Colon Polyp Segmentation Model Based on Mixed-domain Attention ResNeSt
Objective To address the challenges posed by polyps of varied sizes,unclear boundaries,lighting effects,and their relatively small proportions in images that result in lower segmentation accuracy,an improved U-shaped network structure,BMR-Net,was proposed.Methods The model adopted an encoder-decoder architecture.The encoder partially utilized ResNeSt for feature extraction,enhancing the feature extraction performance with only a slight increase in computational cost.Between the encoder and the decoder,a boundary prediction generation module(BPGM)was designed to aggregate high-level features and incorporate a modified spatial pyramid pooling module,in which an attention mechanism was introduced.This promoted multi-scale information fusion,obtaining a more accurate global feature map representation.For unclear edge areas,a reverse attention module was applied to remove previously predicted areas and correct the boundary information.Results Tests were conducted on the CVC-ClinicDB,Kvasir-SEG,CVC-ColonDB,ETIS-Larib,and EndoScene datasets,with mDice values reaching 0.930,0.903,0.743,0.712,and 0.874,respectively.Conclusion This method outperforms other advanced methods in terms of segmentation performance and generalization ability.Furthermore,it can segment small-sized polyps more precisely and completely,providing early prognosis information for patients with colon polyps.
万方数据
