VSR-Net: Vessel-Like Structure Rehabilitation Network With Graph Clustering

Area of Science:

• Medical image analysis
• Computational biology
• Biomedical engineering

Background:

• Accurate segmentation of vessel-like structures (e.g., blood vessels, nerve fibers) is crucial for disease diagnosis.
• Current deep learning methods struggle with rehabilitating ruptures and exhibit overconfidence in segmentation results.

Purpose of the Study:

• To address limitations in current vessel-like structure segmentation methods.
• To propose a novel network for rehabilitating subsection ruptures and improving model calibration.

Main Methods:

• Developed a Vessel-like Structure Rehabilitation Network (VSR-Net).
• Introduced a Curvilinear Clustering Module (CCM) to form subsection rupture clusters.
• Utilized a Curvilinear Merging Module (CMM) to repair ruptures and refine segmentation.

Main Results:

• VSR-Net significantly outperforms state-of-the-art (SOTA) refinement segmentation methods.
• Achieved lower calibration errors compared to existing approaches.
• Demonstrated more effective rehabilitation of vessel-like structures with smaller morphological differences from ground truth.

Conclusions:

• VSR-Net successfully rehabilitates subsection ruptures in vessel-like structures.
• The proposed method enhances segmentation accuracy and model calibration.
• VSR-Net offers a promising advancement for medical image analysis and disease diagnosis.