PROJECT RESULTS

Spectroscopic Identification of Chicken Disease

David A. Prystupa

Spectrum Scientific Inc.

Pinawa, Manitoba  ROE ILO

Table of Contents:

• Background and Objective
• Procedure and Project Activities
• Results and Discussion
• Conclusions

ARDI Project:

#01-516

Project Status:

Completed February, 2005

Infrared spectra of skin, muscle, heart, lung, kidney, liver and spleen tissues from healthy and diseased chickens were obtained.  Principle Component Analysis (PCA) reveals spectral variation with diagnostic value in lung, kidney and liver tissues.  The results do not depend upon the selection of directly affected tissue samples. 

Background and Objective:

The recent emergence of high-profile diseases such as BSE and the Avian Flu have heightened public awareness about potential risks posed by diseased animals in the human food supply.  At the time of writing, human inspectors are the primary means to safeguard the Canadian food supply.  The reliance on human inspectors poses several challenges to the Canadian food inspection system, not the least being the shortage of inspectors.  Human inspectors come at a cost and budgetary constraints ensure that there are never enough inspectors to go around.  Furthermore, consistent standards are difficult to achieve using human inspectors.  Machine vision inspection holds the promise of reduced inspection costs, uniformity, and improved sensitivity for routine inspection tasks, while freeing highly qualified human inspectors for higher value tasks.

Resources are limited for veterinary diagnostic applications, so that a high degree of automation is required.  While it may be practical for an expert to select directly affected tissues for spectral examination for human diagnostics, the cost of such selection is not justified for the veterinary case.  For veterinary examination the inspection system can adopt two approaches.  The first inspection methodology is to examine as much of the carcass as possible to increase the probability that a directly affected region will be included.  The second inspection approach is to identify features that are diagnostic, but are not localized to directly affected body region.  The present study adopts the second approach and tests the hypothesis that disease in one part of the body produces measurable effects at different part of the body that is not directly affected.

Procedure and Project Activities:

A total of 41 chicken carcasses were selected for study from industry sources.  Thirteen samples were healthy and the remaining 28 represented various disease conditions.  The most common disease condition was ascites.  Although samples of other disease conditions were included in the study, many of these chickens had multiple conditions, in many cases including ascites.  Rather than attempting to divide the diseased chickens into several poorly defined classes with poor statistical significance, the diseased chicken samples were pooled as a single class.  Each chicken carcass was dissected and examined by the Provincial Veterinarian.  Sections measuring approximately 1.0 cm² were extracted for spectroscopic examination from skin, muscle, heart, liver, kidney, spleen and lung tissues.  Tissue sections from diseased chickens were selected so as to avoid visibly diseased tissue.

Infrared spectra of each tissue section were obtained at 4.0 cm^-1 resolution with ZnSe internal reflectance crystals and a Nicolet 510P spectrometer.  In each case 100 scans were co-added.  The ZnSe crystals were thoroughly cleaned with acetone and distilled water between each run.  After cleaning, a fresh background was obtained and then the sample tissue was placed on the crystal for the sample run.  Care was taken to ensure good physical contact between the sample and crystal.  Because of the low density and high fluid content of lung tissue, good physical contact was not always achieved for lung tissue samples.

Prior to examination, the tissue samples were stored in petri dishes on ice.  Spectra of all seven tissues from each specimen were examined consecutively and then restored to cold storage, all in less than 30 minutes.  Some of the samples, particularly, skin were visibly desiccated.  The veterinary diagnosis of the tissues was withheld until after the spectra were collected.  This was done to randomize the order of sampling so as to minimize the effects of desiccation and storage time on the overall results.

Results and Discussion:

Typical spectra obtained for each of the chicken tissue types is given in Figure 1.  The total absorbency depends upon the size and geometry of the sample.  All of the raw spectra were scaled using an internal standard prior to further analysis.  The scaled spectra were mean centred and scaled again by the variance at each spectral frequency.  In all cases, the variance was large in the 1640 cm^-1 region, reflecting variability in the water content.  For skin samples, the peak near 1730 cm^-1 is associated with lipid content and is highly variable.  The average spectrum and variance of the skin samples is given in Figure 2.

The mean centred spectra, weighted by the variance, were analyzed by the principle component method.  The PCA takes into account correlations among spectral features that individually may not be distinguished from the random signal fluctuations, but collectively constitute a recognizable signal.  The amount of variance modelled by each principle component (eigenvector) is proportional to the relative magnitude of the corresponding eigenvalue.  The significance of each principle component by tissue type is given in Table 1.  The fractions may not sum to 1.0 due to rounding.  For brevity, only the first five principle components are shown, which represent between 88% and 96% of the experimentally observed variance.  In all cases the first principle component is dominant, accounting for at least 65% of the observed variance.

Table 1.  Fractional of Spectral Variation by Principle Component and Tissue Type

The projections of the mean centred data vectors along each of the principle components gives a reduced dimensionality representation of the original data which describes the variance.  If there are differences between the spectra from tissues of healthy and diseased chickens, then the projections of the spectra will be distributed differently in the principle component space.

In the simplest case, the diseased and healthy tissue spectra might be distinguished by differences along a single principle component. 

The average scores for samples originating from healthy and unhealthy chickens were calculated for each tissue type.  The difference between the healthy and unhealthy averages was tested for statistical significance using the t-Test, double sided distribution.  Results were considered significant if the score is less than 0.05, which corresponds to the 95% confidence level.

Scores for the first principle component for lung and kidney tissues were found to be a statistically significant indicator of the health status of the chicken.  As shown in Table 1, the first principle component models 85% and 68% of the spectral variability for lung and kidney tissues, respectively, and thus can serve as a diagnostic indicator for industrial applications.  For heart tissue, the fifth principle component scores were found to be a statistically significant indicator of the health status of the chicken. Because the fifth principle component for heart tissue models only 2% of the spectral variability, this indicator is not expected to be of use for industrial applications.

Table 2.  t-Test by Principle Component and Tissue Type

Excepting the lung, kidney and heart results noted above, the average scores were not significantly different for healthy and diseased tissues.  However, identical averages do not imply identical distributions.  Visual inspection of the data revealed that the data points from healthy subjects had positive and negative scores with low magnitude and the data points from unhealthy subjects had positive and negative scores with larger magnitudes.  In these cases, the absolute values of the principle component scores were calculated and used as the basis of comparison.  The results of this comparison are summarized in Table 3.  The absolute value of the second principle component scores for lung tissue, and the absolute values of the second and fourth principle component scores for liver tissue are of diagnostic value.  The absolute value of the second principle component scores for skin tissue may also prove to be of diagnostic value, but the results of the present study fall just short of supporting this conclusion.

Table 3.  t-Test by Principle Component and Tissue Type

No statistically significant correlation between health state and principle component scores was found for muscle and spleen tissues.  The result that the composition of muscle tissue is not significantly affected by disease conditions may be mildly reassuring for human consumers.  The results of this study do not, however, preclude the possibility that harmful substances may exist in muscle tissue at concentrations below the detection threshold for this method.

The presence of spectral features indicative of unhealthy conditions in multiple tissues clearly indicates that the underlying diseases are systemic rather than localized.  At least for the most common disease condition, ascites, the consequence is that diagnosis can be made on the basis of a small arbitrarily chosen sample of an appropriate tissue type.

Because the samples were specifically chosen so as to exclude obviously affected tissues, human inspectors in a processing plant would not be able to arrive at a diagnosis if presented only with the same small tissue samples.  The spectral method is more sensitive.  This is not the same as saying that the spectral method is better than a human inspector.  A human inspector assesses much more information with lower sensitivity to arrive at a diagnosis.  The larger volume of information processed by the human inspector may well more than compensate for the human inspector's lower sensitivity.  This remains an unresolved issue for further study.  There is the possibility that spectral measurements may detect a disease condition at an earlier stage than would be apparent to a human inspector.

An inspection of the spectral distribution of the first principle component for both lung and kidney tissues indicates that the spectral variability is concentrated in the Amide II region near 1550 cm^-1 and in the carboxyl stretching region near 1750 cm^-1.  Further investigation is required to determine the underlying causes for this effect.

The results presented in this study are based on a small sample size and are thus preliminary.  Further studies with larger sample sizes are required to establish the range of applicability for the results presented herein.  Based on the current results, further studies should focus on the results holding the greatest promise for industrial application.  Specifically, the first principle component for lung and kidney tissue together with the second principle component for liver and skin tissues hold the greatest promise for industrial application.  Based upon the author's previous experience, the variation in spectral features identified in the current study are likely to be strongly correlated with variations observed using different measurement methods.  This possibility should be investigated further prior to industrial application so that the most effective method is used.

Conclusions:

Statistically significant spectral variation between healthy and diseased chickens was found for lung, kidney and liver tissues.  The results were obtained from tissues that were not directly affected by the disease condition, and thus would not be apparent to a human inspector.  The spectral variation identified may form the basis for a machine vision inspection system.

Acknowledgements:

We are grateful for a grant from the Agri-Food Research and Development Initiative, which provided financial support for this research to Amanda Reinisch, who helped to collect samples and data; to Dr. Mark Swendrowski for examining the chicken carcasses; and to industry sources for providing the samples.

ARDI Home - Contact ARDI