Cerebral stroke classification based on fusion model of 3D EmbedConvNext and 3D Bi-LSTM network

Abstract

Acute stroke can be effectively treated within 4.5 h. To help doctors judge the onset time of this disease as soon as possible, a fusion model of 3D EmbedConvNext and 3D Bi-LSTM network was proposed. It uses DWI brain images to distinguish between cases where the stroke onset time is within 4.5 h and beyond. 3D EmbedConvNeXt replaces 2D convolution with 3D convolution based on the original ConvNeXt, and the downsample layer uses the self-attention module. 3D features of EmbedConvNeXt were output to 3D Bi-LSTM for learning. 3D Bi-LSTM is mainly used to obtain the spatial relationship of multiple planes (axial, coronal, and sagittal), to effectively learn the 3D time series information in the depth, length, and width directions of the feature maps. The classification experiments on stroke data sets provided by cooperative hospitals show that our model achieves an accuracy of 0.83.