Developing machine learning models with pharmaceutically relevant microscopy images to improve classification of multi particulate systems for parenteral delivery.

This study evaluates the development and performance of convolutional neural network (CNN) models for classifying subvisible particles in parenteral products using a comprehensive and pharmaceutically relevant particle image library. Specifically, representative morphologies of glass lamellae were curated for the first time and incorporated into the particle library, along with other intrinsic particle types such as protein aggregates, silicone oil droplets, and rubber. CNN models trained on complete particle libraries achieved high accuracy and stable classification across complex multi particulate mixtures, whereas models trained on limited categories or incomplete morphologies showed reduced robustness and increased misclassification, particularly for transparent or low‑contrast particles. These findings demonstrate that classification performance is highly dependent on the representativeness of the training dataset and the inclusion of challenging particle classes. The results also align with expectations outlined in compendial documents and regulatory guidance, highlighting the importance of well‑curated particle libraries and explainable model behavior to support reliable, scientifically justified particle identification in pharmaceutical quality assessments of parenteral products.