融合边界注意力的特征挖掘息肉小目标网络

A small polyp objects network integrating boundary attention features

从结肠图像中分割息肉小目标病变区域对于预防结直肠癌至关重要,它可以为结直肠癌的诊断提供有价值的信息。然而目前现有的方法存在2个局限性:一是不能稳健捕获全局上下文信息,二是未能充分挖掘细粒度细节特征信息。因此,提出融合边界注意力的特征挖掘息肉小目标网络(transformer feature boundary network,TFB-Net)。该网络主要包括3个核心模块:首先,采用Transformer辅助编码器建立长程依赖关系,补充全局信息;其次,设计特征挖掘模块进一步细化特征,学习到更好的特征;最后,使用边界反转注意力模块加强对边界语义空间的关注,提高区域辨别能力。在5个息肉小目标数据集上进行广泛实验,实验结果表明TFBNet具有优越的分割性能。

Segmentation of small target lesion areas,such as polyps in colon images,is essential for the prevention and diagnosis of colorectal cancer.However,existing methods face two main limitations:either the global context information cannot be captured robustly or the fine-grained detail information cannot be fully mined.To address these issues,this study proposes TFB-Net,a feature mining network for small target polyps that integrates boundary attention.The network consists of three core modules:First,a Transformer is used to establish long-term dependencies and supplement global information.Second,the feature mining module is designed to further optimize and enhance the learned features.Finally,the boundary inversion attention module strengthens attention to the boundary semantic space,which consequently improves regional discrimination.Extensive experiments were conducted on five small polyp target datasets,and the results show that TFB-Net achieves superior segmentation performance.