Computer Vision for Cell Line Development

Abstract

Anomalies in Cell Line Development prove to have significant impact on material and opportunity cost when screening for the Master Cell Bank that is used for all clinical drug development. Cell Line Development scientists spend hundreds of hours collectively identifying anomalies in fluorescent and brightfield imagery to ensure only high-performing cell clones are downselected for testing. The use of computer vision models alleviates this burden on scientists and better standardizes the selection process. Three techniques were tested for classifying anomalous and nominal fluorescent images: an autoencoder, an edge CNN and an RGB SVM. Examining performance through composite metrics such as F1 Score and MCC, the autoencoder (0.8744 and 0.8619, respectively) outperformed the edge CNN (0.8488 and 0.8257) and RGB SVM (0.8343 and 0.8252) for fluorescent anomaly classification. The high performance of the autoencoder came from training solely on anomalous images and using a percentile-based threshold to classify images on their reconstruction error. Data robustness proved to be an issue, with certain test datasets having worse performance due to inherent variability of images within both nominal and anomalous classes. Gathering and labeling more datasets for training and testing will allow models to learn from this variability and provide higher confidence in model performance for real-time screening applications. Adjusting the structure of the traditional autoencoder to that of a variational autoencoder will also help with learning the variability of images within classes, and improve performance on previously unseen data. Overall, the current iteration of the models proves to be beneficial for anomaly detection in Cell Line Development and demonstrates that some modifications to data sourcing and model architecture could see even better performance. These same techniques could be applied to similar biopharmaceutical applications provided care is taken to properly source clean and labeled image data and construct appropriate model architectures for the images' inherent features.