HGNet: A Hypergraph-Enhanced YOLO Framework for Colorectal Polyp Detection in Colonoscopy Images

Colorectal cancer prevention relies on the timely identification and removal of precancerous polyps during colonoscopy. However, automated polyp detection remains challenging because of diminutive lesions, blurred boundaries, and complex mucosal backgrounds. This study proposes HGNet, a hypergraph-enhanced YOLO framework for colorectal polyp detection. HGNet introduces: (i) an Efficient Multi-scale Contextual Attention (EMCA) module, which cascades Efficient Multi-scale Attention (EMA) and Contextual Anchor Attention (CAA) to enhance lesion representation across different receptive-field ranges; and (ii) a Spatial Hypergraph Convolution (HyperConv) module, which constructs hyperedges to capture high-order spatial correlations and contextual cues, thereby providing spatial relationship modeling for polyp localization under boundary ambiguity. The results show that HGNet achieved strong overall performance across the three public datasets and maintained real-time processing capability in the inference-efficiency test. These findings suggest that incorporating multi-scale contextual attention and hypergraph-based relation modeling into a YOLO detection framework can improve polyp detection performance in colonoscopy images.