Multi-Scale Group Agent Attention-Based Graph Convolutional Decoding Networks for 2D Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Automated medical image segmentation is vital for disease diagnosis.
• Feature decoding presents a significant challenge in medical image segmentation.

Purpose of the Study:

• To propose a novel feature decoding network, MSGAA-GCDN, for 2D medical image segmentation.
• To enhance local-global feature representation within graph structures.

Main Methods:

• Developed the multi-scale group agent attention-based graph convolutional decoding networks (MSGAA-GCDN).
• Integrated graph convolutional networks (GCN) with a multi-scale group agent attention (MSGAA) mechanism.
• Introduced an attention-based upsampling convolution fusion (AUCF) module for improved encoder-decoder feature fusion.

Main Results:

• MSGAA-GCDN demonstrated superior performance on abdominal multi-organs, cardiac organs, and polyp lesion segmentation tasks.
• The MSGAA mechanism proved to be a lightweight and effective attention architecture.
• The MSGAA-GCDN achieved state-of-the-art results compared to existing methods.

Conclusions:

• The proposed MSGAA-GCDN effectively addresses feature decoding challenges in medical image segmentation.
• MSGAA-GCDN offers a versatile, plug-and-play decoder for various medical imaging applications.
• The study highlights the potential of graph-based approaches with attention mechanisms for enhanced segmentation accuracy.