Multiscale Feature Fusion Booster Network for Segmentation of Colorectal Polyp

Abstract

Addressing the challenges posed by colorectal polyp variability and imaging inconsistencies in endoscopic images, we propose the multiscale feature fusion booster network (MFFB-Net), a novel deep learning (DL) framework for the semantic segmentation of colorectal polyps to aid in early colorectal cancer detection. Unlike prior models, such as the pyramid vision transformer-based cascaded attention decoder (PVT-CASCADE) and the parallel reverse attention network (PraNet), MFFB-Net enhances segmentation accuracy and efficiency through a unique fusion of multiscale feature extraction in both the encoder and decoder stages, coupled with a booster module for refining fine-grained details and a bottleneck module for efficient feature compression. The network leverages multipath feature extraction with skip connections, capturing both local and global contextual information, and is rigorously evaluated on seven benchmark datasets, including Kvasir, CVC-ClinicDB, CVC-ColonDB, ETIS, CVC-300, BKAI-IGH, and EndoCV2020. MFFB-Net achieves state-of-the-art (SOTA) performance, with Dice scores of 94.38%, 91.92%, 91.21%, 80.34%, 82.67%, 76.92%, and 74.29% on CVC-ClinicDB, Kvasir, CVC-300, ETIS, CVC-ColonDB, EndoCV2020, and BKAI-IGH, respectively, outperforming existing models in segmentation accuracy and computational efficiency. MFFB-Net achieves real-time processing speeds of 26 FPS with only 1.41 million parameters, making it well suited for real-world clinical applications. The results underscore the robustness of MFFB-Net, demonstrating its potential for real-time deployment in computer-aided diagnosis systems and setting a new benchmark for automated polyp segmentation.