Late Mean Fusion Towards Efficient Polyps Segmentation

Accurate and efficient polyp segmentation in colonoscopy images is crucial for early colorectal cancer detection. While traditional methods often rely on one-stage early fusion techniques, this paper explores the advantages of late fusion for polyp segmentation. We propose a novel framework combining advanced deep learning models (DilatedSegNet, TransFuse, and CFANet) to leverage their strengths later in the segmentation process. This approach allows for independent processing of different feature sets before integrating them, improving segmentation accuracy. Evaluations on the Kvasir-SEG and CVC-ClinicDB datasets demonstrate that our late fusion method surpasses existing state-of-the-art techniques, achieving an average score of 0.9609, representing a 2.52% improvement over the second-best option. These findings highlight the potential of late fusion to significantly enhance polyp segmentation for real-time clinical applications in colorectal cancer screening.