Colorectal Polyp Segmentation Based on Group Convolution and Transformer

Abstract

In clinical setting, colorectal polyps are associated with colorectal cancer, so it is critical to segment polyps from colorectal images. However, because their location, shape, and size are not excatly the same, it is difficult to do so precisely and effectively. To address these challenges, this paper proposes Trans-PraNeXt, a colorectal polyp segmentation model based on group convolution and transformer. The model uses the Res2NeXt network as the backbone network and replaces the convolution in the original network with group convolution; extracts features with SE (Squeeze and Excitation); uses MHSA (The Multi-Head self-attention) for spatial recovery and a parallel partial decoder is used to aggregate some of the high-level features. In addition, mining boundary cues with the RA (Reverse Attention) module and PAA-e (Parallel Axial Attention Encoder) reveals the connection between regions and boundary cues. The qualitative and quantitative results of the six metrics on the five datasets reflect the model’s improved learning ability along with its generalization ability to segment polyps more accurately.