Bayesian Particle Instance Segmentation for Electron Microscopy Image Quantification.

Area of Science:

• Materials Science
• Computer Vision
• Data Science

Background:

• Automating materials characterization via electron microscopy (EM) can accelerate scientific discovery.
• Current methods for analyzing EM images are often manual and time-consuming.
• Extracting quantitative particle information is crucial for materials research.

Purpose of the Study:

• To develop an automated method for semantic segmentation and localization of particles in EM images.
• To enable accurate computation of quantitative particle measures, such as size distributions and aspect ratios.
• To integrate uncertainty estimation into the analysis pipeline for improved accuracy and reliability.

Main Methods:

• A Bayesian deep-learning model was employed for semantic segmentation and particle instance localization.
• Epistemic uncertainty from the model was utilized to filter false-positive predictions.
• The developed method was integrated into the ImageDataExtractor package (version 2.0).
• A new dataset, Electron Microscopy Particle Segmentation (EMPS), was created and released.

Main Results:

• The model accurately segments and localizes particles in EM images.
• Quantitative particle measures (size, aspect ratio, etc.) were computed with uncertainty estimates.
• The ImageDataExtractor 2.0 provides a full pipeline for automated particle information extraction.
• The EMPS dataset offers a valuable resource for training and benchmarking segmentation models.

Conclusions:

• Automated particle analysis in EM images significantly accelerates materials discovery.
• The Bayesian deep-learning approach provides accurate quantitative measures with uncertainty quantification.
• The integrated pipeline and public dataset facilitate large-scale, data-driven materials research.