Research on Polyp Segmentation via Dynamic Multi-Scale Feature Fusion and Global-Local Semantic Enhancement

Accurate segmentation of colorectal polyps is crucial for the early screening and clinical diagnosis of colorectal cancer. However, the diverse morphology of polyps, significant variations in scale, and unstable quality of endoscopic imaging pose serious challenges for existing algorithms in achieving precise boundary segmentation. To address these issues, this study proposes a novel polyp segmentation algorithm, GDCA-Net, which is developed based on the You Only Look Once version 12 segmentation model (YOLOv12-seg). GDCA-Net introduces several architectural innovations. First, a Gather-and-Distribute (GD) mechanism is incorporated to optimize multi-scale feature fusion, while Alterable Kernel Convolution (AKConv) is integrated to enhance the modeling of complex geometric structures. Second, the Convolution and Attention Fusion Module (CAF) and Context-Mixing dynamic convolution (ContMix) modules are designed to strengthen long-range dependency modeling and multi-scale feature extraction for polyp boundary representation. Finally, a Wise Intersection over Union-based (Wise-IoU) loss function is introduced to accelerate model convergence and improve robustness to low-quality samples. Experiments conducted on the PolypDB, Kvasir-SEG, and CVC-ClinicDB datasets demonstrate the superior performance of GDCA-Net in polyp segmentation tasks. On the most challenging PolypDB dataset, GDCA-Net achieved a mean Average Precision at 50% IoU threshold (mAP50) of 85.9% and an F1-score (F1) of 85.5%, representing improvements of 2.2% and 0.7% over YOLOv12-seg, respectively. Moreover, on the Kvasir-SEG dataset, GDCA-Net achieved a leading F1 score of 94.9%. These results clearly demonstrate that GDCA-Net possesses strong performance and generalization capabilities in handling polyps of varying sizes, shapes, and imaging qualities.