Machine Learning-Enabled Image Comparability Assessment for Flow Imaging Microscopy Across Platforms

Background/Objectives: The rapid development of biopharmaceuticals has heightened attention from both industry and regulatory agencies toward product quality, particularly regarding subvisible particles as a critical quality attribute. Existing pharmacopoeial methods, Light Obscuration (LO) and Microscopic Particle Count (MC), exhibit limitations in meeting increasingly refined analytical requirements. Flow Imaging Microscopy (FIM) technology shows promise as an alternative, yet its standardized methodologies are still under development. Methods: This study employed polystyrene microsphere standard beads and intravenous immunoglobulin to perform instrument standardization and consistency evaluations on FIM instruments sharing the same operating principles but from different manufacturers. The consistency and transferability of particle counting across platforms were assessed. Additionally, particle images obtained from parallel testing on two platforms were classified using confusion matrices based on convolutional neural networks and the Unified Manifold Approximation and Projection (UMAP) dimensionality reduction method. Results: This study investigated the consistency and developed a transfer strategy for particle counting results across different FIM platforms. Analysis of particle image classification confirmed the consistency of image-based categorization while also revealing the complexity associated with cross-platform image recognition. Conclusions: The findings provide valuable insights for the further standardization of Flow Imaging Microscopy, supporting its potential as a reliable analytical tool for subvisible particle analysis in biopharmaceutical quality control.