Shape-to-graph mapping method for efficient characterization and classification of complex geometries in biological...

Area of Science:

• Biomedical imaging analysis
• Computational biology
• Machine learning applications

Background:

• Increasing volume and quality of biomedical imaging data necessitate advanced computational tools.
• Existing algorithms often require data-specific tailoring, limiting their general applicability.
• Automated extraction of quantitative information from complex biological structures remains a challenge.

Purpose of the Study:

• To develop a broadly applicable computational approach for quantifying and classifying complex shapes and patterns in biological images.
• To overcome the limitations of data-specific algorithms in biomedical image analysis.
• To enable efficient and reliable automated extraction of quantitative information from diverse imaging datasets.

Main Methods:

• Mapping all shape boundaries within an image onto a global, information-rich graph.
• Utilizing machine learning on multidimensional graph measures for analysis.
• Integration of graph theory and machine learning for pattern recognition.

Main Results:

• Successfully extracted subtle structural differences in endothelial tube formations, distinguishing visually similar images.
• Trained an algorithm to identify biophysical parameters governing multicellular network formation in a collective cell behavior model.
• Analyzed cellular responses in U2OS cell cultures to various small molecule perturbations.

Conclusions:

• The presented graph-based machine learning approach offers a broadly applicable solution for complex shape and pattern quantification in biomedical imaging.
• This method demonstrates versatility across different biological assays and simulation models.
• The approach facilitates robust analysis of subtle structural variations and cellular responses to perturbations.