Learning geometric and visual features for medical image segmentation with vision GNN.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation is vital for clinical applications.
• Deep learning methods excel but often overlook inter-object relationships.
• Existing grid-based approaches limit understanding of complex anatomical structures.

Purpose of the Study:

• To introduce MedSegViG, a novel model for medical image segmentation.
• To address limitations of grid-based deep learning methods by incorporating graph structures.
• To improve segmentation accuracy and robustness by modeling relationships between segmented objects.

Main Methods:

• Developed MedSegViG, a model featuring a hierarchical Vision GNN (ViG) encoder and a hybrid feature decoder.
• Represented medical images as graphs to capture object relationships.
• Extracted multi-level graph and image features using the ViG encoder.
• Fused features in the decoder to generate the final segmentation map.

Main Results:

• MedSegViG demonstrated superior segmentation accuracy and robustness.
• The model achieved excellent generalizability across diverse datasets and lesion types.
• Extensive experiments on polyp, skin lesion, and retinal vessel datasets validated effectiveness.

Conclusions:

• MedSegViG offers a significant advancement in medical image segmentation.
• Graph-based representation and hierarchical feature extraction improve performance.
• The model shows strong potential for clinical applications requiring precise segmentation.