Meg-YOLO Powered by Hypergraph: Multiscale Neighborhood Modeling in Medical Image Detection

Object detection in medical images often faces challenges such as small targets, blurred boundaries, and significant scale variations. Existing methods still fall short in integrating multi-level features and capturing cross-region dependencies. To this end, this paper proposes Meg-YOLO, which introduces Hypergraph-based Correlation Adaptive Fusion (HyperCAF) during the feature fusion stage. HyperCAF explicitly models multi-level dependencies by constructing hypergraphs of cross-scale features, and enables global semantic interactions through graph propagation. Additionally, a Feature Allocation Refiner (FAR) is designed within the backbone network to enhance fine-grained feature representations. During the fusion stage, a Feature Refinement Unit (FRMU) and Context-Guided Modulator (CGM) are introduced to facilitate cross-level integration and contextual information supplementation. Experimental results demonstrate that Meg-YOLO yields gains of 1.9%, 0.8%, 2.6%, and 1.3% on Tuberculosis, BCCD, Br35H, and Kvasir-SEG, respectively.