SWMA-UNet: Multi-Path Attention Network for Improved Medical Image Segmentation

Area of Science:

• Deep learning
• Computer vision
• Medical image analysis

Background:

• Convolutional Neural Networks (CNNs) have limitations in capturing long-range dependencies and global context in medical images.
• Transformers offer advantages in understanding global information but can be computationally intensive.
• Integrating Transformers and CNNs is a promising direction, but serial approaches hinder simultaneous processing of local and global features.

Purpose of the Study:

• To propose a novel parallel multi-path attention architecture, SWMA-UNET, for enhanced medical image segmentation.
• To overcome the limitations of serial integration of Transformers and CNNs.
• To improve the accuracy of medical image segmentation by effectively capturing both local details and global context.

Main Methods:

• Developed a parallel multi-path attention architecture (SWMA-UNET) integrating Transformers and CNNs.
• Employed parallel strategies for deep feature mining.
• Simultaneously processed local details and global context information.

Main Results:

• SWMA-UNET demonstrated superior performance compared to existing methods.
• The proposed architecture achieved state-of-the-art results on multiple benchmark datasets: Synapse, ACDC, ISIC 2018, and MoNuSeg.
• The parallel approach effectively enhanced the accuracy of medical image segmentation.

Conclusions:

• The SWMA-UNET architecture offers an effective solution for medical image segmentation.
• Parallel integration of Transformers and CNNs significantly improves feature representation.
• The proposed method sets a new benchmark for accuracy in medical image segmentation tasks.