Classification of Arcobacter species using variational autoencoders

Abstract

Arcobacter (formerly classified as Campylobacter spp.) are curved-to helical, Gram-negative, aerobic/microaerobic bacteria increasingly recognized as human and animal pathogens. In collaboration with Lincoln and Purdue University, we report the first experimental result of laser-based classification method of bacterial colonies of these species. This technology is based on elastic light scatter (ELS) phenomena where incident laser interacts with the whole volume of the colony and generates a unique fingerprint laser pattern. Here we report a novel development and application of deep learning algorithm to classify the scatter patterns of Arcobacter species using variational autoencoders (VAE). VAE creates set of normal distributions. Each of these distributions are responsible for certain properties of the original images. We used VAE to identify features in the features space for several hundred images which includes size of the colony based on scatter size, intensity of the image, and, the number of rings within the image, and so on. Thus each sample within our image database can be coded with sets of features that facilitates fast preliminary search for similar images allowing clustering of similar patterns in feature space. In addition, such initial selection could assist in identifying non-bacterial scatter patterns (i.e. bubbles or dust spots in the agar), or doublets where two colonies are overlapping during the acquisition time thus removing non-biological artifacts prior to analysis. An interesting result was that while VAE created far more realistic synthetic images closer to the original image, a simple autonencoder resulted in better cluster separation.