Principled network extraction from images.

Area of Science:

• Computational imaging
• Network science
• Image analysis

Background:

• Natural systems often exhibit network-like structures crucial for understanding their properties.
• Extracting formal network definitions (nodes, edges) from raw image data remains a challenge.
• Existing image-processing techniques often rely on heuristics and lack principled grounding.

Purpose of the Study:

• To develop a scalable and efficient model for extracting network topologies from images.
• To provide a principled alternative to heuristic-based image-processing methods.
• To validate the model's performance on diverse real-world image datasets.

Main Methods:

• Formulating network extraction as a routing optimization problem.
• Minimizing an energy function representing operational and infrastructural costs.
• Leveraging recent advancements in optimal transport theory.

Main Results:

• The model successfully extracts network topologies from images of retinal vascular systems, slime mould, and river networks.
• Networks extracted from retinal vascular images show higher similarity to hand-labeled ground truth.
• Consistent performance across datasets, even in the absence of ground truth data for rivers and slime mould.

Conclusions:

• The proposed optimal transport-based model offers a principled and efficient method for network topology extraction from images.
• The approach demonstrates superior performance and consistency compared to traditional image-processing routines.
• The method is highly automatable, requiring minimal supervision for application to various datasets.