Automated Visual Inspection of Lyophilized Products via Deep Learning and Autoencoders

Abstract

Manual visual inspection of every sterile parenteral product for defects is costly, cumbersome, and inconsistent. Industry standard currently relies on a semi-automated method, but the percentage of vials that require additional human inspection is high, at 30%. Using deep learning can help reduce this percentage, but there are a number of challenges. In particular, the dataset for defective lyophilized products is not only small, but also suffers from class imbalance because of the small number of defects.

In this thesis, we test the performance of well known deep learning neural network architectures including VGG16 and ResNet50. We compare results from training these architectures from scratch to results using fine-tuning of pretrained variants of the models, and find that the pretrained variants not only help improve accuracies, but help the model learn the correct reasoning for the defect classification decision, in terms of identifying defect location in an image. Furthermore, we show that autoencoders can be used to create classifiers that perform just as well as the pretrained VGG16 and ResNet50 models for our vial image datasets. Lastly, we demonstrate that simple data augmentation techniques do not improve the training of our vial defect classification models.