Abstract

Info

Title:: Medical image segmentation based on multi-scale convolution and parallel reverse attention

Author(s):: CHEN Mengfei1; WANG Raofen1; WANG Hailing1; LI Peng2; GONG Xiaomei2; 1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China 2. Department of Radiotherapy, Shanghai Pulmonary Hospital Affiliated to Tongji University, Shanghai 200433, China

Keywords:: Keywords: medical image segmentation multi-scale convolution multi-layer perceptron partial decoder reverse attention module

Abstract:: Abstract: A medical image segmentation network (RPR-MLP) based on multi-scale convolution and parallel reverse attention is presented. In the encoder, Res2net modules and tokenized multi-layer perceptron modules are used as the backbone structure to extract multi-scale information and enhance the diversity of semantic features. Meanwhile, the accuracy of semantic information extraction in the decoder is improved through parallel partial decoder. Additionally, reverse attention module re-emphasizes the focus on important regions for further improving the accuracy of segmentation results. The proposed method achieves Dice scores of 0.896 7 and 0.876 2 on the Kvasir and ISIC 2018 public datasets, respectively, demonstrating its effectiveness and generalization ability in medical image segmentation. Furthermore, when applied to the lung tumor CT image dataset (LungCancer dataset) collected in the study, the proposed network has Dice score, IoU and F1 score of 0.727 8, 58.83% and 67.85%, respectively, outperforming baseline network (UNeXt) and common CNN (U-Net, AttU-Net, U-Net++ and PraNet) by 0.030 1-0.057 8、3.16%-4.70%, and 6.72%-18.53%, respectively. The study confirms the effectiveness and generalization ability of RPR-MLP network on different datasets, providing important technical support for lung tumor image segmentation.