus.