Automating Visual Inspection of Lyophilized Drug Products With Multi-Input Deep Neural Networks

摘要

In the increasingly automated world of biotech manufacturing, the inspection of lyophilized drug products remains a cumbersome and manual process, relying on human operators to inspect finished vials for glass defects, contaminants, and other possible issues. Automating this procedure is challenging due to “fogging” on glass vials and low occurrence of defective samples for training decision models. In this work, we apply deep neural networks to automated classification of lyophlized product vials via computer vision. The proposed approach utilizes shared convolutional layers to account for multiple images acquired of the same vial at various rotations, and transfer learning is examined as a tool to partially overcome the lack of defective data in industrial manufacturing applications. The method is tested using a real-world industrial product as a representative case study. We find that 85-90% of defects can be detected from a single camera angle, demonstrating the potential of multi-input neural networks in lyophilized drug product inspection. Moreover, the results suggest that transfer learning enhances the generalization ability of learned models, even when using data designed for a very different task.