Design and Validation of Hyperspectral Elastic Light Scatter Phenotyping Instrument for Bacterial Colonies

Iyll-Joon Doh, Purdue University

Abstract

An optical technique that discriminates microbial organisms using an elastic light scatter (ELS) pattern, known as BActeria Rapid Detection using Optical Technology (BARDOT), has shown excellence in pathogen screening which effectively saves time and cost during the identification. Owing to the successful implementation of the light-scattering technique in microbiology, a series of studies on the light scatter pattern have been conducted to improve the technology. As an extended study of the multispectral application in BARDOT, a hyperspectral elastic light-scatter phenotyping instrument (HESPI) was developed to increase the ability to discriminate and detect foodborne pathogens. The newly designed instrument integrated a supercontinuum (SC) laser into the traditional BARDOT system to provide a broad spectrum of the light source. An acousto-optic tunable filter (AOTF) was utilized to select the wavelength of interest, allowing multiple spectral patterns in a single measurement. Owing to the filtering mechanism of AOTF, the wavelength of the laser was shifted rapidly so the overall acquisition time of 80 hyperspectral patterns was less than30 seconds. A pair of optical lenses were used to compensate for the beam spot movement caused by the wavelength-dependent separation angle at the exit of AOTF. To capture the transmitted scattering patterns, a complementary metal-oxide-semiconductor (CMOS) sensor was placed under the bacteria sample plate.For a comprehensive understanding of the ELS patterns, at first, the diverse nature of bacterial colony morphology was explored. Using the optical scatter model based on the scalar diffraction theory, the forward light-scatter patterns were simulated with respect to various colony shapes. The numerical predictions were then compared to the scattering patterns that were experimentally obtained from the colonies with various elevation profiles. The experimental verification proved a strong correlation between the colony morphology and the ELS pattern, as an excellent agreement between the simulation and the experiment observed. Second, the wavelength-dependent characteristics of the ELS patterns were investigated. Based on the theoretical and experimental interrogation, the wavelength of the incident laser beam affected the shape of ELS patterns by the overall size, the number of diffraction rings, and the gap distance between the rings.The performance of HESPI was validated by differentiating green leafy microflora using the hyperspectral ELS patterns. A group of bacteria that were poorly classified with the traditional single-wavelength method was selected to prove the improvement of the classification by the hyperspectral application. HESPI was utilized to measure the hyperspectral ELS patterns of their colonies, and for the classification, the descriptive features were extracted from the patterns at 70 selected wavelengths within the 473 – 709 nm region. A classification model was constructed for every wavelength, and the classification accuracy of the individual model ranged from 88.7% to 93.2%. The classification result also showed that colonies of varied species produced distinctive scatter features at a particular spectral band. When employing the entire wavelengths for the classification, the more number of wavelengths consequently led to an increase in the number of scatter-pattern features. This could cause the classifier's overfitting and negatively affect the classification. Therefore, the presented work incorporated various feature reduction and selection procedures to enhance the robustness and ultimately lessen the complexity of data collection. A classification model with feature reduction improved the overall classification rate to 95.9% after selecting meaningful predictors.

Degree

Ph.D.

Advisors

Bae, Purdue University.

Subject Area

Optics|Agronomy|Artificial intelligence|Computer science|Food Science|Microbiology

COinS