Label-Decoupled Medical Image Segmentation With Spatial-Channel Graph Convolution and Dual Attention Enhancement

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning excels in medical image segmentation but struggles with global context and edge details.
• Existing methods often fail to integrate long-range spatial information and channel correlations.
• Medical images frequently present challenges like blurred target boundaries.

Purpose of the Study:

• To propose a novel medical image segmentation framework, LADENet.
• To address limitations in capturing global information and topological correlations.
• To improve segmentation accuracy, especially for blurred edges.

Main Methods:

• Developed LADENet, incorporating spatial-channel graph convolution for global and topological feature extraction.
• Implemented a label-decoupled strategy using distance transformation for body and edge segmentation.
• Integrated a dual attention enhancement mechanism with dedicated body and edge attention blocks.
• Introduced a feature interactor for enhanced information exchange between segmentation branches.

Main Results:

• LADENet demonstrated superior performance on benchmark medical image segmentation datasets.
• The proposed methods effectively captured global context and improved edge segmentation.
• Experimental results confirmed the framework's advantage over state-of-the-art approaches.

Conclusions:

• LADENet offers a significant advancement in medical image segmentation.
• The combination of graph convolution, label decoupling, and dual attention enhances segmentation accuracy.
• The framework shows promise for clinical applications requiring precise segmentation.