Abstract

Info

Title:: MFMANet: a multi-attention medical image segmentation network fused with multi-scale features

Author(s):: YUAN Jinli; LI Bohua; CHEN Muxuan; JIANG Rending; JUI SHANAZ SHARMIN; GUO Zhitao; School of Electronic Information Engineering, Hebei University of Technology, Tianjin 300401, China

Keywords:: Keywords: medical image segmentation multi-scale information fusion attention mechanism

Abstract:: The research on medical image segmentation is of great significance in advancing efficient and accurate automated image processing techniques. To address the problem of inaccurate segmentation results caused by significant variations in organ tissue shapes and blurred boundaries present in medical images, a novel network named MFMANet is proposed. Built upon a "U"-shaped architecture, the network integrates multi-scale information fusion modules and multi-attention modules. Specifically, multi-scale information modules capture multi-scale information in the shallow layers of the network to bridge the semantic gap between encoder and decoder features, thereby enhancing the networks ability to handle large variations in organ sizes. Regarding the issue of blurred boundaries, multi-attention mechanism utilizes Swin Transformer as the deep encoder-decoder network, employing channel and spatial attention instead of traditional skip connections to achieve finer feature extraction. Experimental results on the ACDC and Synapse public datasets show that the proposed method achieves improvements of 1.51% and 1.29% in Dice similarity coefficient as compared with MTUNet, fully demonstrating its effectiveness in enhancing segmentation network accuracy.