MDFF-Net:Multi-Attention Dual-Branch Feature Fusion Network for Polyp Segmentatio

In routine diagnostics, accurately segmenting polyp regions in colonoscopic images is of paramount importance for clinicians, as it significantly contributes to the prevention of colorectal cancer. The challenge arises from the diverse sizes and shapes of polyps, as well as their indistinct boundaries with the mucosa. To address the limitations of convolutional neural networks (CNNs) in focusing primarily on local feature extraction, we propose a novel Multi-Attention Dual-Branch Feature Fusion Network (MDFF-Net). Our approach employs a dual-branch encoder that combines Transformers and CNNs to extract polyp features from both a global and a local perspective. Subsequently, the aforementioned features are subjected to processing by a Multi-Scale Attention Fusion module (MSAF), which employs Parallel Self-Attention (PSA) and Self-Attention (SA) to adaptively highlight pertinent features. For low-level features, a Feature Complementation Module (FCM) filters out noise and captures fine details. Finally, an Aggregation Module (FAM) integrates the enhanced features to yield precise segmentation results. The results of experiments conducted on two publicly available datasets demonstrate that our method exhibits superior performance in terms of both accuracy and generalization ability compared to existing techniques.