A trainable clustering algorithm based on shortest paths from density peaks

Area of Science:

• Biomedical data analysis
• Computational biology
• Machine learning

Background:

• Clustering is vital for analyzing empirical data, particularly in biomedical research.
• Traditional clustering methods rely on point-to-point similarity and local rules, which can introduce artifacts with heterogeneous data structures.
• Identifying distinct groups in complex datasets remains a challenge.

Purpose of the Study:

• To propose a novel clustering algorithm that overcomes limitations of traditional methods.
• To develop a trainable algorithm capable of adapting to specific datasets and applications.
• To enhance the accuracy and robustness of group identification in heterogeneous biomedical data.

Main Methods:

• The proposed algorithm evaluates path properties between data points, rather than direct point-to-point similarity.
• It employs a global optimization approach, avoiding local choices that can lead to suboptimal solutions.
• A trainable path classifier is incorporated, allowing adaptation through examples of valid and invalid paths.

Main Results:

• The algorithm successfully identifies heterogeneous groups in challenging synthetic datasets.
• It demonstrates efficacy in segmenting highly nonconvex immune cells from confocal microscopy images.
• The method accurately classifies arrhythmic heartbeats in electrocardiographic signals.

Conclusions:

• The novel path-based clustering algorithm offers a robust alternative to traditional methods for complex data.
• Its trainability and global optimization approach enable superior performance in diverse biomedical applications.
• This technique holds significant potential for advancing data analysis in fields like immunology and cardiology.