A feature fusion module based on complementary attention for medical image segmentation

Automated segmentation algorithms are a crucial component of medical image analysis, playing an essential role in assisting professionals to achieve accurate diagnoses. Traditional convolutional neural networks (CNNs) face challenges when dealing with complex and variable lesions: limited by the receptive field of convolutional operators, CNNs often struggle to capture long-range dependencies of complex lesions. The transformer’s outstanding ability to capture long-range dependencies offers a new perspective on addressing these challenges. Inspired by this, our research aims to combine the precise spatial detail extraction capabilities of CNNs with the global semantic understanding abilities of transformers. Unlike traditional fusion methods, we propose a fine-grained feature fusion strategy based on complementary attention, deeply exploring and complementarily fusing the feature representations of the encoder. Moreover, considering that merely relying on feature fusion might overlook critical texture details and key edge features in the segmentation task, we designed a feature enhancement module based on information entropy. This module emphasizes shallow texture features and edge information, enabling the model to more accurately capture and enhance multi-level details of the image, further optimizing segmentation results. Our method was tested on multiple public segmentation datasets of polyps and skin lesions,and performed better than state-of-the-art methods. Extensive qualitative experimental results indicate that our method maintains robust performance even when faced with challenging cases of narrow or blurry-boundary lesions.