Local-global multi-scale attention network for medical image segmentation

Abstract

With the continuous advancement of deep learning technologies, deep learning-based medical image segmentation methods have achieved remarkable results. However, existing segmentation approaches still face several key challenges, including the insufficient extraction of local and global information from images and the inaccurate selection of core features. To address these challenges, this article proposes a novel medical image segmentation architecture-local-global multi-scale attention network (LGMANet). LGMANet introduces an innovative local-global information processing block (LGIPB) to effectively facilitate the deep mining of both local and global information during the downsampling process. In addition, an efficient multi-scale reconstruction attention (EMRA) module is designed to help the model accurately extract core features and multi-scale information while effectively suppressing irrelevant content. Experiments on the ISIC2018, CVC-ClinicDB, BUSI, and GLaS datasets demonstrate that LGMANet achieves IoU scores of 85.28%, 82.67%, 70.07%, and 88.90%, respectively, showcasing its superior segmentation performance.