Algorithm for polyp segmentation with local encoding and decoding fusion and multi-scale attention

Abstract

In recent years, the application of medical image semantic segmentation tasks in medical diagnosis and treatment planning has received widespread attention from the research community. The High-Resolution Network (HRNet) has good adaptability to high-resolution and high-scale medical images. In this paper, a novel high-resolution serial feature fusion encoding and decoding structure is proposed, and a CBAM attention mechanism is fused to construct a module that can jointly focus on spatial, channel, and multi-scale hierarchical information, which can improve the feature representation ability of the model and effectively reduce parameter complexity. We use the HRNet architecture to construct our model. Experimental results show that our method achieves MIoU coefficient of 98.44% on the Kvasir-SEG dataset, which is 1.43 percentage points higher than the original HR-Net model, validating the effectiveness and reliability of our method.