Journal of Spectral Imaging,   Volume 7   Article ID a6   (2018)

Peer reviewed paper

Unsupervised classification of individual foodborne bacteria from a mixture of bacteria cultures within a hyperspectral microscope image

Matthew Eadya and Bosoon Parkb,*
aThe US National Poultry Research Center, US Department of Agriculture — Agricultural Research Services, Athens, GA 30606, USA
bThe US National Poultry Research Center, US Department of Agriculture — Agricultural Research Services, Athens, GA 30606, USA. E-mail: [email protected]

Salmonella is a leading cause of foodborne illness. Traditional detection methods require lengthy incubation periods or expensive reagent kits. Hyperspectral microscope images (HMIs) have been previously investigated as a method for early and rapid detection of bacteria by using a spectral signature that is unique to the organism. Previous HMI use with bacteria has consisted of supervised classification with hypercubes collected for single culture images isolated from highly selective growth media. In order to move forward with HMI as a detection tool in the food industry, unsupervised classification of bacteria cells in mixed culture HMIs was investigated. Four foodborne bacteria cultures, S. Typhimurium (ST) E. coli (Ec), S. aureus (Sa) and L. innocua (Li) were combined in seven different culture combinations with HMIs collected between 450 nm and 800 nm. A k-means divisive cluster analysis (CA) was implemented and mixed culture image sets were found to contain between two and four clusters. CA cluster accuracy was obtained by assigning a dummy variable of the proposed CA classification, then carrying out a discriminant analysis. From the mixed culture HMIs, 700 bacteria cells were classified and accuracies were between 91.92% and 100%, with six of the seven HMI sets resulting in > 97% accuracies. A distance measure between clusters was applied to identify unknown clusters based on single culture reference samples of the four bacteria used. Results showed that the CA has potential for unsupervised classification of bacteria cells, but the distance metric was not an adequate method for identifying the unknown cluster based on reference spectra, potentially due to the collinearity amongst bacteria spectra.

Keywords: foodborne bacteria, hyperspectral microscope images, cluster analysis, discriminant analysis, distance classifier