Sub-Visible Particle Classification and Label Consistency Analysis for Flow-Imaging Microscopy Via Machine Learning...

Area of Science:

• Pharmaceutical analysis
• Biopharmaceutical quality control
• Particle characterization

Background:

• Sub-visible particles are critical quality attributes in parenteral pharmaceuticals.
• Flow-imaging microscopy generates particle images, requiring characterization.
• Machine learning (ML) is used for particle detection and classification, often needing manual expert labeling.

Purpose of the Study:

• To develop and evaluate ML techniques for characterizing sub-visible particles in biopharmaceuticals.
• To assess the utility of unsupervised learning and one-class classifiers for particle analysis.
• To investigate the consistency of expert-based particle labeling.

Main Methods:

• Generated artificial datasets mimicking real-world particle populations (silicone oil, protein, glass).
• Applied unsupervised learning to describe particle composition within samples.
• Trained independent one-class classifiers to detect specific particle types (silicone oil, glass).
• Evaluated model performance against expert-labeled data, assessing inter-expert agreement.

Main Results:

• Unsupervised learning effectively described particle composition.
• One-class classifiers demonstrated suitability for heterogeneous flow-imaging data.
• Low inter-expert agreement was observed, particularly for particles smaller than 8 µm.
• Unsupervised learning can assist in the particle labeling process.

Conclusions:

• One-class classifiers are a viable approach for analyzing complex flow-imaging microscopy data.
• Unsupervised learning offers potential to streamline the particle identification workflow.
• The subjectivity and inconsistency in expert labeling, especially for small particles, necessitate further investigation and reporting of label confidence.