Improved sub-visible particle classification in flow imaging microscopy via generative AI-based image synthesis.

Area of Science:

• Pharmaceutical analysis
• Biotechnology
• Artificial Intelligence

Background:

• Sub-visible particle analysis is crucial for drug quality control, often using flow imaging microscopy and deep learning.
• Data scarcity and class imbalance, especially for rare particles like silicone oil and air bubbles, hinder accurate classification.
• Existing methods struggle with imbalanced datasets, impacting the reliability of identifying critical impurities.

Purpose of the Study:

• To develop a diffusion model for generating high-fidelity synthetic particle images.
• To address data imbalance issues in sub-visible particle datasets for improved deep learning model training.
• To enhance the accuracy of multi-class classifiers for pharmaceutical particle identification.

Main Methods:

• Development of a state-of-the-art diffusion model for image generation.
• Augmentation of training datasets with high-fidelity, synthetically generated particle images.
• Large-scale experimental validation using a dataset of 500,000 protein particle images.

Main Results:

• Generated synthetic particle images closely resemble real images in visual quality and structure.
• Diffusion model-generated data effectively augmented training datasets, enabling robust multi-class classifier training.
• Significant improvement in classification performance was observed without any negative impact on accuracy.

Conclusions:

• Diffusion models offer a powerful solution for data augmentation in imbalanced particle analysis datasets.
• This generative AI approach enhances the reliability and accuracy of deep learning models in pharmaceutical quality control.
• The study promotes open research by releasing diffusion models, trained classifiers, and an integration interface.