Multi-feature Fusion for Epileptic Focus Localization Based on Tensor Representation

Epileptic focus localization based on intracranial electroencephalogram (iEEG) signal is a key task before the patient's surgery, which is time-consuming by visual diagnosis of clinical experts and its accuracy directly affects the effectiveness of surgery. Recently, many machine learning methods have been applied to automatic epileptic focus localization. However, most studies have performed feature extraction and classification based on single type of features, leading to less robustness to noises and degraded performance. In this paper, inspired by making a diagnosis from multiple angles in practical situation, we proposed a multi-set feature fusion strategy by using tensor representation technique. As compared to the existing linear fusion strategy, our proposed tensor fusion approach can significantly enhance the expressive power of the model by taking into account the feature interaction information. We integrate tensor fusion strategy into deep convolutional neural network, yielding a new method for automatic epileptic focus localization. Experimental results demonstrate that our proposed model based on multi-set tensor fusion can achieve the best performance among single type features based models and the linear fusion model.