DBMA-Net: A Dual-Branch Multiattention Network for Polyp Segmentation

In the early prevention stage of colorectal cancer (CRC), the utilization of automatic polyp segmentation techniques from colonoscopy images has demonstrated efficacy in mitigating the misdiagnosis rate. Nonetheless, accurate polyp segmentation is always against with various challenges, including the presence of inconsistent size and morphological changes within polyp classes, limited interclass contrast, and high levels of interference. In recent years, much methodologies based on convolutional neural networks (CNNs) have been widely introduced to enhance the precision of polyp segmentation. However, two significant hurdles persist: 1) these methods frequently suffer from an inadequate acquisition of contextual features, causing insufficient feature representation and 2) there is a deficiency in recognizing intricate information, such as precise polyp boundaries. Addressing these issues, this article introduces a novel dual-branch multiattention network, denoted as DBMA-Net. Specifically, proposed DBMA-Net primarily introduces a dual-encoding path that combines CNN and Transformer-based approaches to enrich feature representation. Additionally, an attention-based fusion module (AFM) is incorporated between the dual-encoding path, aimed at optimizing features by supplementing local information with global insights. Subsequently, two distinct attention mechanisms are introduced to enhance features: the attention-based enhancement module (AEM) and the multiview attention module (MAM), to acquire stronger local features. These modules serve to enrich the finer details while extensively exploring and enhancing the lesion region, thereby further elevating segmentation accuracy. Following the above feature optimization, the enhanced feature maps are hierarchically integrated across multiple scales based on the proposed multiscale feature integration module (MFIM) for accurate feature reconstruction. This strategy not only curtails feature loss but also aids in restoring feature resolution. Ultimately, comprehensive experiments, including comparative and ablation studies across various datasets, validate the superior segmentation performance of the proposed network compared to most state-of-the-art (SOTA) models.