Identifying Epilepsy Based on Deep Learning Using DKI Images

Abstract

Epilepsy is a serious hazard to human health. Minimally invasive surgery is currently an extremely effective treatment to refractory epilepsy. However, it is challenging to localize the lesion for most patients because they are MRI negative. The identification of epileptic foci in local brain region will be helpful to the localization of epileptic foci because we can infer whether there is a lesion from the results of the classification. For the sake of simplicity and the data we collected, only the hippocampus was segmented as a local brain region and classified in this paper. We recruited 59 children with hippocampus epilepsy and 70 age- and sex-matched normal controls, and diffusion kurtosis images (DKI) for all subjects were collected because DKI can understand the pathological changes of local tissues and other regions of epileptic foci at the molecular level. Then, a mask of hippocampus was made to segment the hippocampus of FA, MD, and MK images for all subjects, which are the parameter images of DKI and were used to perform the independent-sample t-test and the classification task. At last, a convolutional neural network (CNN) based on transfer learning technique was developed to extract features of FA, MD, MK, and the fusion of FA and MK, and support vector machine was employed to classify epilepsy and normal control. Finally, the classifier produced 90.8% accuracy for patient vs. normal controls. Experimental results showed that the features extraction based on CNN is very effective, and the high accuracy of classification means that FA and MK are two remarkable features to identify epilepsy, which indicates that DKI images can act as an important biomarker for epilepsy from the point of view of clinical diagnosis.