SCAGC-UNet: Graph Convolutional Network with Spatial and Channel Attention for Medical Image Segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning for healthcare

Background:

• Existing medical image segmentation methods, CNNs and Transformers, face limitations in capturing both local details and global context.
• A trade-off exists between local receptive fields in CNNs and semantic dilution in Transformers for global context modeling.
• Accurate medical image segmentation is crucial for effective clinical diagnosis and treatment planning.

Purpose of the Study:

• To introduce SCAGC-UNet, a novel region-aware graph convolutional network designed to overcome the limitations of existing segmentation methods.
• To effectively bridge local detail extraction and global dependency modeling through structured region-level reasoning.
• To enhance the accuracy and robustness of medical image segmentation across various clinical applications.

Main Methods:

• Developed SCAGC-UNet, a region-aware graph convolutional network featuring a dual-layer residual encoder for hierarchical feature extraction.
• Incorporated a Spatial-Channel Graph Convolution (SC-GCN) module at the bottleneck for simultaneous inter-region spatial topology and intra-region channel semantics modeling.
• Enhanced decoder feature refinement using Context-Corrected Modules and Backward-Aided Modules to minimize semantic gaps in skip connections.

Main Results:

• SCAGC-UNet achieved state-of-the-art performance on three public benchmarks: Kvasir-SEG (Dice 92.28%), BUSI (IOU 78.10%), and COVID-19 CT (DSC 82.51%).
• The model significantly outperformed strong CNN-based (CCBANet) and Transformer-based (TransUNet) baselines across all datasets.
• Demonstrated robust segmentation capabilities, particularly for irregular lesion morphologies in COVID-19 CT scans.

Conclusions:

• SCAGC-UNet offers a robust and effective solution for medical image segmentation, overcoming the limitations of prior CNN and Transformer approaches.
• The region-aware graph convolutional network architecture successfully integrates local and global information, leading to improved segmentation accuracy.
• The consistent and superior performance across diverse medical imaging datasets suggests SCAGC-UNet's strong potential for clinical relevance and application.