LWAH-Net: Light weight Attention-Driven Hybrid Network for Polyp Segmentation in Endoscopic Images.

Abstract

Polyp segmentation is vital for the early detection and diagnosis of colorectal cancer, challenges such as variability in polyp morphology, low contrast, and imaging artifacts demand advanced segmentation solutions. LWAH-Net is a light-weight, attention-driven hybrid network combining CNN and transformer-based attention modules to effectively capture local and global contextual features. The architecture includes booster encoders for multiscale feature extraction, attention-based bottleneck for attentiondriven global feature modeling, transformer attention-based residual connection and a combined loss function employing Dice, Jaccard, and surface losses to enhance boundary accuracy. With only 0.82 million parameters, LWAH-Net achieved state-of-the-art performance across five datasets. It attains Dice scores ranging from 78.8% (ETIS dataset) to 93.8% (CVC-ClinicDB dataset) and mean Intersection over Union (mIoU) scores ranging from 70.4% to 90.1%, surpassing existing models in accuracy and computational efficiency. The model demonstrates excellent generalization on diverse datasets, highlighting its adaptability for clinical applications in resource-constrained environments. LWAH-Net is a robust and efficient tool that is a new addition for real-time diagnostic systems for polyp segmentation. https://github.com/manansandila/LWAH-Net.