Classification of Salmonella enterica serotypes with selective bands using visible/NIR hyperspectral microscope images

Optical detection of foodborne bacteria such as Salmonella classifies bacteria by analysing spectral data, and has potential for rapid detection. In this experiment hyperspectral microscopy is explored as a means for classifying five Salmonella serotypes. Initially, the microscope collects 89 spectral measurements between 450 and 800 nm. Here, the objective was to develop correct classification of five serotypes with optimal spectral bands selected through multivariate data analysis (MVDA), thus reducing the data processing and storage requirement necessary for practical application in the food industry. An upright digital microscope is equipped with an acousto-optical tuneable filter, electron multiplying charge-coupled device, and metal halide lighting source. Images for each of the five serotypes were collected, and informative bands were identified through a principal component analysis, for four abbreviated spectral ranges containing 3, 7, 12 and 20 spectral bands. The experiment was repeated with an independent repetition and images were collected at each of the reduced band sets, identified by the first repetition. A support vector machine (SVM) was used to classify serotypes. Results showed that with the first repetition, classification accuracy decreased from 99.5% (89 bands) to 84.5% (3 bands), whereas the second repetition showed classification accuracies of 100%, possibly due to a reduction in spectral noise. The support vector machine regression (SVMR) was applied with cross-validation, and had R-2 calibration and validation values >0.922. Although classification accuracies through SVM classification showed that as little as 3 bands were able to classify 100% of the samples, the SVMR shows that the smallest root-mean squared-error values were 0.001 and 0.002 for 20 and 12 bands, respectively, suggesting that the 12 band range collected between 586 and 630 nm is optimal for classifying bacterial serotypes, with only the informative HMI bands selected. Lay Description Optical detection of pathogenic bacteria can offer the potential for early and rapid identification without the need for high cost molecular testing or traditional plating techniques that typically require between several days and a week to complete. Optical characteristics assess bacteria by collecting a spectral fingerprint that is unique to the organism. Here, hyperspectral microscopy is used to detect five serotypes within the Salmonella species on a cellular level. The hyperspectral microscope collects images at 89 wavelengths along the visible and near infrared light spectrums. These images are stacked on top of each other, creating a spectral fingerprint for each cell in the image. The spectral information from the cell is then processed through multivariate data analyses. A drawback of hyperspectral imaging is the large datasets that are accumulated. Theseanalyses reduce the amount of information generated from hyperspectral image cubes to easily interpreted plots and graphs. In addition to classification, the second objective was to reduce the number of wavelengths collected by the hyperspectral microscope to as few as possible, while maintaining high classification accuracy for the Salmonella cells. Images were collected for the five serotype samples, with important spectral bands identified through the multivariate analyses. Then, a second independent repetition of images were collected, but only at the spectral ranges identified as informative. Overall, images for all samples were collected with 89, 20, 12, 7 and 3 spectral bands. Multivariate classification analyses were performed on each repetition, with the first repetition seeing an overall decrease in accuracy as the bands are reduced, from 99.5% (89 bands) to 84.5% (3 bands). Eliminating the spectral bands deemed uninformative showed classification accuracies to be 100% for all reduced variable sets. Using multivariate data analyses to compare similarities between the two repetitions showed the strongest correlation with the reduced spectral range of 586-630 nm, containing 12 measured spectral bands. This, in combination with 100% classification when collecting only the informative spectral bands suggests that accurate classification can be achieved through hyperspectral microscopy within the spectral range of 586-630 nm. This is a 86.5% reduction in the data processing and storage requirement, moving towards application of this detection method.