Noise-free microbial colony counting method based on hyperspectral features of agar plates

Highlights

• A noise-free colony count method identifying noise with similar colors or shapes to those of colonies was proposed.

• A cluster-segmenting calibration model based on the spectral features of colonies and their background was built.

• A colony-separating calibration model based on the spectral features of colony centers and its borders was built.

• The proposed method achieved high correlation (R² = 0.9998) with the human vision method.

Abstract

A noise-free bacterial colony counting method identifying noise (i.e., sausage, bacon, and millet fragments) with similar colors or shapes to those of colonies was developed for food quality assessment. First, spectral features corresponding to colony cluster regions and background regions (agar medium and food fragments) were extracted after collection of hyperspectral images. A cluster-segmenting calibration model that could identify colony clusters and background regions was developed. Second, spectral features of colony centers and borders were extracted, and a colony-separating calibration model that could separate single colonies from clusters (multiple colonies contacting each other) was developed. Third, each pixel of an agar plate hyperspectral image was identified using established calibration models, enabling the colonies on the agar plate to be counted successfully (R² = 0.9998). The results demonstrated that the proposed method could identify the noises caused by food fragments with similar colors or shapes to those of colonies.

Introduction

A method of colony counting using solid agar plates has been widely employed to quantify viable microbial cells for food quality assessment (Corry, Jarvis, & Hedges, 2010). The colony counting method involves smearing a diluted bacterial suspension of food products on a solid agar plate (Chiang, Tseng, He, & Li, 2015). Because each viable cell on the plate grows and forms a single colony, the number of viable microbial cells in food products can be evaluated by counting the colonies (Lobete, Fernandez, & Van Impe, 2015). Colony counting provides essential indicators regarding probiotics that reflect food nutrition and harmful organisms that affect food safety (Feroz et al., 2016, Piepers et al., 2014, Rysanek et al., 2009). Numerous reference methods and national standards concerning food quality related to various organisms are based on this type of colony counting (AOAC International, 1989, China Food and Drug Administration, 2016, Marshall, 1992). Conventional colony counting is manually performed by well-trained operators; thus, it is low throughput, laborious, and time consuming in practice (Ferrari, Lombardi, & Signoroni, 2017).

Various automated colony counting methods using computer vision have been proposed to alleviate the disadvantages associated with manual counting (Chen and Zhang, 2009, Dahle et al., 2004, Puchkov, 2010). An image-capture system is designed to collect two dimensional color or gray-scale images of an agar plate. Color features or shape features of the colony are extracted from the image and employed to distinguish the colony cluster from the background, separate single colonies from clusters composing touching colonies, and provide colony counting results automatically. These types of automated counting methods can obtain satisfactory results as long as interference noise, such as various food fragments, does not appear in the agar plate (Chiang et al., 2015, Ogawa et al., 2012).

Preparing pure samples of target cells do not contain food fragments is difficult. As a result, the food fragments may be incorrectly identified by computer vision as regular colonies. Researchers reported to the challenge of preparing pure samples of target microbial cells without residual food fragments (Araki, Matsuzaki, Sekita, Saito, & Matsuoka, 2010). Technicians working in government institutions of disease control and prevention also reported that food fragments frequently appeared on/in agar medium (Li, 2010). Because the color and shape of food fragments may be quite similar to those of regular colonies, food fragments are easily wrongly identified as colonies by computer vision and even by well-trained employees in practice.

Time-lapse imaging technology and triphenyl tetrazolium chloride (TTC) have been employed to reduce or eliminate the influence of food fragments with colors similar to those of colonies. Time-lapse imaging technology collects a series of images to record the dynamic growth of colonies from single cells. Because the signals of food fragments can be captured before those of the colony, this method can clearly distinguish colonies from food fragments (Ogawa et al., 2012). However, it requires many repetitions of the image acquisition process during the colony culture, and a complex plate holder that can automatically transport the agar plate from the incubation area to the imaging area must be installed. Researches determined that TTC can be used as an indicator because only colonies containing living cells can be marked with a red color. The use of TTC significantly increases the contrast between the colonies and the background; thus, segmentation of colonies becomes pretty simpler (Clarke et al., 2010). However, Wei reported that TTC affected the growth of microbial cells, which indicated that errors can be caused by the use of TTC in the colony growth stage (Wei, 2008).

The present study developed a novel method of separating colonies from their background based on spectral features caused by changes in chemical composition. Many studies have demonstrated that spectral features are sensitive to the chemical components of biological samples, and spectral techniques using ultraviolet (UV) (Tang et al., 2016), visible (VIS) (Bao et al., 2014), near infrared (NIR) (Li, Huang, Zhao, & Zhang, 2013) and mid infrared (MIR) (Botelho, Reis, Oliveira, & Sena, 2015) spectra have been successfully used for quantitative and qualitative analysis of chemical composition in various biological samples. Changes exist in the chemical composition between colony area and agar medium, between colony area and food fragments, and between colony center area and colony border area. Therefore, employing spectral features to identify colonies and food fragments with colors similar to those of the colonies is reasonable.

To utilize the new method, obtaining spectral data of the whole agar plate pixel by pixel is essential. Thus, hyperspectral imaging technology was employed to record the whole agar plate in this manner. Unlike conventional spectral technologies, such as UV, VIS, NIR and MIR, that rely on spot measurement, hyperspectral imaging technology combines conventional spectroscopy and imaging techniques to acquire a spectrum for each pixel in the two dimensional image of an object (Du et al., 2016, He et al., 2014). The chemical composition of the whole sample must be evaluated, and acquiring spectral data and spatial data from the sample surface simultaneously is essential. Hyperspectral imaging technology meets these requirements and has been successfully used for full assessment of chlorophyll, flavonoids, moisture, soluble solids and other chemical composition in various biological samples (Cheng et al., 2015, Cheng and Sun, 2015, Shi et al., 2012, Shi et al., 2011, Zhu et al., 2016). It is reasonable to assume that the differences in chemical composition caused by colonies, agar medium, and various noises can be characterized using hyperspectral imaging technology.

As discussed, this study aimed to develop a noise-free, high-precision method for automatic colony counting. Agar plates with colonies and food fragments were employed to acquire hyperspectral image data. Spectral features corresponding to colonies and backgrounds were employed to count the colonies automatically. The performance of the proposed method was compared with that of computer vision, and the practical feasibility of the proposed method is discussed.