SCANet: Dual Attention Network for Alzheimer's Disease Diagnosis Based on Gated Residual and Spatial Asymmetry Mechanisms

Convolutional neural networks, combined with attention mechanisms, can effectively extract global and local features from structural magnetic resonance images to aid in the diagnosis of Alzheimer's disease (AD). However, the attention mechanism still presents challenges in AD diagnosis. First, channel attention degradation and feature correction processes lead to the loss of important features. Second, capturing directional information during spatial attention correction is difficult. Therefore, this study proposes the Spatial and Channel attention Network (SCANet) based on Gated Residual Channel Attention (GRCA) and Spatial Asymmetric Attention (SAS) blocks. The GRCA block is based on the normalized attention jumping mechanism, which reduces attentional decay that occurs when the network is too deep, and the block can be calibrated to further enhance important features and suppress non-important features. The SAS block uses asymmetric convolution to model the horizontal and vertical direction-related attention information generated by different pooling methods. It adopts a cross-fertilization strategy to fuse the attention direction information generated by different asymmetric convolutions with different pooling methods, obtaining an attention vector with direction information. The SCANet model was validated through various experiments. The results of five-fold cross-validation showed that SCANet has an average classification accuracy of 97.84% for AD and normal control, which is better than that of the comparison models.