MFFNet:基于多尺度特征有效融合的息肉分割网络  

MFFNet: Multi-Scale Feature Effective Fusion Network for Polyp Segmentation

作　　者：杨虹霞 祝汉灿[1] 

机构地区：[1]绍兴文理学院数理信息学院,浙江 绍兴

出　　处：《理论数学》2025年第1期198-210,共13页Pure Mathematics

摘　　要：息肉的准确分割对结直肠癌的治疗具有重要意义。虽然现有的方法已经取得了良好的分割效果,但仍然存在一些挑战。为此,我们提出了一个新的多尺度特征有效融合网络(MFFNet),用于精确分割息肉。具体来说,考虑到息肉的尺寸差异,我们使用改进的Pvt-v2作为编码器(TC编码器,TC encoder),提取丰富的多尺度特征。然后,应用通道–空间模块(Channel Spatial Module, CSM)来抑制背景信息,防止信息的冗余。为了使多尺度特征进行有效融合,我们提出了融合注意力模块(Fusion Attention Block, FAB),该模块充分学习多层次特征之间的上下文相关性,以进一步精确定位息肉区域。在5个公共数据集上的实验表明,我们的MFFNet比其他方法具有更好的学习和泛化能力。Accurate segmentation of polyps is important in the management of colorectal cancer. Although existing methods have achieved good segmentation results, there are still some challenges. To this end, we propose a new Multi-Scale Feature Effective Fusion Network (MFFNet) for accurate polyp segmentation. Specifically, considering the size difference of polyps, we use the improved Pvt-v2 as an encoder (TC encoder) to extract rich multi-scale features. Then, the Channel-Spatial Module (CSM) is applied to minimize background interference and prevent the redundancy of information. To enable effective fusion of multi-scale features, we propose the Fusion Attention Block (FAB), which fully learns the contextual correlations between multi-level features to further pinpoint the polyp region. Experiments on five public datasets show that our MFFNet has better learning and generalization capabilities than other methods.

关 键 词：息肉分割 多尺度特征 特征融合 注意力机制 

分 类 号：G63[文化科学—教育学]