Multi-Scale Feature Integrated Double Aggregation Network for Enhanced Segmentation

Over 90% of colorectal cancers gradually develop from colorectal polyps. Accurate polyp segmentation offers crucial information for the early detection of colorectal cancer. Nevertheless, the existing research on polyp segmentation in colonoscopy images encounters two major difficulties: the aggregation of polyp boundaries is challenging and the segmentation of small objects is difficult. To handle these problems and strike a fine balance between the quality of image segmentation and network parameters, this paper presents a novel polyp segmentation framework to address these two challenges. Specifically, a transformer encoder is employed, which is capable of learning a more potent and robust representation. Two standard modules are introduced, including the boundary aggregation module (BA) and the Deep secondary fusion module (GLRA). Among them, BA is utilized to aggregate the boundary features of low-level features and the semantic information of high-level features to better preserve boundary details and locate recalibrated targets. GLRA is adopted to collect the semantic and location information of polyps from high-level features, which possesses the ability to aggregate and represent global and local spatial features, which is beneficial for locating large and small targets respectively. Our proposed MFDA has carried out extensive experimental studies on Kvasir, ColonDB, and ETIS datasets. Experiments indicate that our network architecture outperforms the majority of the state-of-the-art methods under different evaluation metrics.