基于双聚合与代理注意力的结肠息肉分割方法研究

Research on colon polyp segmentation method based on dual aggregation and agent attention

摘要:

针对结肠息肉区域误识别、边界定位模糊及复杂样本分割困难的问题，本文提出一种双聚合与代理注意力的结肠息肉分割方法。首先通过 Agent-PVT（Agent Pyramid Vision Transformer, AgentPVT）提取上下文信息，利用代理注意力（Agent Attention）在减少计算量的同时保留了全局建模能力。然后设计全局到局部聚合注意力模块（Global-to-Local Aggregation Module, GLAM），捕捉特征图的全局和局部纹理特征，同时设计边界聚合模块（Boundary Aggregation Module, BAM），高效的聚合边界信息和语义信息，准确地定位息肉轮廓。最后在损失函数中引入批量核范数最大化（Batch Nuclear-norm Maximization, BNM）方法，增强模型对复杂样本的判别能力。提出的方法在五个数据集上进行了实验分析，实验结果表明该方法具有良好的息肉分割性能，其中在 Kvasir-SEG 数据集上，mDice 和 mIoU 分别达到了 92.83 和 88.16%。

Abstract:

To solve the problems of misidentification of colon polyp areas, fuzzy boundary localization, and difficulty in segmenting complex samples, this paper proposed a colon polyp segmentation method using dual aggregation and proxy attention. Firstly, contextual information is extracted through Agent-PVT (Agent Pyramid Vision Transformer), where Agent Attention reduced computational complexity while retaining global modeling capabilities. Then designed a Global to Local Aggregation Module (GLAM) to capture the global and local texture features of the feature map, while designing a Boundary Aggregation Module (BAM) to efficiently aggregate boundary and semantic information and accurately locate polyp contours. Finally, the Batch Nuclear-norm Maximization (BNM) method is introduced into the loss function to enhance the model's discriminative ability for complex samples. The proposed method was experimentally analyzed on five datasets, and the experimental results showed that the method has good polyp segmentation performance. Among them, on the Kvasir-SEG dataset, mDice and mIoU reached 92.83% and 88.16%, respectively.