<p>Gastrointestinal polyps are important precursors of gastrointestinal cancer, making their early detection and precise localization during endoscopy clinically critical. Yet reliable automated detection remains challenging because polyps are often small, poorly demarcated and embedded in complex mucosal backgrounds, leading to missed lesions and false positives under variable illumination and texture conditions. Here we present a lightweight detector based on YOLOv11n that integrates three complementary components. First, C3k2-WTConv introduces wavelet-transform convolution into the C3 backbone, enabling the network to jointly capture global morphology and multi-frequency local details for improved recognition of small and boundary-ambiguous polyps. Second, AFGC-Attention models long-range contextual dependencies and adaptively recalibrates channel responses, enhancing polyp-relevant features while suppressing background noise. Third, CA-HSFPN combines coordinate attention with hierarchical feature fusion to align low-level spatial information with high-level semantic representations across scales. Together, these designs strengthen multiscale feature representation while maintaining a compact computational profile with potential for real-time computer-aided detection. The proposed network contains only 1.81 million parameters and requires 5.5 GFLOPs, achieving 625 frames per second on an RTX 4090 GPU under an offline image-level testing protocol. On a gastrointestinal polyp dataset, it achieves 94.6% mAP@0.5, 92.1% precision and 89.9% recall. Supplementary experiments on VisDrone and PASCAL VOC further evaluate non-medical object-detection robustness, but they are not interpreted as evidence of clinical generalization. These results suggest that the proposed framework is a promising lightweight candidate for future AI-assisted endoscopic polyp screening, while prospective video-based validation and hardware-specific deployment testing remain necessary before clinical use.</p>