SegNext框架下融合边缘特征的息肉分割方法

A Polyp Segmentation Method Fusing Edge Features under the SegNext Framework

摘要: 结直肠息肉的精准分割是结直肠癌早期诊断与治疗的关键技术环节，现有融合卷积神经网络(Convolutional Neural Network, CNN)与视觉Transformer (Vision Transformer, ViT)的双分支架构，在息肉医学图像分割中仍面临特征融合不充分、边缘细节捕捉不足及语义表征鲁棒性欠佳等问题。针对上述问题，本文提出一种基于SegNext框架的多模态边缘感知提示息肉分割方法。该方法以多模态边缘感知适配器(Multi-modal Edge-Aware Adapter, MEAA)为核心，协同提取CNN局部纹理特征、小波高频分量(Wavelet High-Frequency Component, WHFC)边缘特征及ViT全局语义特征，通过通道–空间双注意力机制实现多模态特征的自适应融合与增强，并引入密集图视觉提示策略以弥补细节丢失。在Kvasir-SEG、CVC-ClinicDB等5个公开息肉分割数据集上的验证实验表明，所提框架在0次、1次、2次交互提示设置下，平均Dice系数分别达到0.854、0.910、0.935，平均交并比(Intersection over Union, IoU)分别达到0.781、0.858、0.895，分割性能显著优于现有主流方法，具备临床实时应用潜力，为结直肠息肉计算机辅助诊断提供了高效精准的技术方案。

Abstract: Accurate segmentation of colorectal polyps is a key technical step in the early diagnosis and treatment of colorectal cancer. Existing dual-branch architectures that combine Convolutional Neural Networks (CNN) and Vision Transformers (ViT) still suffer from insufficient feature fusion, inadequate capture of edge details, and poor robustness of semantic representation in polyp medical image segmentation. To address these issues, this paper proposes a multi-modal edge-aware prompt polyp segmentation method based on the SegNext framework. The method is centered on a Multi-modal Edge-Aware Adapter (MEAA), which jointly extracts local texture features from CNN, edge features from Wavelet High-Frequency Components (WHFC), and global semantic features from ViT. A channel-spatial dual attention mechanism is adopted to achieve adaptive fusion and enhancement of multi-modal features, and a dense map visual prompt strategy is introduced to compensate for detail loss. Experimental results on five public polyp segmentation datasets, including Kvasir-SEG and CVC-ClinicDB, show that the proposed framework achieves mean Dice coefficients of 0.854, 0.910, and 0.935, and mean Intersection over Union (IoU) scores of 0.781, 0.858, and 0.895 under 0, 1, and 2 interactive prompt settings, respectively. The segmentation performance is significantly superior to existing mainstream methods, demonstrating potential for real-time clinical application and providing an efficient and accurate technical solution for computer-aided diagnosis of colorectal polyps.