Separable Residual Attention U-Net (Sraunet) and Stacked Multi-Scale Correlation Denoising Autoencoder (Smcdae) for Epileptic Seizure Detection

This is the authors' abstract. We don't add key points for this paper.

EEG -electroencephalogram is the most common way to find seizures since it monitors the brain's electrical activity.Interictal EEG in individuals with epilepsy may exhibit asymmetric spikes or sharp waves, signifying the presence of epileptic activity.Recent methods are clearly time-consuming and tiresome, making proper diagnosis crucial to mitigate the risk of subsequent seizures and seizure-related problems.Because of the residual block and SCBA -Separable Convolutional Block Attention, the residual attention block is added to the U-Net architecture.This paper utilizes the IEWT -Improved Empirical Wavelet Transform on EEG recordings for signal pre-processing.The method works by splitting the spectral domain into flexible bounds, which makes it possible to accurately find the important areas in the Fourier spectrum that together make up the signal's underlying trend component.In a few specific frequency bands, the Separable Residual Attention U-Net -SRAUNet model figure out the properties of EEG signals.The residual attention block was added to the U-Net architecture because of the residual block and SCBA.Stacked Multi-scale Correlation Denoising Autoencoder (SMCDAE), the training process is executed by an encoder and a decoder is introduced for seizure detection.The layers are stacked sequentially, second layers' input is the output of the first layer and so on.This step-by-step learning process helps the system capture strong, noise removed patterns which are closely related to seizure activity.The experimental assessment utilizes the CHB-MIT scalp EEG dataset, a renowned collection produced through collaboration between Boston Children's Hospital and the Massachusetts Institute of Technology, this proposed method has obtained highest recall/sensitivity, F-measure, accuracy and precision than other methods.