Convolutional neural network-based image classification to understand breakup of oat cell structures due to high-pressure homogenization.

High-pressure homogenizers are used to disrupt oat cell aggregates in beverage processing applications. The mechanism of disruption remains poorly understood. Partially, this can be due to that previous studies either draw conclusions from individual micrographs or measure aggregate properties without accounting for the diverse morphologies of cells and tissues. This investigation combines experiments conducted over a range of homogenizing conditions with static image analysis, automatic particle identification, and a convolutional neural network (CNN) to quantitatively study how different oat cell morphologies respond to high-pressure homogenization. The CNN can track the evolution of seven distinct cell tissue morphologies. Homogenization results in breakup across all seven morphologies. However, different cell morphologies are differently susceptible to breakup, as seen from a shift in the relative percentage of particles in different classes. More specifically, whereas class I (possibly identified as aleurone cell clusters) is the most common category of large particles in the premix, class VII (possibly identified as pericarp cell clusters) dominate after homogenization. Investigations of the evolution of particle size distributions across the number of passages suggest differences in terms of breakup mechanisms. Methodologically, the contribution opens for future studies that combine the directness of microscopy with the strength of quantitative techniques.