Accurate Gastric Ulcer and Intestinal Polyp Segmentation Using Attention Network and Multiscale Information Integration

Abstract

The accurate segmentation in endoscopic images is essential in aiding the management of pathology analysis. This paper proposes a composite segmentation algorithm working on the segmentation task in gastric ulcer and intestinal polyp. Technically, this problem suffers from indistinct boundaries and ignorance of contextual information. We address this segmentation task by utilizing a framework that works on different medical images, containing self-attention modules and encoder-decoder structure. Alongside using multi-scale feature fusion methods to obtain multi-level features and changing convolutional kernels to increase the receptive field, two more customized attention modules – position attention module and channel attention module are applied to promote capturing contexts. The attention modules integrate local features using their global information. Image processing methods are also used to optimize image quality and remove noise information. Data augment methods are applied to increase the amount of training data and improve the model's generalization ability. Experimentally, we conducted sufficient ablation experiments, and the proposed network has been validated on two datasets: the CVC-ClinicDB and the Jiading District gastric ulcer dataset, which is rationally pixel-level annotated by medical specialists. The results indicate our proposed network concurrently outperforms other methods in these two types of endoscopic images.