Abstract
Colorectal cancer remains a leading cause of cancer-related mortality worldwide, emphasizing the importance of early detection through accurate polyp identification. However, colonoscopy relies heavily on precise polyp segmentation in endoscopic images, yet this task remains challenging due to morphological variability, low contrast, and imaging artifacts. In this study, we propose HSSAM-Net, a lightweight deep learning framework that integrates a Hyper-Scale Shifted Aggregation Module to capture multi-scale contextual information while preserving fine-grained details, Progressive Reuse Attention mechanism that strengthens feature propagation across the encoder-decoder pathway, and Max-Diagonal Pooling/Unpooling (MaxDP/MaxDUP) a novel dual-branch sampling scheme to improve texture representation, feature alignment to enhance feature aggregation, context learning, and boundary refinement. The proposed model is evaluated on five benchmark datasets (Kvasir, CVC-ClinicDB, ETIS, CVC-300, EndoCV2020). Experimental results show that HSSAM-Net consistently outperforms state-of-the-art methods across benchmark datasets, HSSAM-Net consistently achieves state-of-the-art accuracy (Dice: 0.949-0.952, mIoU: 0.924-0.930), while maintaining real-time efficiency at 24.1 FPS with only 0.9 M parameters. Furthermore, an analysis of trainable parameters and inference speed confirms its suitability for real-time clinical applications. Our findings demonstrate that HSSAM-Net achieves a favorable trade-off between accuracy and efficiency, advancing the development of practical and reliable computer-aided colonoscopy systems.