Assessment of subvisible particles in biopharmaceuticals with image feature extraction and machine learning

Abstract

An image classification tool was developed to classify subvisible particles, namely silicone oil (SO) and non-silicone oil (NSO; protein aggregate, rubber closure, and air bubble) particles, present in biopharmaceuticals using feature extraction in FlowCam (R) images, and the outcomes were validated with six machine learning (ML) classifiers. The image classification tool set at specific configuration: condition 1 - CDEP {(compactness, diameter, elongation, perimeter) = (8.06, 10.44, 13.30, 27.20)} identified SO, while the configuration set at condition 2 - CDEL {(compactness, diameter, elongation, length) = (3.14, 24.48, 1.97, 4.28)} detected NSO. The classification tool was particularly useful in detecting the release of SO after exposure to the stress sources. Additionally, the morphological features-based classification tool (p < 0.05) enhanced the predictive accuracy of the ML classification tools (>= 97.2 %). Specifically, CNN (100 %) outperformed na & iuml;ve Bayes (99.3 %), linear discriminant analysis (98.4 %), artificial neural network (98.1 %), support vector machine (SVM 97.2 %). Bootstrap forest was excluded because it failed to classify SO in a large dataset. The developed classification tool could be an alternative in classifying the image datasets without the burden of complex ML tools. Such image-based classification tool can be computationally economical solution in quality control of the protein formulations.