Gramian angular fields for leveraging pretrained computer vision models with anomalous diffusion trajectories.

Area of Science:

• Physics
• Chemistry
• Biology
• Ecology
• Computer Science

Background:

• Anomalous diffusion occurs across various scales, from atomic to ecological systems.
• Characterizing diffusion provides crucial insights into system dynamics and enables interdisciplinary study.
• Identifying diffusion regimes and exponents is vital for multiple scientific fields.

Purpose of the Study:

• To develop a novel data-driven method for analyzing anomalous diffusion trajectories.
• To leverage computer-vision models for characterizing diffusion and inferring anomalous diffusion exponents.
• To improve the analysis of short, challenging trajectories common in single-particle tracking.

Main Methods:

• Encoding one-dimensional trajectories into images (Gramian matrices) using Gramian angular fields (GAF).
• Utilizing pretrained computer-vision models (ResNet, MobileNet) to process GAF images.
• Applying the method to short raw trajectories (lengths 10-50) for characterization.

Main Results:

• The GAF method effectively preserves the spatiotemporal structure of trajectories.
• The approach successfully characterizes underlying diffusive regimes and infers the anomalous diffusion exponent (α).
• GAF-based analysis demonstrates superior performance compared to current state-of-the-art methods, especially for short trajectories.

Conclusions:

• Gramian angular fields offer a powerful new approach for analyzing anomalous diffusion.
• This method enhances the accessibility and application of machine learning in diffusion studies.
• The technique shows promise for improving the analysis of challenging single-particle tracking data.