A Pixel-based Confidence Weighted Transfusion model for Colonoscopy and Colon Capsule Endoscopy image-based Polyp Segmentation

Abstract Polyps one of the most frequently occurring lesions in the human gastrointestinal system and play a critical role in colorectal cancer development, often serving as early indicators or precursors of malignant transformation. Timely and early detection of polyps is critical for lowering the risk of colorectal cancer and increasing survival. Polyp segmentation is crucial in improving computer-aided detection and diagnosis systems, which may support early intervention. CNN-based segmentation models often struggle to capture global contextual information, leading to suboptimal performance. These limitations are further amplified in endoscopic imaging scenarios, where variable resolution in colon capsule endoscopies (CCE) and conventional colonoscopy images—introduces inconsistencies in feature representation and contextual awareness. Therefore, we develop a comprehensive TransFuseNet model that captures local and global receptive fields at the encoder and decoder levels. The pixel-based confidence-weighted fusion combines features from two models, which facilitates knowledge and domain shift features that allows segmentation of low- and high-resolution images. We validated the model's performance on seven diverse optical colonoscopies, CCE, and synthetic datasets. Analysis reveals that our developed model has yielded an intersection over the union of 84.67\%, 89.05\%, 88.91\%, 89.46\%, 89.58\%, 82.07\%, and 86.41\% on the Kvasir, CVC-300, CVC-ColonDB, CVC-ClinicDB, ETIS, CCE CareForColon2015 dataset, and Synthetic datasets, respectively. The external validation also confirms the robustness and effectiveness of our developed model, providing a new pathway for polyp segmentation using optical colonoscopy and CCE images with explainable characteristics and polyp pixel measurement. Thus, our proposed TransFuseNet effectively handles both low- and high-resolution endoscopic images by fusing local and global contextual features. This supports more reliable polyp segmentation across diverse imaging conditions, enhancing diagnostic accuracy in clinical practice.