A Cascaded Semantic Enhancement Network Based on Attention Mechanism for Blurred Small Polyp Segmentation

In some polyp feature extraction algorithms based on deep learning, there is a problem of polyp semantic information degradation caused by frequent operations of feature down-sampling, which reduces the accuracy of polyp segmentation. To solve the problem of polyp semantic information degradation, a Cascaded Semantic Enhancement Network (CSENet) based on attention mechanism is proposed. There are two main parts in CSENet, Cascaded Partial Decoder (CPD) and Polyp Semantic Enhancement (PSE) module. The CPD is used to aggregate multiscale features of polyps. The PSE module is constructed based on channel attention and spatial attention to enhance the degraded semantic information of polyps. The PSE module improves CSENet's ability to segment polyps in both cases of small polyp targets and blurred polyp edges, thus improving CSENet's polyp segmentation accuracy. Experiment results show that the CSENet has best performance compare with five methods (U-Net, UNet++, SFA, PraNet and SANet), under four benchmark polyp segmentation datasets (the Kvasir dataset, the CVC-ClinicDpB dataset, the CVC-ColonDB dataset and the CVC-T dataset). In particular, compared with the SANet, CSENet improves mIoU and Fβw by 2.3% and 1.6% on the CVC-ClinicDB dataset.