A novel medical image segmentation approach by using multi-branch segmentation network based on local and global information synchronous learning

Abstract

In recent years, there have been several solutions to medical image segmentation, such as U-shaped structure, transformer-based network, and multi-scale feature learning method. However, their network parameters and real-time performance are often neglected and cannot segment boundary regions well. The main reason is that such networks have deep encoders, a large number of channels, and excessive attention to local information rather than global information, which is crucial to the accuracy of image segmentation. Therefore, we propose a novel multi-branch medical image segmentation network MBSNet. We first design two branches using a parallel residual mixer (PRM) module and dilate convolution block to capture the local and global information of the image. At the same time, a SE-Block and a new spatial attention module enhance the output features. Considering the different output features of the two branches, we adopt a cross-fusion method to effectively combine and complement the features between different layers. MBSNet was tested on five datasets ISIC2018, Kvasir, BUSI, COVID-19, and LGG. The combined results show that MBSNet is lighter, faster, and more accurate. Specifically, for a $$320 \times 320$$ 320 × 320 input, MBSNet’s FLOPs is 10.68G, with an F1-Score of $$85.29\%$$ 85.29 % on the Kvasir test dataset, well above $$78.73\%$$ 78.73 % for UNet++ with FLOPs of 216.55G. We also use the multi-criteria decision making method TOPSIS based on F1-Score, IOU and Geometric-Mean (G-mean) for overall analysis. The proposed MBSNet model performs better than other competitive methods. Code is available at https://github.com/YuLionel/MBSNet.