基于通道注意力与图注意力网络的脑部影像分类研究OA

To address brain image classification for distinguishing patients with Alzheimer's Disease(AD)from Cognitively Normal(CN)individuals,this study proposes a model based on channel attention and graph attention networks.First,the original Diffusion Tensor Imaging(DTI)data were preprocessed and tensor fitted to obtain a three-dimensional Fractional Anisotropy(FA)map.Then,it was sliced layer by layer along the axial direction,and all slices were uniformly resampled to the same size using bilinear interpolation.A channel attention convolution module was designed to extract spatial detail features through parallel convolutions.At the same time,the channel weights were dynamically recalibrated using the inter-channel attention mechanism with adaptive kernel length,ef-fectively learning and compressing local texture features.The feature vector was obtained through mapping in the fully connected layer and embedded as a node in the subsequent graph.Next,the adjacency matrix was constructed according to the spatial connectivity of the slices.The slices were regarded as nodes in the graph.The node feature vectors and the adjacency matrix were combined into graph data,which were input into three stacked graph attention convolution layers.The information of adjacent and distant slices was aggregated layer by layer through the learnable attention coefficient to achieve the fusion of the global structure across slices.Finally,the classification results were output through global average pooling and fully connected layers.Experimental results show that the proposed meth-od not only outperforms traditional image classification models in accuracy and stability,but also verifies the advan-tages of the local-global attention architecture.Furthermore,it offers a novel approach for converting slice-based data into graph signals.