Detection of Campylobacter colonies using ... - PubAg - USDA

Sens. & Instrumen. Food Qual. (2010) 4:35–49 DOI 10.1007/s11694-010-9094-0

ORIGINAL PAPER

Detection of Campylobacter colonies using hyperspectral imaging Seung Chul Yoon • Kurt C. Lawrence • John E. Line • Gregory R. Siragusa • Peggy W. Feldner • Bosoon Park • William R. Windham

Received: 6 November 2009 / Accepted: 25 February 2010 / Published online: 12 March 2010 Ó US Government 2010

Abstract The presence of Campylobacter in foods of animal origin is the leading cause of bacterially induced human gastroenteritis. Isolation and detection of Campylobacter in foods via direct plating involves lengthy laboratory procedures including enrichments and microaerobic incubations, which take several days to a week. The incubation time for growing Campylobacter colonies in agar media usually takes 24–48 h. Oftentimes the problem is the difficulty of visually differentiating Campylobacter colonies from non-Campylobacter contaminants that frequently grow together with Campylobacter on many existing agars. In this study, a new screening technique using non-destructive and non-contact hyperspectral imaging was developed to detect Campylobacter colonies in Petri dishes. A reflectance spectral library of Campylobacter and non-Campylobacter contaminants was constructed for characterization of absorption features in wavelengths from 400 to 900 nm and for developing classification methods. Blood agar and Campy-Cefex agar were used as culture media. The study found that blood agar was the better culture medium than Campy-Cefex agar in terms of Campylobacter detection accuracy. Classification algorithms including single-band thresholding, band-ratio thresholding and

S. C. Yoon (&) K. C. Lawrence J. E. Line P. W. Feldner B. Park W. R. Windham U.S. Department of Agriculture, Agricultural Research Service, Richard Russell Research Center, 950 College Station Road, Athens, GA 30605, USA e-mail: [email protected] G. R. Siragusa Agtech Products, Inc, W227 N752 Westmound Drive, Waukesha, WI, USA

spectral feature fitting were developed for detection of Campylobacter colonies as early as 24 h of incubation time. A band ratio algorithm using two bands at 426 and 458 nm chosen from continuum-removed spectra of the blood agar bacterial cultures achieved 97–99% of detection accuracy. This research has profound implications for early detection of Campylobacter colonies with high accuracy. Also, the developed hyperspectral reflectance imaging protocol is applicable to other pathogen detection studies. Keywords Hyperspectral imaging Pathogen detection Campylobacter Food safety Poultry

Introduction Campylobacter bacteria are widespread in warm-blooded food-producing animals, and the presence of Campylobacter in foods of animal origin has been the most common cause of bacterially induced human gastroenteritis in the United States and other developed countries [1–3]. Poultry and poultry products, if incorrectly treated during processing or insufficiently cooked, are considered as the major cause of human Campylobacter infections although unpasteurized milk and untreated water are also frequently associated with human infections [1]. The majority of human campylobacteriosis cases are caused by Campylobacter jejuni, followed by Campylobacter coli and other species [4]. Despite the fact that it is time-consuming, direct plating of samples on agar media has been an effective technique for detection, isolation and enumeration of Campylobacter from a variety of different food samples. Typical colonies of Campylobacter on bloodsupplemented or charcoal-based agar media appear

123

36

smooth, convex, and glistering with a distinct edge or flat, shiny, translucent, and spreading with an irregular edge; colorless to grayish or light cream. The Campylobacter colonies can range pinpoint to 4–5 mm in diameter although growth may be confluent without distinct colonies [2, 3]. Isolation is usually done by a highly proficient laboratory technician who can discern Campylobacter-like colonies. The isolated presumptive colonies are further examined by dark-field or phase-contrast microscopy to find out whether the tested colony has the characteristic cell morphology and motility of Campylobacter species [2, 4]. The microscopic analysis is not a confirmatory identification test. The definite identification of Campylobacter colonies involves additional laboratory tests such as immunological techniques using antibody/antigen interactions such as a latex agglutination test, or molecular methods using polymerase chain reaction (PCR) or nucleic acid [2]. Oftentimes, it is difficult to visually differentiate Campylobacter colonies from non-Campylobacter contaminants that frequently grow on many existing agars [1, 5]. In the past, Fourier transform infrared (FT-IR) spectroscopy, which has been applied to solve various microorganism detection/identification problems, was studied to distinguish Campylobacters (C. jejuni and C. coli) at the species level [6]. However, FT-IR spectroscopy is not readily applicable to imaging of Petri dishes. Typically, FT-IR spectroscopy needs an infrared transparent window to mount a biological sample [7]. Recently, as a more readily accessible imaging solution, a visible and nearinfrared (VNIR) hyperspectral imaging technique was developed for detection of Campylobacter and non-Campylobacter organisms grown in Petri dishes [8]. Similarly, VNIR hyperspectral imaging was studied for fungi detection [9]. The hypothesis of the hyperspectral imaging technology for Campylobacter detection is that, while the colony morphology may provide important visual cues, spectral signatures of Campylobacter and non-Campylobacter colonies may provide vital information about the biochemical compositions of the organisms and thus facilitate the detection and identification of the organisms on agar media. However, the previous hyperspectral imaging study for Campylobacter detection had its limitation for early detection of Campylobacter cultures because the study was conducted with the cultures incubated for 48 h. Although the 48-h incubation time was often used for growing Campylobacter cultures in various studies such as the Food Safety and Inspection Service (FSIS) baseline study [10] and the International Organization for Standardization (ISO) standard for detecting Campylobacter [11], there was a need for a hyperspectral imaging study with the cultures incubated

123

S. C. Yoon et al.

for less than 48 h for more rapid detection of Campylobacter. A hyperspectral imaging technique that can rapidly distinguish Campylobacter colonies from nonCampylobacter cultures can serve as a screening method for identifying presumptive Campylobacter colonies nondestructively before performing any confirmatory identification tests. In this paper, we present a new screening technique based on non-destructive and non-contact hyperspectral imaging technology for detecting Campylobacter colonies incubated for 24 h. We address the challenges faced with the reduced incubation time. We also report the development of a hyperspectral image classification algorithm using continuum removal for Campylobacter detection. The specific objectives of the study were (1) to compare imaging effects of 24- and 48-h incubation times and (2) to develop hyperspectral image processing/classification algorithms and to compare their performances with the performances of the algorithms developed from the 48 h culture study.

Materials and methods Bacterial cultures Frozen cultures of Campylobacter species and frequently encountered non-Campylobacter species were maintained at the USDA-ARS Poultry Microbiological Safety Research Unit’s culture collection in Athens, Georgia. The list of the bacterial species used in the study is in Table 1. With the exception of the American Type Culture Collection (ATCC) strains, all other bacterial cultures were isolated from poultry samples consisting of either whole-carcass rinses or fecal/cecal specimens from conventionally reared broiler chickens or processing plants [5, 12]. Eleven Campylobacter strains (5 Campylobacter jejuni strains, 5 Campylobacter coli strains and 1 Campylobacter lari strain) and six different nonCampylobacter contaminants were used for the study. The initial liquid culturing that took 72 h was performed to revive the live cells from frozen stocks [8]. Following the initial liquid culturing step, 5 lL spots were inoculated onto agars and then incubated in a microaerophilic atmosphere at 42 °C for a total of 24 h. Each 5 lL spot was made from 108 cells/mL fluid suspension. The plates showing the best colony growth were selected from three dilution plates for the imaging experiments. The plates showing confluent or pinpoint growth were not selected. Agars used were blood agar (5% sheep’s blood agar, Remel, Inc., Lenexa, Kans.) and Campy-Cefex [13]. Campy-Cefex (Cefex) agars were prepared in-house in standard 100 mm 9 15 mm Petri dishes [12]. Cefex agar

Detection of Campylobacter colonies Table 1 Microorganisms used in this study

a

American Type Culture Collection, Manassas, VA, USA

37

Lab designation

Organism (genus and species)

Strain designation

Source

1

Campylobacter jejuni

ATCCa 49943

ATCC

2

Campylobacter coli

ATCC 49941

ATCC

3

Campylobacter lari

ATCC 43675

Human feces

4


CT-epi #5

Poultry

5


CT-epi #65

Poultry

6


Ty-16C

Poultry carcass Rinse

7


Ty-78


8

Campylobacter coli

CT-epi #8

Poultry

9

Campylobacter coli

CT-epi #18

Poultry

10

Campylobacter coli

Ty-19C


11 12

Campylobacter coli Sphingomonas paucimobilis

Ty-66 Contam. 1

Poultry carcass Rinse Poultry

13

Acinetobacter baumannii

Contam. 2

Poultry

14

Brevundimonas diminuta

Contam. 3

Poultry

15

Ochrobacterium sp.

Contam. 4

Poultry

16

Flavobacterium odoratum

Contam. 5

Poultry

17

Acinetobacter baumannii

Contam. 6

Poultry

is one of commonly used agar media to isolate Campylobacter species from food sources. Cefex is more selective than blood-supplemented agars. Spot plating and 24-h incubation To avoid confluent growth and cross-contamination, agar plates were inoculated with 5 lL spots at pre-determined and well-spaced locations on the agar surface. Figure 1 shows a schematic of the imaged spot locations. The 17 spots were inoculated on two separate plates with nine and eight spots, respectively. The ‘‘Plate A’’ plates contained six spots of Campylobacter subspecies (1 through 6) and three spots of non-Campylobacter contaminants (12 through 14). The ‘‘Plate B’’ plates contained five spots of Campylobacter subspecies (7 through 11) and three spots of non-Campylobacter contaminants (15 through 17). One experiment took 4 days from sample preparation (72 ? 24 h) to imaging (1–5 h). At the end of the 24-h incubation process, plates were taken out of the incubator and stored at room temperature for 1–5 h before being imaged. Five replicates of experiments were carried out over the period of 6 months from August 2007 until February 2008. The experimental protocols for the sample preparation and the imaging remained the same throughout the experiments. The two duplicate plates were prepared per plate type and per experiment. Therefore, a total of 20 Petri dishes per agar media were prepared for imaging.

System set up and imaging protocol The hyperspectral imaging system used for the experiments consisted of a hyperspectral imaging camera (ITD, Stennis Space Center, MS) covering a usable spectral range from 400 to 900 nm, a copy stand to attach the camera, a computer to control the camera and acquire images, an enclosure to block unwanted light, two 50 W tungsten halogen lamps, and a Petri dish holder. The hyperspectral imaging camera consisted of a CCD sensor and a diffraction grating spectrograph (ImSpector V10E, Specim, Oulu, Finland) and associated optical and mechanical components. A hyperspectral image cube was constructed from line scans of a stationary Petri dish by moving the frontal optic lens in front of the hyperspectral image camera via a motorized translation stage. For more details of the imaging system, refer to Yoon et al. [8]. For reflectance imaging, two halogen lamps having a color temperature of 4,700 K laterally illuminated a Petri dish at about 45° pointing down from the left and right sides. It was important to maintain the lateral illumination in order to reduce the glare effect (glints) from the shiny surface of agar media and colonies. A white Teflon plate was put under the Petri dish holder so as to increase the reflectivity of semi-transparent agar and colonies with thin layers. Without the Teflon plate, most of the light energy that passed through these materials was not sensed by the camera because the inside of the enclosure was painted in black and the angle of the light incidence was about 45°. Two-dimensional spectral images (i.e., line-scan images)

123

38

S. C. Yoon et al.

Fig. 1 Location map of 17 inoculation spots on two Petri dishes of agars. This location map can be used as a reference for interpreting the images presented in this paper. Refer to Table 1 for the detailed information about the organisms and their identification numbers (top/bottom rows: Campylobacter, middle rows: non-Campylobacter)

Plate A Inoculated Spots

Non-Campylobacter

Data processing and analysis The acquired hyperspectral images were pre-processed to reduce the image size and to suppress spectral noise. All images were spatially cropped down to 421 (W) 9 475 (H) and spectrally to 193 spectral bands ranging between 400 nm and 900 nm. The Savitzky–Golay smoothing filter (window size: 25; order of moment: 4) was applied to a spectrum at each pixel position independently to reduce the spectral random noise [14]. Intensity calibration was performed with a 75% reflectance Spectralon target (13 9 13 cm, SRT-75-050, Labsphere, North Sutton, N.H.) and a reflectance calibration model [8, 15]. In a comparative study between 75 and 100% reflectance Spectralon targets, we found that the 75% reflectance Spectralon target produced higher dynamic range calibration images. The percent reflectance value R at each pixel (x,y) of the zth wavelength band was obtained by the following calibration model: Im ðx; y; zÞ Id ðx; y; zÞ 75 Ir ðx; y; zÞ Id ðx; y; zÞ

2

1

9

8

7

14

13

12

17

16

15

6

5

4

11

10

B’’ type plate was put in the second row. The third and fourth rows were replicates of the top two rows, respectively. Figures 2 and 3 show the image mosaics of 24-h cultures grown in Cefex and blood agar media, respectively. The bacteria on agar may suffer from no growth, confluent growth or even cross-contamination. As shown in the first column of Fig. 2, the images of Petri dishes that showed no growth or confluent growth were excluded from the mosaic and thus from the data analysis. Also note that the images of the fourth column in Fig. 3 were taken with the plates rotated from the reference positions shown in Fig. 1. Some of Campylobacter cultures were cross-contaminated (circled spots in Fig. 2 and 3). In the last column plates of Fig. 3, the circled Campylobacter cultures (No. 8) were cross-contaminated by unidentified non-Campylobacter contaminants. The other circled Campylobacter cultures in Figs. 2 and 3 were cross-contaminated by unknown source which could be either different Campylobacter species or any other non-Campylobacter cultures.

ð1Þ

where Im is a measured raw value, Ir is a reference value on the surface of the 75% reflectance panel, and Id is a dark current. A mosaic-based image representation was used to facilitate data analysis and algorithm development, where a mosaic of multiple images can be treated as a single hyperspectral image. The calibrated hyperspectral images were tessellated without overlap into an image mosaic by date and type. In the mosaic, the images taken on the same day were stitched into the same column. The columnwise images were then added to the mosaic in chronological order (latest right). In each column of the mosaic, the ‘‘Plate A’’ type plate was put in the first row. The ‘‘Plate

123

3

Campylobacter

were captured by 2 (spatial) 9 4 (spectral) binning and 90-ms exposure time. The resulting pixel resolution of a line-scan image was 640 (spatial) 9 256 (spectral). A total of 475 lines were scanned. Hence, the dimension of the 3-D image data cube was 640 (W) 9 475 (H) 9 256 (wavelength).

Rðx; y; zÞ ¼

Plate B

Fig. 2 Hyperspectral image mosaic of 24 h cultures on Cefex: RGB color-composites. Cross-contaminated cultures were marked with circles

Detection of Campylobacter colonies

39 Campy. Non-Campy. Blood Agar

Fig. 4 Image mosaic of ground truth ROIs showing Campylobacter (magenta; ROIs in top/bottom rows of each plate), non-Campylobacter (green; ROIs in middle rows of each plate) and blood agar (blue; box shaped ROIs) Fig. 3 Hyperspectral image mosaic of 24 h cultures on blood agar: RGB color-composites. Cross-contaminated cultures were marked with circles

Ground-truth regions-of-interest (ROIs) for all 17 organisms plus agar media were manually prepared in such a way that only pure cultures could be selected in the ROIs. Glints and rim shadows were excluded in the ROIs. Although the lateral lighting reduced surface glare, glints were observed around colony edges (rims) along the direction of the lateral illumination. If possible, pixels that might have contained mixed spectra of agar media and organisms were also excluded. Mixed pixels were typically observed at the center of a spot with a translucent colony surface and at outside boundaries of a spot. Figure 4 shows the ground-truth ROIs of cultures on blood agar media categorized by three classes: Campylobacter, non-Campylobacter, and blood agar. They were manually obtained from the mosaic image shown in Fig. 3. A binary mask image (mosaic) was also generated in order to suppress the background outside and around the rim of each Petri dish. The ground-truth ROIs and the mask image were used to build a spectral library of pure organisms and to develop and evaluate classification algorithms. The data analysis and algorithm development were performed using IDL 6.3/ENVI 4.3 (ITT Visual Information Solutions, Boulder, Colo.) and MATLAB R2009a (The Mathworks, Natick, MA). Intensity thresholding of single-band image A hyperspectral imaging study with 48-h Campylobacter cultures found that there was a wavelength showing the statistical separability large enough to differentiate Campylobacter species from non-Campylobacter

microorganisms [8]. According to the study, the wavelength was 503 nm for blood agar media and 501 nm for Cefex media. The same intensity thresholding algorithm developed from the study was also adopted in this 24-h culture study. The single-band intensity thresholding algorithm classified a pixel’s relative reflectance value (Ir) by the following rule: Agar class;

if Ir T1

Campy: class; non Campy: class;

if T1 \ Ir \ T2 if Ir T2 ;

ð2Þ

where Ti (i = 1, 2) is a threshold value. Continuum-removal and feature extraction The continuum-removal analysis is aimed to quantify absorption bands depart from a common baseline [16]. The common baseline (i.e., continuum) is defined as the convex hull surrounding the data points of a reflectance spectrum [16–18]. In other words, the continuum consists of piecewise continuous lines connecting local maximum points of the reflectance spectrum (Fig. 5a). Continuum removal is a procedure to isolate a particular absorption feature for analysis by dividing the reflectance spectrum at each wavelength with the continuum at the corresponding wavelength: RCR = R/C where RCR is the continuumremoved spectrum, R is the reflectance spectrum and C is the continuum (Fig. 5). The continuum-removed spectral values range from 0 to 1. The first and last points in the reflectance spectrum are always local maxima for the continuum. Hence, the first and last points become 1 in the continuum-removed spectrum (Fig. 5b). After continuum removal was applied, the parametric quantification of

123

40

S. C. Yoon et al.

20

affect the continuum-removed spectra as long as the value at 700 nm is a local maximum for the continuum curve and the values at k1 nm, k2 nm are not local maxima. Reduction of input wavelengths to compute the continuum may be useful for multispectral imaging. We compared both the hyper-spectrum (all of 193 bands) and the multi-spectrum (4 bands at 401 nm, k1 nm, k2 nm and, and 701 nm) for the continuum removal.

15

Spectral feature fitting

(a) 40

Relative Reflectance (%)

35 30

Continuum C

25

Reflectance Spectrum R 10 5 400

450

500

550

600

650

700

750

800

850

900

850

900

Wavelength (nm)

Continuum Removed Reflectance

(b)

1 0.9

Band Depth = 1- RCR λ

0.8

RCR = R / C

2

0.7

Bandwidth λ

0.6

1

Band Ratio: R CR(λ 2) / RCR(λ 1)

0.5 0.4

400

450

500

550

600

650

700

750

800

Wavelength (nm)

Fig. 5 Example of continuum removal and absorption features. a Continuum over the mean reflectance spectrum of non-Campylobacter. b Continuum-removed spectrum, RCR = R/C

the absorption band features can be done with the calculation of band depth, bandwidth and wavelength position [19, 20]. As shown in Fig. 5b, the band depth D of each band can be calculated by subtracting the continuumremoved reflectance from 1: D = 1 - RCR [20]. In this study, the slope information of the continuum-removed spectrum was utilized to enhance differences in absorption features of Campylobacter and non-Campylobacter organisms. This slope information was represented by a band ratio (Fig. 5b). For band ratio computation of continuum-removed spectra, it is not always necessary to apply the continuum removal to all wavelengths. For instance as in Fig. 5, the continuum for the band ratio computation at k1 nm and k2 nm may only need reflectance values at four wavelengths: 401 nm (approximation), k1 nm, k2 nm and 701 nm (approximation), where k1 nm, k2 nm will be determined later. The continuum removal at these four points does not

123

Spectral feature fitting (SFF) is a method of matching the complete shape of measured spectra to reference library spectra in terms of similarity [18, 21, 22]. SFF requires to have the reference library spectra, to remove the continuum from both the library and unknown pixel spectra, and to superimpose a continuum-removed library spectrum over the unknown pixel spectrum by a scale factor and a subtraction. SFF determines a single multiplicative scaling factor that makes the reference library spectrum match the unknown spectrum. Assuming that a reasonable spectral range has been selected, a large scaling factor is equivalent to a deep spectral feature, while a small scaling factor indicates a weak spectral feature. A least-squares fit is calculated band-by-band between the unknown pixel spectrum and each of the reference library spectra. The total root-mean-square (RMS) error is used to form an RMS error image for each reference library spectrum. A ratio image of Scale/RMS was used to create a fit image that is a measure of how well the unknown spectrum matches the reference spectrum on a pixel-by-pixel basis. In this study, 3 classes (Campylobacter, non-Campylobacter and agar) were used for the SFF. Finally, the SFF fit image was applied to the maximum likelihood classifier. Classification using continuum-removed band ratio In hyperspectral image processing, a band ratio usually means the division of pixel values in one band by the values of the corresponding pixels in another band. The beauty of a band ratio technique is in its simplicity of algorithm design and its efficacy for differentiating materials. The band ratio is also invariant to shading because relative values are used. However, there is no golden rule or universal technique to select optimal bands for a band ratio from hyperspectral image bands. The bands for obtaining ratio features are typically chosen from domain knowledge specific to an application or empirical observations of spectral data. In this study, a band ratio technique was applied to two continuum-removed spectral bands that revealed the prominent slope differences between Campylobacter and non-Campylobacter classes. Specifically, two bands were selected corresponding to a


41

local minimum (at k1 nm) and a local maximum (at k2 nm) of the continuum-removed mean spectrum of the nonCampylobacter class. After computing ratios, a simple thresholding operation was applied to the band ratio values in order to classify them into two classes: Campylobacter and non-Campylobacter. The classified image was further processed for determining the sole identity of colonies showing mixed classification results. In practice, if an object was round and the rim of the object was classified differently from its inner object (we call it a ‘‘ring’’ pattern), then the rim could be merged into the inner object because the reflectance response of the rim of a colony was usually mixed with that of the background agar. To do this, we estimated the roundness of each object and designed decision rules that determined the merge of a rim into its inner object. The roundness of an object was computed by either of the following metrics: R1 ¼ 1 p r2 area =area ; ð3Þ R2 ¼ 4 p area=perimeter2 ; where ‘‘r’’ refers to the estimated radius of a virtual circle encompassing the object, and ‘‘area’’ and ‘‘perimeter’’ refer to the measured size of the object and the measured length of the object boundary. For obtaining the estimated radius ‘‘r’’ for R1 computation, we used the average value of the major axis length and the minor axis length of the object, divided by 2. The major (or minor) axis length is the length of the major (or minor) axis of the ellipse that has the same normalized second central moments as the object. For the roundness computation, holes in the object were filled. The metric R1 measured the closeness of an object shape to a circle. The metric R2 was introduced in Matlab example codes estimating the roundness of an object [23]. The metric R2 was sensitive to the roughness of object boundary. Both metrics R1 and R2 were equal to one only for a perfect circle. For a round object with smooth boundaries, R1 and R2 were typically between 0.86 and 1 (or 1.1). When the overall shape of an object was round but the boundaries were rugged, the metric R2 was not a good metric. Nonetheless, either metric was not good enough as a sole metric for estimating the object roundness. In reality, the shapes of colonies were too diverse and complex to be described sufficiently in terms of roundness only. The focus of the study was to post-process circular objects (not touching each other) showing mixed classification results (two classes). Therefore, the following decision rule was designed to classify an object pattern into ‘‘circular’’ or ‘‘broken ring’’. If ð0:85\R1 \1j0:85\R2 \1:1Þ & ðR2 [ 0:3 & R1 [ R2 Þ is true; the object is circular: If R1 [ 5 & R2 \0:2 is true; the object is a broken ring pattern:

ð4Þ

Only when an object was circular and consisted of two classes, the following rule was applied to determine the merger of two classes into one class. If the relative size of one class in the object is less than 10% of the object; the small size class is merged to the other class: ð5Þ The specifics of the developed classification algorithm are described in the below. Step 1: Background segmentation via single-band thresholding at 503 nm. Mask out background pixels outside of the plates. Step 2: Colony object segmentation via single-band thresholding. Extract objects of colony/culture spots using a single threshold at 503 nm. Pixels of agar media and cultures will be separated in Step 2. Step 3: Colony identification via band ratio thresholding. Apply continuum removal to the entire spectrum (hyperspectral CR) or a few bands (multispectral CR). Compute band ratio [k2 (458 nm)/k1 (426 nm)] and classify the ratio data by a threshold value. Step 4: Merge round objects with ring patterns. Step 4 is the post-processing for changing the classes of thin outer rings. If the size of a ring was smaller than 10% of the total size of the entire object, the class of the ring was changed to the class that the majority of the object had.

Results and discussion Spectral responses of cultures The mean, normalized reflectance spectra of 24 and 48 h cultures from both blood agar and Cefex agar are presented in Fig. 6. The mean reflectance spectra of 24 h cultures were obtained from pixels in the ground-truth regions of interest whereas them of 48 h cultures were obtained from the spectral library built by Yoon et al. [8]. Although the locations of local maxima and minima were not much different between 24 and 48 h cultures, the overall reflectivity tended to decrease as the incubation time reduced from 48 to 24 h. In the visible spectral range from 400 to 650 nm especially at around 425 and 460 nm, there were pronounced local minimum and maximum features. In the red spectral range from 650 to 700 nm, reflectance responses increased sharply. Finally, in the near-infrared range from 700 to 900 nm, no prominent features were noticeable except that 24 h cultures on blood agar showed a weak absorption at around 760 nm. This weak absorption feature disappeared in the mean spectra of the 48 h cultures nonetheless.

123

42

S. C. Yoon et al.

(a) 45

(a)


40 35 30


48h NonCampy 24h NonCampy 48h Campy 24h Campy 48h BA 24h BA

25 20 15 10 5 0 400

450

500

550

600

650

700

750

800

850

1

0.9 0.8 0.7 0.6 0.5 0.4 48h NonCampy 24h NonCampy 48h Campy 24h Campy 48h BA 24h BA

0.3 0.2 0.1 0 400

900

450

500

550

Wavelength (nm)

(b)

45 48h NonCampy 24h NonCampy 48h Campy 24h Campy 48h Cefex 24h Cefex


40 35 30


(b)

25 20 15 10 5 0 400

450

500

550

600

650

700

600

650

700

750

800

850

900

Wavelength (nm)

750

800

850

1

0.9 0.8 0.7 0.6 0.5 0.4

0.2 0.1 0 400

900

48h NonCampy 24h NonCampy 48h Campy 24h Campy 48h Cefex 24h Cefex

0.3

450

500

600

650

700

750

800

850

900

Wavelength (nm)

Wavelength (nm)

Fig. 6 Mean reflectance spectra of Campylobacter and non-Campylobacter spots grown on different agar media: a Blood agar: continuum-removed mean spectra of 24 and 48 h cultures. b Cefex: continuum-removed mean spectra of 24 and 48 h cultures

550

Fig. 7 Continuum-removed mean reflectance spectra of Campylobacter, non-Campylobacter spots grown on different agar media: a Blood agar: Continuum-removed mean spectra of 24 and 48 h cultures and b Cefex: Continuum-removed mean spectra of 24 and 48 h cultures

Continuum-removal analysis Figure 7 shows the continuum-removed mean reflectance spectra of Campylobacter and non-Campylobacter cultures (24 and 48 h) in two different agar media (blood agar and Cefex). The spectra shown in Fig. 7 were obtained by continuum removal directly applied to the mean reflectance spectra shown in Fig. 6. The most prominent characteristic observed from the continuum-removed spectra was the difference in slopes between Campylobacter and nonCampylobacter spectra of blood agar cultures. Specifically, in the range between approximately 420 and 465 nm of the spectra of the blood agar cultures, the slope of the nonCampylobacter class was positive whereas that of the Campylobacter class was negative (Fig. 7a). The 1-D line plots in Fig. 8 were obtained by sub-sampling the

123


1 0.9 0.8 0.7 0.6

Positive slope

0.5

Negative slope

0.4 NonCampy Campy

0.3 0.2 400 426

458

500

550

600

650

700

Wavelength (nm)

Fig. 8 Slope differences revealed in continuum-removed spectra of blood agar cultures (24 h)


43

(a) 3000 Campy NonCampy

2500 threshold (0.92)

Frequency

2000

1500

1000

500

0 0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

Band Ratio (458 nm / 426 nm)

(b) 1000 Campy NonCampy

900 800

threshold (1.4)

700

Frequency

continuum-removed mean reflectance spectra of 24 h blood agar cultures, shown in Fig. 7a. The sub-sampling was done at four wavelengths (401, 426, 458 and 701 nm). These plots clearly highlight the slope difference between the non-Campylobacter and the Campylobacter cultures. From this observation, it was straightforward to adopt a band ratio technique using two bands. The bands for the band ratio were chosen from wavelengths corresponding to local extrema of the non-Campylobacter spectra. The extrema were at 426 nm (k1, minimum) and 458 nm (k2, maximum). The same effect, however, was not observed from the Cefex-based cultures. Nonetheless, we applied the same band selection strategy to the case of Cefex 24 h cultures where the local extrema of the non-Campylobacter continuum-removed spectra were observed at 423 nm (minimum) and 461 nm (maximum). Table 2 shows the results of the band depth and ratio measurements at a few different wavelengths. Overall, the band depth decreased as the incubation time increased. The decreased band depth meant that the more light was reflected as the incubation time increased. Note that classification directly utilizing the band depth index was not pursued in this study. As aforementioned, the band ratio was used in this study. In the case of 24 h blood agar cultures, the average band ratio value was 0.7411 (a negative slope) for the Campylobacter class and 1.1275 (a positive slope) for the non-Campylobacter class. The ratio data were thresholded by a single value for classifying the data into two groups. The mean of 0.7411 and 1.1275 was approximately 0.93 that could be used as a threshold value for classifying the band-ratio data. Obviously, we did not know whether the mean value was optimal. To find the optimal threshold value, we used a histogram analysis. Figure 9 shows the histograms of the band ratio data. The histograms of Campylobacter and non-Campylobacter band-ratio data in 24 h blood agar showed unimodal distributions without much overlap. Thus, from Fig. 9a, a threshold value (0.92) dividing two distributions optimally was selected. The selected threshold value (0.92) was very close to the mean value (0.93). The case of Cefex cultures was not simple because of the bimodal distribution of the non-Campylobacter cultures as shown in Fig. 9b.

600 500 400 300 200 100 0

0

1

2

3

4

5

6

7

8

Band Ratio (461 nm / 423 nm)

Fig. 9 Histogram of band ratio data. a Blood agar: 24 h cultures. b Cefex: 24 h cultures

Graphically, this bimodality indicates the inefficiency of the use of band ratio in the case of Cefex cultures. Thus, it was not desirable to use this band-ratio feature for classification of 24-h Cefex culture data. The selection of the threshold at 1.4 for the case of Cefex cultures was based on the constraint to get the better detection rate of Campylobacter. One can mathematically obtain the optimal threshold value minimizing the classification errors or

Table 2 Results of band depth and band ratio measurements Index

Wavelength

Non-Campy. (24 h)

Campy. (24 h)

Non-Campy. (48 h)

Campy. (48 h)

Band depth

426 nm

0.3118

0.6290

0.2565

0.4550

Band ratio

458/426 nm

1.1275

0.7411

1.3444

0.9222

Band depth

423 nm

0.5939

0.5220

0.5803

0.4263

Band ratio

461/423 nm

2.3420

1.0801

2.3742

1.0649

(a) Blood agar (24 h)

(b) Cefex (24 h)

123

44 Table 3 Classification accuracy: single-band thresholding of Cefex 24 h cultures

S. C. Yoon et al.

Ground truth (Pixels) Class

Campylobacter

Non-Campy.

Cefex

Campylobacter

42,610 (92.39%)

11,062 (28.91%)

0 (0%)

Non-Campy.

404 (0.88%)

27,040 (70.68%)

0 (0%)

Cefex

3,108 (6.74%)

156 (0.41%)

20,454 (100%)

46,122

38,258

(a) Confusion matrix

Total (Pixels) Class

Commission (%)

Omission (%)

Commission (Pixels)

20,454 Omission (Pixels)

(b) Commission and omission errors Campylobacter Non-Campy. Cefex

20.61

7.61

11,062/53,672

3,512/46,122

1.47

29.32

404/27,444

11,218/38,258

3,264/23,718

0/20,454

13.76

theoretically use a receiver operating characteristic (ROC) curve which is a graphical plot of the true positive rate versus the false positive rate for a binary classifier system as a function of its discrimination threshold. However, we found after experiments with different thresholds around heuristically chosen values that the selection of the true optimal threshold using a theoretic optimization framework such as the ROC curve and the Bayesian classifier was not necessary. We determined the best threshold values after these preliminary experiments.

0

Campy. Non-Campy. Cefex

Single band thresholding algorithm (Cefex 24 h and BA 24 h) The parameters for the single-band thresholding algorithm applied to the 24-h Cefex culture images at 501 nm were as follows: T1 = 3 and T2 = 7 (T1 = 3 and T2 = 10 were recommended for 48 h cultures). The overall classification accuracy was 85.95% whereas the overall classification accuracy of the 48 h Cefex cultures was 96.81% [8]. Table 3 summarizes the performance of the single-band thresholding algorithm. The classification accuracy of the Campylobacter class alone was 92.39% with 20.61% of the commission error (i.e. false positive) rate. The classification accuracy of the non-Campylobacter class was 70.68% with 1.47% of the commission error rate. The classification accuracy of the Cefex agar class was 100% with 13.67% of the commission error rate. Figure 10 shows the classification result of the single-band thresholding algorithm applied to the 501-nm band of the Cefex 24-culture images. Fifteen out of 49 non-Campylobacter spots were missclassified as Campylobacter class (31%). Two out of 75 Campylobacter spots were miss-classified as non-Campylobacter (3%). Although small portions of three Campylobacter spots (triangles in Fig. 10) were incorrectly classified, the majority of the three spots were correctly classified. Overall, 107 out of all 124 spots (86%) was

123

Fig. 10 Classification of Cefex 24 h cultures by single-band thresholding at 501 nm. The markers indicate the spots that were classified incorrectly (box), partially incorrectly (Campy: triangle and nonCampy: diamond). A circle encloses a cross-contaminated spot

correctly (at least by majority) classified. All cross-contaminated cultures (circled in Fig. 10) were correctly classified. The parameters of the single-band thresholding algorithm applied to the 24 h blood agar (BA) culture images at 503 nm were as follows: T1 = 2 and T2 = 5 (T1 = 3 and T2 = 7 were recommended for 48-h cultures). The overall classification accuracy was 94.54% whereas the overall classification accuracy of the 48-h BA cultures was 98.07% [8]. Table 4 summarizes the performance. The classification accuracy of the Campylobacter class alone was 99.57% with 9.31% of the commission error rate. The classification accuracy of the non-Campylobacter class was 85.91% with 0.68% of the commission error rate. The classification accuracy of the BA class was 100% with 0% of the commission error rate. Figure 11 shows the classification result of the single-band thresholding algorithm applied to the 501-nm band of the BA 24-h culture images.

Detection of Campylobacter colonies Table 4 Classification accuracy: single-band thresholding of blood agar 24 h cultures

45


Campylobacter

Non-Campy.

Blood agar

Campylobacter

183,478 (99.57%)

18,836 (14.09%)

0 (0%)

Non-Campy.

784 (0.42%)

114,853 (85.91%)

0 (0%)

Blood agar

0 (0%)

0 (0%)

41,386 (100%)


Total (Pixels) Class

184,262 Commission (%)

133,689 Omission (%)

41,386

Commission (Pixels)

Omission (Pixels)

(b) Commission and omission errors Campylobacter

9.31

0.43

18,836/202,314

784/184,262

Non-Campy.

0.68

14.09

784/115,637

18,836/133,689

Blood agar

0

0/41,386

0/41,386

Campy. Non-Campy. Blood Agar

Fig. 11 Classification of blood agar 24 h cultures by single-band thresholding at 503 nm. The markers indicate the spots that were classified incorrectly (box), partially incorrectly (non-Campy: diamond). A circle encloses a cross-contaminated spot

Three out of 60 non-Campylobacter spots were missclassified as the Campylobacter class (5%). All 109 Campylobacter spots were correctly classified (100%). Eight non-Campylobacter spots (enclosed with diamonds) had ring patterns but their inner regions were incorrectly classified to Campylobacter. Overall, 158 out of all 169 spots (93%) was correctly (at least by majority) classified. Compared to the Cefex agar (85.95% of pixels), the blood agar (94.54% of pixels) was the better agar medium in differentiating Campylobacter colony spots from nonCampylobacter contaminants with 4.5%. All four crosscontaminated Campylobacter spots (circled in Fig. 11) were correctly classified. Spectral feature fitting classification (BA-24 h) The spectral feature fitting (SFF) classification algorithm was applied to the entire spectral range (193 bands from

0


Fig. 12 SFF-maximum likelihood classification result (BA 24 h). The markers indicate the spots that were classified almost completely incorrectly (box), partially incorrectly (Campy: triangle and nonCampy: diamond). A circle encloses a cross-contaminated spot

400 to 900 nm) of the blood agar 24-h cultures only. The background outside of the Petri dishes was suppressed by the binary mask. The continuum removal for SFF was applied on the fly. Figure 12 shows the maximum-likelihood classification result from a fit image (Scale/RMS). Overall classification accuracy was 90.43%. Table 5 summarizes the performance. Qualitatively, a non-Campylobacter organism (No. 13, Acinetobacter baumannii) was miss-classified twice in the first experiment (the first column in the mosaic). The rest of the misclassification errors partially due to freshness differences of the agar media were from the images in the third and fourth column plates. In the future, there is a need to control the agar freshness variable by using agars with the same storage period or also to study effects of agar freshness on spectral imaging. Four out of 60 non-Campylobacter spots were misclassified either entirely or partially as the Campylobacter class (7%).

123

46 Table 5 Classification accuracy: SFF-ML classification of blood agar 24 h cultures

S. C. Yoon et al.


Campylobacter

Non-Campy.

Blood agar

Campylobacter

154,791 (84.01%)

4,914 (3.68%)

0 (0%)

Non-Campy.

18,629 (10.11%)

128,771 (96.32%)

0 (0%)

Blood agar

10,842 (5.88%)

4 (0.00%)

41,386 (100%)

184,262

133,689


Total Class

Commission (%)

Omission (%)

Commission (Pixels)

41,386 Omission (Pixels)


3.08

15.99

4,914/159,705

29,471/184,262

Non-Campy.

12.64

3.68

18,629/147,400

4,918/133,689

Blood agar

20.77

0

10,846/52,232

0/41,386

Fig. 13 Performance examples of post-processing (Step 4 of the developed classification algorithm) for merging mixed classification results having ring patterns (middle row spots). a Before post-processing. b After post-processing

Fifteen out of 109 Campylobacter spots were misclassified entirely or partially as the Campylobacter class either (14%). Some spots showed salt-and-pepper classification errors (see boxed spots in the third column plates). In this study, we did not attempt to remove the salt-and-pepper classification errors. Overall, 152 out of all 169 spots (90%) was correctly classified. All four cross-contaminated Campylobacter spots (circled in Fig. 12) were correctly classified. Results of developed band ratio algorithm (24 h cultures on blood agar) The overall classification accuracy of the developed band ratio algorithm on the 24-h blood agar-based cultures was 99.38% (the hyperspectral continuum removal case) and 97.21% (the multispectral continuum removal case). Among the techniques studied in this paper, the band ratio algorithm showed the highest classification accuracy both qualitatively and quantitatively. In the hyperspectral

123

continuum removal (CR) case, two bands for band ratio were directly obtained from the continuum removed spectral bands. However, in the multispectral continuum removal (CR) case, note that four bands were first obtained from reflectance hyperspectral bands at 401 nm, 426 nm (k1), 458 nm (k2), and 701 nm without continuum removal, and then the continuum-removal was applied to these four reflectance spectral bands. The band image at 503 nm was used for segmentation of locations of blood agar, grown colonies and plates (Step 1 and 2). In Step 3, the band ratio data were classified by a single threshold value. Figure 13 shows performance examples of the post-processing technique (Step 4). The technique successfully identified round objects with ring patterns and converted the class of the rings to the majority inner class (non-Campylobacter here). Tables 6 and 7 summarize the performance of the developed band ratio algorithm in both hyperspectral and multispectral imaging modes. The classification accuracy of the Campylobacter class was 99.84% (hyperspectral CR) and 95.31% (multispectral CR). The classification

Detection of Campylobacter colonies Table 6 Classification accuracy: band ratio algorithm using hyperspectral continuumremoval of blood agar 24 h cultures

47


Campylobacter

Non-Campy.

Blood agar

Campylobacter

183,032 (99.33%)

1,007 (0.75%)

0 (0%)

Non-Campy.

1,230 (0.67%)

132,682 (99.25%)

0 (0%)

Blood agar

0 (0.0%)

0 (0%)

41,386 (100%)


Total Class


133,689

41,386

Omission (%)

Commission (Pixels)

Omission (Pixels)

(b) Commission and omission errors

Table 7 Classification accuracy: band ratio algorithm using multispectral continuumremoval of blood agar 24 h cultures

Campylobacter

0.55

0.67

1,007/184,039

1,230/184,262

Non-Campy.

0.92

0.75

1,230/133,912

1,007/133,689

Blood agar

0.00

0.00

0/41,386

0/41,386


Campylobacter

Non-Campy.

Blood agar

Campylobacter

175,618 (95.31%)

1,394 (1.04%)

0 (0%)

Non-Campy.

8,644 (4.69%)

132,295 (98.96%)

0 (0%)

Blood agar

0 (0%)

0 (0%)

41,386 (100%)


Total Class


133,689

41,386

Omission (%)

Commission (Pixels)

Omission (Pixels)


0.79

4.69

1,394/177,012

8,644/184,262

Non-Campy.

6.13

1.04

8,644/140,939

1,394/133,689

Blood agar

0

0

0/41,386

0/41,386

accuracy of the non-Campylobacter class was 99.25% (hyperspectral CR) and 98.96% (multispectral CR). In the case of the hyperspectral CR mode, both commission and omission errors were \1%. In the case of the multispectral CR mode, the commission error of the Campylobacter class was still \1% but that of the non-Campylobacter class was 6.13%. The omission rate of the Campylobacter class also increased to 4.69% in the case of the multispectral mode. The results suggest that the developed band ratio technique can detect Campylobacter species and nonCampylobacter contaminants with 97–99% of accuracy in 24 h of incubation. Figure 14a shows the classification result images of the developed band ratio algorithm in the hyperspectral CR mode. All 169 spots were correctly classified either entirely or at least by vast majority (100%). The post-processing (Step 4) correctly identified and successfully converted the class of 58 non-Campylobacter spots out of 60. Only nonCampylobacter spots had the thin outer ring patterns in their classification results (see Fig. 13 for typical sizes of

rings). Thus, the post-processing technique did not affect the classification results of Campylobacter spots. Two nonCampylobacter spots (enclosed with diamonds in Fig. 14) were still unchanged after the post-processing. Four Campylobacter spots (enclosed with triangles in the third column) showed a small amount of non-Campylobacter pixels. All four cross-contaminated Campylobacter spots (circled) were correctly classified. Figure 14b shows the classification result images of the developed band ratio algorithm in the multispectral CR mode. As in the hyperspectral CR mode, all 169 spots were correctly classified either entirely or at least by vast majority (100%). The performance of the post-processing (Step 4) was almost the same between the hyperspectral CR mode (97% of success rate) and the multispectral CR mode (95% of success rate). The lower performance (2–3% in overall accuracy) of the multispectral CR modes can be explained by salt-and-pepper classification errors observed in the multispectral CR mode results. The salt-and-pepper classification errors were more pronounced in the spots

123

48

S. C. Yoon et al.

(a)


(b)


thresholding, the spectral feature fitting, and the continuum-removed band ratio were used as classification methods for Campylobacter detection. The single-band thresholding algorithm showed 95% classification accuracy whereas the spectral feature fitting method and the continuum-removed band ratio method showed 90 and 97–99% classification accuracy, respectively. Blood agar was the better culture medium than Campy-Cefex agar in terms of Campylobacter detection accuracy. The experimental results suggest the developed band ratio algorithm can detect Campylobacter species and non-Campylobacter contaminants with up to 99% of accuracy in 24 h of incubation. The developed imaging protocol is applicable to direct spread plating techniques using chicken carcass rinses or other pathogen detection studies.

References

Fig. 14 Blood agar 24 h cultures: classification results of the developed band ratio algorithm using two bands obtained from continuum removal of (a) hyperspectral bands (193 bands) and (b) multispectral bands (401, 426, 458, and 701 nm) (a Hyperspectral image classification result. b Multispectral image classification result). The markers indicate the spots that were classified incorrectly (box), partially incorrectly (Campy: triangle and non-Campy: diamond). A circle encloses a cross-contaminated spot

with markers. All four cross-contaminated Campylobacter spots (circled) were correctly classified.

Conclusion In this paper, we attempted to develop a hyperspectral image processing algorithm for detecting colonies of Campylobacter species and non-Campylobacter contaminants incubated for 24 h. The continuum-removed mean reflectance spectra of blood agar-based cultures showed prominent difference in slopes between 426 and 458 nm. The band ratio algorithm using two continuum-removed spectral bands at 426 and 458 nm was a key detection technique proposed in this paper. The single-band

123

1. W. Jacobs-Reitsma, U. Lyhs, J. Wagenaar, in Campylobacter, ed. by I. Nachamkin, C.M. Szymanski, M.J. Blaser (ASM Press, Washington, DC, 2008), p. 627 2. N.J. Stern, J.E. Line, H.C. Chen, Compendium of Methods for the Microbiological Examination of Foods (American Public Health Association, Washington, DC, 2001), p. 301 3. N.J. Stern, J.E. Line, The Microbiological Safety and Quality of Food, vol. II (ASPEN Publishers, Gaithersburg, MD, 2000), p. 1040 4. National Advisory Committee on Microbiological Criteria for Foods (NACMCF), J. Food Prot. 70, 241 (2007) 5. J.E. Line, J. Food Prot. 64, 1711 (2001) 6. D.J.M. Mouwen, M.J.B.M. Weijtens, R. Capita, C. AlonsoCalleja, M. Prieto, Appl. Environ. Microbiol. 71, 4318 (2005) 7. D. Naumann, Encyclopedia of Analytical Chemistry (Wiley, Oxford, UK, 2000), p. 102 8. S.C. Yoon, K.C. Lawrence, G.R. Siragusa, J.E. Line, B. Park, P.W. Feldner, Trans. ASABE 52, 651 (2009) 9. H. Yao, Z. Hruska, R. Kincaid, R.L. Brown, T.E. Cleveland, Sens. Instrum. Food Qual. Safety 2, 215 (2008) 10. Food Safety and Inspection Service (FSIS), Detection and enumeration method for Campylobacter jejuni/coli from poultry rinses and sponge samples, available at http://www.fsis.usda. gov/PDF/Baseline_Campylobacter_Method.pdf (2006) 11. International Organization for Standardization (ISO), Microbiology of food and animal feeding stuffs—Horizontal method for detection and enumeration of Campylobacter spp.—Part 1: detection method (ISO 10272-1, Geneva, Switzerland, 2006) 12. G.R. Siragusa, J.E. Line, L.L. Brooks, T. Hutchinson, J.D. Laster, R.O. Apple, J. Food Prot. 67, 901 (2004) 13. N.J. Stern, B. Wojton, K. Kwiatek, J. Food Prot. 55, 515 (1992) 14. W.H. Press, S.A. Teukolsky, W.T. Vettering, B.P. Flannery, Numerical Recipes in C??: The Art of Scientific Computing, 2nd edn. (Cambridge University Press, New York, 2002) 15. K.C. Lawrence, B. Park, W.R. Windham, C. Mao, Trans. ASAE 46, 513 (2003) 16. R.N. Clark, T.L. Roush, J. Geophy. Res. 89, 6329 (1984) 17. P.M. Mather, Computer Processing of Remotely-Sensed Images: An Introduction, 3rd edn. (Wiley, New York, 2004) 18. ITT Visual Information Solutions, in ENVI Tutorial, available at http://www.ittvis.com/portals/0/tutorials/envi/Mapping_Methods. pdf

Detection of Campylobacter colonies 19. F.A. Kruse, Rem. Sens. Environ. 24, 31 (1988) 20. R.F. Kokaly, R.N. Clark, Rem. Sens. Environ. 67, 267 (1999) 21. R.N. Clark, A.J. Gallagher, G.A. Swayze. in Proceedings of the Second Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Workshop. JPL Publication 90-54, Pasadena, pp. 176– 186 (1990)

49 22. R.N. Clark and G.A. Swayze, in The USGS Tricorder Algorithm. Summaries of the Fifth Annual JPL Airborne Earth Science Workshop, JPL Publication 95-1, p. 39 (1995) 23. The MathWorks, Identifying round objects, available at http://www.mathworks.com

123