摘要：

针对视网膜血管结构细微复杂、边界模糊、计算成本高等问题, 提出一种基于多注意力机制的视网膜血管分割模型(GAC-UNet)。首先, 在跳跃连接中嵌入用于提取通道间关系与空间位置信息的CASP(Channel Attention Spatial Pooling)注意力模块, 将其与残差连接相结合, 构成注意力残差单元(ARU), 以优化编解码器之间的特征处理, 突出重要特征; 然后, 在编码器结构中加入用于合理分配注意力的新型图注意力网络(NGAT), 将其与CASP注意力模块相结合以构建GACA综合注意力模块, 从多个角度对血管细节和边缘进行关注; 接着, 在编码器中堆叠多个GACA模块, 实现NGAT模块内部图注意力信息的累积, 以加强模型建模全局信息的能力, 弥补并丰富边缘特征信息; 最后, 在解码器结构相应层级中聚合不同注意力模块所提取的特征信息, 通过上采样操作得到最终的分割结果。在3个视网膜公共数据集DRIVE、CHASE_DB1、STARE上进行实验, 结果表明, GAC-UNet的特异性分别为97.76%、99.16%和98.66%, 准确率分别为96.80%、96.81%和96.34%, 表明GAC-UNet能够很好地识别出细微复杂、边界模糊的血管结构, 且具有较小的模型参数量。

关键词: 视网膜血管分割, CASP注意力模块, 新型图注意力网络, GACA综合注意力模块, 注意力机制

Abstract:

To address the challenges of subtle and complex structures, blurred boundaries, and high computational costs associated with retinal vasculature, this study proposes a retinal vessel segmentation model named GAC-UNet, based on a multi-attention mechanism. First, a Channel Attention Spatial Pooling (CASP) attention module, designed to extract interchannel relationships and spatial position information, is embedded into skip connections. By integrating this module with residual connections, an Attention Residual Unit (ARU) is formed to optimize feature processing between the encoder and decoder, thereby highlighting important features. Subsequently, a New Graph Attention Network (NGAT) is introduced into the encoder architecture for rationally allocating attention. This NGAT is combined with the CASP attention module to construct the GACA integrated attention module, which enables multi-faceted attention to vessel details and edges. Multiple GACA modules are stacked within the encoder to internally accumulate graph attention information within the NGAT modules, thereby enhancing the ability of GAC-UNet to model global information and enrich the edge feature information. Finally, the feature information extracted by the different attention modules is aggregated at the corresponding levels in the decoder architecture and the final segmentation result is obtained using upsampling operations. Experimental evaluations conducted on three public retinal datasets—DRIVE, CHASE_DB1, and STARE—demonstrate that the proposed model achieves specificities of 97.76%, 99.16%, and 98.66%, and accuracies of 96.80%, 96.81%, and 96.34%, respectively. These results indicate that GAC-UNet effectively identifies subtle and complex vessel structures with blurred boundaries while maintaining a relatively small model parameter size.

Key words: retinal blood vessel segmentation, CASP attention module, the New Graph Attention Networks (NGAT), GACA composite attention module, attention mechanism