Morphology Edge Attention Network and Optimal Geometric Matching Connection model for vascular segmentation

Yuntao Zhu, Yuxuan Qiao, Xiaoping YangarXiv 2023 · DOI · arXiv

Abstract

There are many unsolved problems in vascular image segmentation, including vascular structural connectivity, scarce branches and missing small vessels. Obtaining vessels that preserve their correct topological structures is currently a crucial research issue, as it provides an overall view of one vascular system. In order to preserve the topology and accuracy of vessel segmentation, we proposed a novel Morphology Edge Attention Network (MEA-Net) for the segmentation of vessel-like structures, and an Optimal Geometric Matching Connection (OGMC) model to connect the broken vessel segments. The MEA-Net has an edge attention module that improves the segmentation of edges and small objects by morphology operation extracting boundary voxels on multi-scale. The OGMC model uses the concept of curve touching from differential geometry to filter out fragmented vessel endpoints, and then employs minimal surfaces to determine the optimal connection order between blood vessels. Finally, we calculate the geodesic to repair missing vessels under a given Riemannian metric. Our method achieves superior or competitive results compared to state-of-the-art methods on four datasets of 3D vascular segmentation tasks, both effectively reducing vessel broken and increasing vessel branch richness, yielding blood vessels with a more precise topological structure.

Introduction

Vascular segmentation—the process of digitally isolating blood vessels from medical scans—is a cornerstone of modern diagnostics and surgical planning. However, it is notoriously difficult. Blood vessels are often tiny, have low contrast against surrounding tissue, and follow complex, winding paths.

Even advanced deep learning models frequently produce "fractured" results, where a single continuous vessel appears as several disconnected fragments. These breaks ruin the topological structure of the vascular tree, making it impossible to perform accurate blood flow simulations or comprehensive disease analysis. Existing solutions either try to fix this during training (which is computationally heavy) or use simple post-processing that ignores the underlying geometric "logic" of how vessels grow and flow.

Methodology

The proposed framework consists of a two-stage approach: an initial segmentation stage using a specialized neural network, followed by a geometry-driven connection model to repair topological fractures.

The proposed two-stage framework: MEA-Net for initial segmentation followed by the OGMC model for topological repair.

1. MEA-Net (Morphology Edge Attention Network)

Built on a U-Net architecture, MEA-Net introduces a novel Edge Attention Module to address the difficulty of segmenting fine boundaries and small vessels.

  • Differentiable Morphology: Unlike standard networks that rely solely on pixel intensity, this module uses differentiable erosion and dilation operators to explicitly calculate vessel boundaries at multiple scales. This allows the network to learn edge-detection strategies that were previously only available in traditional computer vision.
  • Multi-scale Loss: The network calculates loss at every layer of the decoder, forcing the model to pay attention to both large arteries and tiny capillaries simultaneously.
Detailed view of the Edge Attention Module and the process of boundary extraction using morphology operations.

2. OGMC (Optimal Geometric Matching Connection)

Once the initial segmentation is complete, the OGMC model steps in to fix breaks by treating the problem as a geometry puzzle:

  • Curve Touching: Using Taylor expansions and Frenet frames from differential geometry, the model analyzes the endpoints of broken vessels to determine if two curves would meet smoothly if continued.
  • Minimal Surfaces: To decide which fragments belong together, the model calculates the area of a "minimal surface" that would bridge the gap between two curves. A smaller surface area indicates a more natural, higher-priority connection.
  • Geodesic Repair: Finally, it uses a Riemannian metric and the "fast marching method" to draw the most efficient path (a geodesic) between the points, effectively "regrowing" the missing vessel segment.

Key Results

The proposed method produces more mathematically accurate vascular trees by leveraging the physical properties of vessel structures:

  • Topological Accuracy: The model improved absolute error rates for Betti Numbers (a mathematical measure of connectivity), tree length, and branch richness by over 10% compared to previous state-of-the-art (SOTA) methods.
  • Robustness: It achieved competitive scores on standard metrics like the Dice Similarity Coefficient (DSC) while significantly reducing the number of vessel fractures.
  • Qualitative Improvement: Visual results show much richer, more connected vascular trees, especially in the fine, "scarce" branches that other models typically miss, such as those found in complex liver vessel structures.

Conclusion

By integrating the laws of differential geometry with deep learning, this framework moves beyond simple distance-based matching. While the OGMC model currently relies on an initial skeletonization of the vessels, it demonstrates that geometry is key to quantifying "topological likelihood." Future work may involve integrating this geometric logic directly into the neural network's training loop to create a fully end-to-end system.

Summary updated

Abstract

There are many unsolved problems in vascular image segmentation, including vascular structural connectivity, scarce branches and missing small vessels. Obtaining vessels that preserve their correct topological structures is currently a crucial research issue, as it provides an overall view of one vascular system. In order to preserve the topology and accuracy of vessel segmentation, we proposed a novel Morphology Edge Attention Network (MEA-Net) for the segmentation of vessel-like structures, and an Optimal Geometric Matching Connection (OGMC) model to connect the broken vessel segments. The MEA-Net has an edge attention module that improves the segmentation of edges and small objects by morphology operation extracting boundary voxels on multi-scale. The OGMC model uses the concept of curve touching from differential geometry to filter out fragmented vessel endpoints, and then employs minimal surfaces to determine the optimal connection order between blood vessels. Finally, we calculate the geodesic to repair missing vessels under a given Riemannian metric. Our method achieves superior or competitive results compared to state-of-the-art methods on four datasets of 3D vascular segmentation tasks, both effectively reducing vessel broken and increasing vessel branch richness, yielding blood vessels with a more precise topological structure.
BibTeX
@misc{zhu2023morphology,
  author = {Yuntao Zhu and Yuxuan Qiao and Xiaoping Yang},
  title = {{Morphology Edge Attention Network and Optimal Geometric Matching Connection model for vascular segmentation}},
  year = {2023},
  eprint = {2306.01808},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.01808},
}
Figure