BayeTopo: Bayesian-Based Topology-Guided Learning for Vascular Imaging Segmentation.

Area of Science:

• Medical Image Analysis
• Computational Biology
• Artificial Intelligence

Background:

• Vascular segmentation is crucial for diagnosing vascular diseases but is hindered by variability in vessel representations.
• Existing topology-guided methods struggle to model global-to-local dependencies, creating a trade-off between global topology and local geometry.
• This limitation impacts the accuracy and semantic consistency of automated vascular segmentation.

Purpose of the Study:

• To propose a Bayesian-based topology-guided (BayeTopo) learning approach for enhanced vascular segmentation.
• To address the challenge of modeling global-to-local dependencies in vascular networks.
• To improve the accuracy and semantic consistency of automated medical image segmentation.

Main Methods:

• Developed a Bayesian-based topology-guided (BayeTopo) learning approach incorporating a prior for local geometry conditioned on global topology.
• Implemented a topology-guided diffusion model to optimize conditional probability and infer local geometry from global topology.
• Utilized an inhomogeneous diffusion process for orderly information degradation from global topology to local geometry.

Main Results:

• BayeTopo demonstrated superior performance and generalization across six datasets covering three vascular network types and four imaging modalities.
• The method outperformed previous topology-guided learning and diffusion-based models in vascular segmentation tasks.
• Case studies confirmed enhanced semantic consistency in local vascular regions and improved topological accuracy.

Conclusions:

• The proposed BayeTopo approach effectively captures global-to-local dependencies for robust vascular segmentation.
• This method offers significant improvements in accuracy and semantic consistency for medical image analysis.
• BayeTopo shows strong potential for clinical applications in diagnosing vascular-related diseases.