LSSNet: A Method for Colon Polyp Segmentation Based on Local Feature Supplementation and Shallow Feature Supplementation

Accurate polyp segmentation methods are essential for colon polyp screening and colorectal cancer diagnosis. However, polyp segmentation faces the following challenges: (1) Small-sized polyps are easily lost during the identification process. (2) The boundaries separating the polyp from its surroundings are fuzzy. (3) Additional distracting information is introduced during the colonoscopy procedure, resulting in noise in the colonoscopy image and influencing the segmentation outcomes. To cope with these three challenges, a method for colon polyp segmentation based on local feature supplementation and shallow feature supplementation (LSSNet) is proposed by incorporating feature supplementation structures in the encoder-decoder structure. The multiscale feature extraction (MFE) module is designed to extract local features, the interlayer attention fusion (IAF) module is designed to fuse supplementary features with the current layer features, and the semantic gap reduction (SGR) module is designed to reduce the semantic gaps between the layers, which together form the local feature supplementation structure. The shallow feature supplementation (SFS) module is designed to supplement the features in the fuzzy areas. Based on these four modules LSSNet is proposed. LSSNet is evaluated on five datasets: ClinicDB, KvasirSEG, ETIS, ColonDB, and EndoScene. The results show that mDice scores are improved by 1.33%, 0.74%, 2.65%, 1.08%, and 0.62% respectively over the compared state-of-the-art methods. The codes are available at https://github.com/heyeying/LSSNet .