Graph-Based Region and Boundary Aggregation for Biomedical Image Segmentation.

Area of Science:

• Biomedical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Segmentation is crucial in biomedical imaging, but existing methods often focus on either region or boundary information.
• Integrating both region and boundary features comprehensively remains a challenge for accurate segmentation.

Purpose of the Study:

• To develop a novel deep learning framework using graph neural networks (GNNs) for end-to-end biomedical image segmentation.
• To explicitly leverage both region and boundary features by integrating them within a unified GNN architecture.

Main Methods:

• A graph neural network (GNN) framework with multiple graph reasoning modules was developed.
• An Attention Enhancement Module (AEM) extracts discriminative region and boundary node embeddings.
• Multi-level feature node embeddings and iterative message passing enable concurrent reasoning across different feature levels.

Main Results:

• The proposed GNN method achieved state-of-the-art performance in segmenting polyps in colonoscopy images.
• The framework also demonstrated superior results in segmenting the optic disc and optic cup in color fundus images.
• Experiments on challenging datasets validated the effectiveness of integrating region and boundary features.

Conclusions:

• The novel GNN-based framework successfully integrates region and boundary information for enhanced biomedical image segmentation.
• This approach offers a significant advancement over existing methods by concurrently reasoning across multiple feature levels.
• The developed models are made publicly available to facilitate further research and application.