- Email: [email protected]
Near-infrared spectroscopy detects woody breast syndrome in chicken fillets by the markers protein content and degree of water binding
Near-infrared spectroscopy detects woody breast syndrome in chicken fillets by the markers protein content and degree of water binding Jens Petter Wold,∗,1 Ingrid M˚ age,∗ Atle Løvland,† Karen Wahlstrøm Sanden,∗ and Ragni Ofstad∗ ∗
Nofima AS, Norwegian Institute for Food and Fisheries Research, Muninbakken 9-13, Breivika, NO-9291 Tromsø, Norway; and † Nortura SA, Lørenveien 37, NO-0513 Oslo, Norway sification of normal versus WB on the training set, and 96% correct classification on a test set of 52 fillets. The main reason for why NIRS can successfully discriminate between WB and normal fillets is the methods sensitivity to both protein content and degree of water binding in the muscle, both established markers for WB. The classification model can be based on NIR spectra only, calibration against protein is not needed. The affected muscle tissue associated with the WB syndrome is unevenly distributed in the fillets, and this heterogeneity was characterized by NIRS and NMR. Clear differences in water binding properties were found between the superficial 1 cm layer and the deeper layer at 1 to 2 cm depth. Significant differences in protein estimates by NIRS at different measurement points along the chicken fillets were obtained for WB fillets. The findings suggest how to obtain optimal sampling with NIRS for best possible discrimination between WB and normal breast fillets.
ABSTRACT The muscle syndrome woody breast (WB) impairs quality of chicken fillets and is a challenge to the poultry meat industry. There is a need for online detection of affected fillets for automatic quality sorting in process. Near-infrared spectroscopy (NIRS) is a promising method, and in this study we elucidate the spectral properties of WB versus normal fillets. On a training set of 50 chicken fillets (20 normal, 30 WB), we measured NIR, nuclear magnetic resonance (NMR) T2 relaxation distributions, and crude chemical composition. NIRS could estimate protein in the fillets with an accuracy of ±0.64 percentage points. T2 distributions showed that there was a larger share of free water in WB fillets. This difference in water binding generated a shift and narrowing of the water absorption peak in NIR around 980 nm, quantified by a bound water index (BWI). The correlation between BWI and T2 distributions was 0.78, indicating that NIRS contains information about degree of water binding. Discriminant analysis showed that NIRS obtained 100% correct clas-
Key words: woody breast, chicken fillet, NIR spectroscopy, NMR relaxation, classification 2018 Poultry Science 0:1–11 http://dx.doi.org/10.3382/ps/pey351
INTRODUCTION
broiler chickens (Kuttapan et al., 2016). The situation calls for a rapid method for detection, so that affected fillets can be automatically sorted out from the production line. Such a method can also be used to map incidences in large numbers of fillets, in order to trace the main causes responsible for the syndrome, and to work with improvement strategies. Near-infrared spectroscopy (NIRS) has been demonstrated to work well for rapid and non-destructive detection of WB fillets (Wold et al., 2017). The method was applied online in an industrial process and could classify fillets passing by on a conveyor belt at high speed according to protein content. Woody breast fillets have, on average, lower protein content compared to normal breast muscle (Soglia et al., 2016a). During the study, we got indications that NIR could measure not only protein, but probably also water binding properties in the fillets. It has been reported that WB tissue contains more loosely bound water than normal breast muscle (Soglia et al., 2016a). It is important to know if the spectra contain information about both protein
During the last years, the muscle syndrome woody breast (WB) has become a challenge to the poultry meat industry. Woody breast is a myopathy affecting the muscle tissue in the chicken breast (Pectoralis major), making the muscle appear as pale, hard, and outbulging (Sihvo et al., 2014). Because the appearance of this meat is unpleasant and the functional properties are impaired, WB fillets have to be downgraded and used for less valuable products. With typical incidences of 5 to 10% of WB in markets with heavy broilers, this represents significant economic losses (Gee, 2016; Zanetti et al., 2018). The causes for WB are still not clear, but are most likely multifactorial where an important part is related to the fast growth of modern
C 2018 Poultry Science Association Inc. Received January 31, 2018. Accepted July 16, 2018. 1 Corresponding author: [email protected]
1 Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
2
WOLD ET AL.
and water binding properties, in order to understand and apply the method for WB detection in the best possible way. Soglia et al. (2016a) reported mean crude protein content for WB fillets of 22.6% compared to 24.65% in normal muscle tissue, whereas Wold et al. (2017) found average protein values of 18.4, 18.9, and 23.5% in severe WB, moderate WB, and normal fillets, respectively, in the upper 1 cm surface layer of the fillets. The lacking muscle protein in WB muscle is replaced by mainly moisture, but also higher levels of fat and collagen are found (Soglia et al., 2016b, 2017). Note that the protein content in WB fillets is not evenly distributed. The surface part of the fillet tends to be more affected with low protein content, whereas deeper into the fillet, the protein content can be as in normal fillets. Also muscle texture varies unevenly; the superficial layer is harder than the deeper layer, which again does not differ from normal muscle (Soglia et al., 2017). NIRS is widely used for rapid, online and non-destructive determination of typically fat, protein, and moisture in a variety of food products (Porep et al., 2015), and the method is able to estimate the amount of protein in intact chicken fillets (Cozzolino et al., 1996; Wold et al., 2017). Wold et al. (2017) used a protein calibration based on NIRS to successfully detect WB fillets. Fillets with estimated protein below about 21.9% were classified as WB. It has been reported that the water holding capacity in WB fillets is poorer than in normal fillets, shown by increased drip loss and cooking loss (Mudalal et al., 2015; Tijare et al., 2016). Soglia et al. (2016a) measured water mobility by the use of low field nuclear magnetic resonance (NMR) and reported that there are significantly higher shares of loosely bound water in WB tissue, probably due to muscle fiber degeneration. Water binding properties affect the spectral features of the water absorption peak at 980 nm in the NIR spectrum. Hydrogen bonding between water and other molecules such as proteins causes a spectral shift toward longer wavelengths as well as a peak broadening (Chung et al., 2008). This shift, as well as other spectral properties, has been suggested as a basis for measuring water holding capacity and drip loss in meat by NIRS, but it has been difficult to establish solid calibrations with causal explanations for these applications (Prieto et al., 2009). Within medical diagnostics, the spectral water band shift has been used to detect human breast cancer, because cancerous tissue contains more loosely bound water compared to normal tissue (Chung et al., 2012). Chung et al. (2008) introduced a bound water index (BWI) that quantifies the shift observed in tissue water absorption spectra measured by diffuse NIR spectroscopy, and this index correlated well with the occurrence of malignant breast tissue. The index can also be calculated in NIR images of breast tissue, highlighting regions with loosely bound water indicating tumor areas.
Low-field NMR gives well-founded estimates of water distribution and mobility in meat (Bertram et al., 2002a,b). The spin-spin relaxation time, (T2 ), indicates shares of bound water located within highly organized myofibrillar protein matrix, and more loosely bound extramyofibrillar water. Tasoniero et al. (2017) found that there is a rather low correlation (R = 0.41–0.48) between T2 signals and hardness measured on WB fillets. The objective of this paper is to better understand the spectral properties of chicken muscle tissue affected by the WB syndrome. We focus on the NIR wavelength region 760 to 1,040 nm, because this region has been proven well suited to distinguish WB from normal fillets. NIR spectra in interaction mode was recorded from 50 chicken fillets, normal and WB. The NIR spectra were interpreted based on the T2 relaxation distributions obtained by low-field NMR as well as the crude chemical composition. Different approaches for classification of WB fillets based on NIR spectroscopy were tested, including the BWI and prediction of protein content, with the aim of finding the most reliable method. The different approaches were then evaluated on a test set of 52 new fillets. The effects of sample heterogeneity and spectroscopic sampling are discussed.
MATERIAL AND METHODS Chicken Fillets Calibration Set A total of 50 skin and boneless breast fillets (Musculus pectoralis major) were sampled directly from a Norwegian processing line approximately 3 h after CO2 stunning, bleeding, and slaughter of the birds. The birds were as hatched, 32 to 34 d old, and of the strain Ross 308. Average standard broiler live weight at the processing plant is 2.1 kg. The samples were classified by an experienced veterinarian based on visual inspection and palpation of consistency (normal, hard, and very hard). Breast fillets with hard consistency and limited distribution of very hard parts were classified as moderate WB, whereas fillets with extensive areas of very hard consistency were considered severe WB. Woody breast fillets were pale in color, but no chicken breasts had significant amounts of serous fluid at the surface described from some studies in markets with higher slaughter weights. A total of 24 normal fillets and 26 WB fillets were taken out. The WB fillets were classified in 13 moderate and 13 severe WB. Woody breast fillets were from 2 different flocks, whereas the normal fillets were from the same flock. The fillets were packed in plastic bags and stored overnight at 4◦ C, before they were measured with NIR and NMR, and the fat and protein contents were determined. NIR and NMR were measured about 24 h after slaughter. Tasoniero et al. (2017) reported that there were no significant changes in NMR T2 parameters, nor in measured hardness, in normal and WB fillets from 10 to 24 h post mortem.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
NIR SPECTROSCOPY DETECTS WOODY BREAST
Figure 1. Sampling regime of chicken fillets. Circular region 2 was used for NIR and NMR measurements for calibration set. For the test set, region 2 was used for NMR, and NIR was measured in regions number 1 to 4. The excised cylinder was used for NMR measurements. The fillet part above the black horizontal line was used for determination of crude chemical composition of the samples in calibration set. The 2 yellow rectangular regions indicate size and shape of the illuminated area during an NIR measurement. The absorption spectrum was measured between the 2 illuminated regions, in this case in the center of the circular region 2.
Test Set A total of 52 skin and boneless breast fillets were taken from the production line in the same way as the calibration set, but 4 mo later. The WB fillets were from several flocks, whereas the normal fillets were from 2 different flocks. Twenty normal fillets and 32 WB fillets were taken out. The WB fillets were classified in 20 moderate and 12 severe WB. The fillets were packed in plastic bags and stored overnight at 4◦ C before measurements with NIR. NMR was measured on 5 samples from each group to determine T2 relaxation distributions in superficial and deep layers of the fillets. Water content was determined in the same samples.
NIR Measurements Each fillet was measured by an NIR instrument collecting spectra in interaction mode. It was originally designed to measure fat in the muscle of live salmon (Folkestad et al., 2008). Two halogen light sources of 50 W each illuminate the sample in 2 rectangular regions of 5 mm × 20 mm size (indicated in Figure 1). Distance between the 2 illuminated regions is 10 mm. The system collects the light that has penetrated down into the sample and comes up again in a small area of 4 × 4 mm between the 2 illuminated rectangles. The collected light has then traveled from an approximate square area of typically 2 cm × 2 cm, and a depth in the muscle of approximately 10 to 15 mm is probed (Wold,
3
2015). The receiving part of the instrument is in physical contact with the sample in order to reduce stray light. The mode of measurement is called interaction and makes it possible to measure through the skin of a salmon to sample the muscle below. It is therefore also possible to measure rather deep into chicken fillet to probe chemical properties. The system measures at 15 wavelengths in the region 760 to 1,040 nm, and spectral bandwidth is 20 nm. The system does about 70 measurements per second and these were averaged over the total measurement time of 1 s per sample. The system is thoroughly described in O’Farrell et al. (2011). Each intact fillet in the calibration set was measured on the skin side at the location indicated in Figure 1. Temperature of fillets was 4◦ C. A spectrum of pure water at 4◦ C was also measured. This was done by filling a square glass beaker with 1 L of water. At the bottom of the beaker, there was a polished aluminium plate working as a mirror for the NIR radiation, resulting in a transflection mode spectrum. The measurement principle, interaction, is the same as the one used for detection of WB fillets by Wold et al. (2017). That was however an imaging system, capturing the entire width of a 80 cm wide conveyor belt, measuring on fillets in motion without physical contact, whereas the system used in this study measured only a smaller square-like sample area of samples in the steady state. This point system captures spectra with higher signal to noise ratio and penetrates a few mm deeper into the samples.
NMR Relaxation Measurements For the calibration set, a cylindrical sample of diameter 8 mm and height approximately 20 mm was excised from each fillet at the center of where the NIR measurement was performed (Figure 1). The entire cylinder was weighted and placed in a sealed Teflon container, which was inserted in the NMR probe. Before measurement, the samples were thermostated to 25◦ C for 30 min. For the calibration set, the relaxation measurements were performed with a Maran Ultra Resonance 0.5 tesla (Oxford Instruments, Oxfordshire, UK). T2 was measured using the Carr-Purcell-Meiboom-Gill pulse sequence (CPMG) (Carr and Purcell, 1954; Meiboom and Gill, 1958). The T2 measurements were done with a tau-value of 150 μs. Data from 8,000 echoes were collected during 16 scans. The CPMG data were analyzed using WinDXP software (Resonance instruments, Oxfordshire, UK). From the test set, a cylindrical sample of diameter 8 mm and height 20 mm was excised from 5 fillets from each class (normal, moderate, and severe WB) from the corresponding site as for the calibration set (indicated region 2 in Figure 1). The cylinders were cut in 2 equally sized parts, an upper and a lower sub-sample, representing the superficial (0 to 1 cm) and deeper layer (1 to 2 cm) of the fillets. The pieces were weighted and
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
4
WOLD ET AL.
placed in a sealed Teflon container, which was inserted in the NMR probe. Before measurement, the samples were thermostated to 25◦ C. For each subsample, we determined the T2 relaxations distribution. The CPMG analysis of the test set was performed by a 20 MHz NMR instrument (Antek Instruments Inc., Houston, TX). The measurements were done with a tau-value of 150 μs. Data from 8,000 echoes were collected during 4 scans. Another NMR system was used for the test set samples due to technical problems with the one used for the calibrations set.
Crude Chemical Composition Calibrations Set The upper part of the fillet, where the NIR and NMR measurements were performed, was used for determination of crude protein and fat (indicated in Figure 1). The inner fillet (Pectoralis minor) was removed, and the remaining outer fillet (Pectoralis major) was used for analyses. This sample was homogenized and sent to external lab for analysis. Fat was determined based on extraction with diethyl ether and petroleum ether according to the Schmid-BondzynskiRatslaff method (NMKL 131, 1989). Crude proteins were determined according to the Kjeldahl method by the use of copper sulfate as catalyst (NMKL 6, 2003). Test set Water content in upper and lower part of excised subsamples (cylinders) from test set was determined by weighing, drying at 105◦ C for 24 h, and then weighting again (NMKL 23, 1991).
Data Analysis Spectral Pre-processing The NIR spectra were linearized using the logarithm of the inverse of the interactance spectrum (T), log10(1/T). To reduce light scattering effects, the log10(1/T) spectra were normalized by standard normal variate (SNV) (Barnes et al., 1989). That is, for each spectrum the mean value was subtracted and the spectrum was then divided by the standard deviation of the spectrum. Bound Water Index BWI for each sample was calculated based on the NIR spectrum and the spectrum of pure water, both SNV corrected. The wavelength region used for this calculation was 960 nm to 1,040 nm because it covers the water absorption peak. BWI was calculated as BWI =
i
|Sfillet (λi ) − Swater (λi )|
where Sfillet (λi ) is the spectral value at wavelength λi, and Swater (λi ) is the corresponding spectral value in the water spectrum. A large spectral shift from pure water results in a high BWI and vice versa. Protein Calibration An NIR calibration for protein was made by partial least squares regression (PLSR) (Martens and Næs, 1993). The calibration model was
based on one spectrum from each fillet. Performance of the model was estimated by the root mean squared error (RMSE) and the correlation (R) between measured and estimated protein, where the estimated values were obtained by full cross-validation. The cross-validated protein estimates were then used for classification of WB fillets. This calibration was also used to estimate protein values in the test set. Linear Discriminant Analysis Linear discriminant analysis (LDA) (Duda and Hart, 1973) was used to test how well different measures could discriminate WB from normal fillets. Whole NIR spectra, the BWI, and the NIR estimated protein were tested separately. NMR T2 relaxation distributions were also tested as a discriminator. Because the variables in both the NIR spectra and NMR profiles were highly correlated, we used the score values from a principal component analysis (PCA) (Marten and Næs, 1993) of these data sets. These score values are orthogonal to each other and well suited as input variables in LDA. We used the Mahalanobis distance as metric in the discriminant function. The functions were validated by full cross validation on the training set and then evaluated on the independent test set. One-way analysis of variance was performed to analyze group differences between normal and WB. If the P-value was less than 0.05 the differences were considered significant. The software The Unscrambler ver. 9.8 (CAMO Software AS, Oslo, Norway) was used for regression analysis. Calculation of BWI, LDA, spectral pre-processing and significance testing were carried out by the use of MATLAB version 7.10 (The MathWorks Inc., Natick, MA).
RESULTS AND DISCUSSION Crude Chemical Composition Mean values of fat and protein content for the 3 groups, normal, moderate WB, and severe WB, in the calibration set are shown in Figure 2. There were no significant differences in fat content between the 3 groups, whereas the means for protein were significantly different for all groups. The mean values for protein were comparable with those obtained by for instance Soglia et al., (2016a), who obtained 24.65 and 21.60% protein for normal and WB, respectively. The protein values for moderate and severe WB were higher than those obtained for moderate and severe WB by Wold et al. (2017) (18.9 and 18.4%, respectively), which is reasonable, because those values were measured in the 1 cm superficial layer of the fillets. Wold et al. (2017) found that a protein content of 21.9%, estimated by NIR, was a suitable limit to separate between normal and WB fillets. In this study, there were 6 fillets from the moderate WB group with protein content above this threshold. This result can be explained by the sampling. Because the upper 2 cm
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
NIR SPECTROSCOPY DETECTS WOODY BREAST
5
Figure 2. Sample distribution of chemically measured fat and protein in the calibration set. The horizontal line represents the mean, whereas the colored areas correspond to ±1 standard deviation (light area) and ±1.96 standard deviations (= 95% confidence interval, dark area).
layer was measured, higher protein values will be obtained compared to the previous study when the upper 1 cm sample layer was used. When only upper 1 cm is used, a more clear difference between normal and WB can be expected. This is further discussed below.
Spectral Properties The main visible difference in the NIR spectra was a subtle shift in the absorption peak at about 980 nm (Figure 3). At about 1,000 nm there was an apparent isosbestic point, whereas systematic differences occurred at shorter and longer wavelengths. The main peak at 980 nm originates mainly from water absorption. Relatively small contributions are expected from fat, protein, and myoglobin (Jacques, 2013) for chicken breast muscle in this spectral region. For simplicity, the peak is referred to as the “water peak”. The water peak in spectra from WB fillets was slightly blueshifted toward shorter wavelengths compared to normal fillets. The peak was also more narrow. It was close to the peak of pure liquid water, suggesting that much of the water in WB fillets is free, and very weakly bound to other macromolecules. The BWI quantifies this shift, and it is clear that it varied systematically with the quality of the muscle tissue (Figure 4). Normal fillets had high BWI values (above 0.6), whereas WB fillets had lower values (below 0.6). Low values indicate that the spectral difference from pure water is small. There was no difference in BWI between moderate and severe WB fillets. When compared with the distribution of protein in the same samples (Figure 2), it is clear that the BWI was a better discriminator of normal and WB fillets than the protein content for this particular data set. So although the shift observed in Figure 3 looks subtle to the naked eye, it is rather pronounced for a stable NIR spectroscopic system of high quality. The observed variation is much larger than the noise level in such systems, which gives a
good basis for discrimination between normal and WB fillets. A PCA of the NIR spectra (760 to 1,040 nm) resulted in a first principal component (PC1), which explained 97% of the spectral variation. The score values of PC1 correlated highly with BWI (R = 0.997), which means that PC1 expressed more or less the same shift variation. It should be noted that the calculation of BWI in this case is simplified compared to the original procedure proposed by Chung et al. (2008). Before calculating BWI they corrected the spectra by removing contributions from lipids and hemoglobin. These contributions are probably substantial and can vary a lot in and between human breasts. The spectral variation due to fat and myoglobin in the chicken fillets is probably small enough to be ignored, but this should be investigated further before using BWI alone to classify WB. Fat content is very low and would most likely not affect the spectra much, but the amount of remaining blood in the fillets might vary and could affect the spectral shape and also the calculated BWI. Protein has an absorption peak at about 1,020 nm (not discernable in the spectra). A lower protein content and a corresponding increase in moisture, as the case is for WB muscle, could also affect the observed spectral shift. Another important parameter to consider is sample temperature, because temperature also induces a shift in the water peak. As temperature increases, the fraction of hydrogen-bound water molecules is reduced, causing the peak to increase in intensity, narrow in bandwidth, and blueshift to shorter wavelengths (Libnau et al., 1994). Actually, NIR spectroscopy as measured in this work can be used to determine core temperature in heat treated or chilled food products (Wold, 2015). Temperature changes will therefore directly affect the value of BWI. In a poultry processing plant, the temperature of the fillets should therefore be stable within a few degrees. If this is not the case,
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
6
WOLD ET AL. 2
2
1.5
1.5
Arbitrary unit
Arbitrary unit
1
0.5
1
0.5
0 0
-0.5
-1
800
950 900 850 Wavelength, nm
1000
-0.5 960
980
1000 Wavelength, nm
1020
1040
Figure 3. Left panel: SNV corrected NIR spectra from one selected filet of each class normal (blue), moderate WB (green), severe WB (red). Right panel: zoom-in on water absorption peak with spectrum of liquid water (black) added for comparison.
Figure 4. Sample distributions of derived properties from NIR spectra and NMR T2 relaxation distributions. The horizontal line represents the mean, whereas the colored areas correspond to ±1 standard deviation (light area) and ±1.96 standard deviations (= 95% confidence interval, dark area).
then it should also be possible to correct the spectra for temperature effects before calculation of BWI.
Protein Calibration based on NIR The protein calibration based on NIR spectra from the chicken fillets made by PLSR required 5 latent variables. By cross validation we estimated the root mean square error of cross validation (RMSECV) to 0.64 percentage points and a correlation, R, between estimated and measured protein of 0.89. This result is close to what we previously obtained (RMSECV = 0.57, R = 0.76) (Wold et al., 2017). The higher R in this study was due to a greater protein content range. For classification of WB fillets, Wold et al. (2017) found an optimal decision limit of 21.9%. Fillets with protein estimated under this limit were classified as WB. When the same limit was used in this study, 5
of the 26 WB samples were erroneously classified as normal (Figure 4). These samples were all from the moderate WB group. All normal samples were correctly classified.
NMR Relaxation Figure 5 shows T2 relaxation distributions measured in chicken breast muscle from 1 normal and 2 WB samples from the calibration set, the same samples as used to illustrate NIR spectra in Figure 3. The distributions consist of 2 major components, which are characteristic for meat. The peak between 0.02 and 0.1 s is often referred to as T21 , and the smaller peak in the region 0.18 and 0.50 s is called T22 (Bertram et al, 2002b). T21 indicates the share of water bound in myofibrils, whereas T22 indicates share of more loosely bound water outside the muscle fibers.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
7
NIR SPECTROSCOPY DETECTS WOODY BREAST 0.08 0.07
Amplitude
0.06 0.05 0.04 0.03 0.02 0.01 0 10
10
10
T2 relaxation time, sec Figure 5. NMR T2 relaxation distributions from normal fillet (blue), moderate WB (green), and severe WB (red).
The T2 distribution for the normal chicken muscle in this study resembles similar measurements reported for pork and chicken muscle, a large T21 component and a much smaller T22 component separated by a time region with no signals (Bertram et al., 2002a,b; Li et al., 2014). T2 distributions from WB muscles had less contribution from the T21 component and more from the T22 component. The 2 components were also broader and connected to one continuous curve, indicating that the water binding properties in WB muscle span a range of different states. The distributions clearly illustrate that there is a much larger share of loosely bound water with higher mobility in WB muscle. This corresponds with earlier observations of water relaxation times in WB muscle (Soglia et al., 2016a). Bertram et al. (2002a) observed a third population of water (T2b ) at very short relaxation times (0.001 to 0.005 s), and proposed that the signal originated from water tightly bound to macromolecules. We did also observe signals in this region; however, they were of very low intensity and did not vary systematically with the degree of WB. These signals could possibly vary with for instance changes in protein secondary structure, but we did not study this structure, and it was therefore difficult to interpret the variation in T2b . Because T2 relaxation expresses the water binding properties of water in the meat, it is of interest to elucidate how the NIR spectra and NMR profiles related to each other. This was done by PLSR analysis and showed that there was a correlation (R) between T2 relaxation distributions and BWI of 0.78. Note also that there was a corresponding correlation (R) between NMR T2 and protein content of 0.92. The water binding properties were hence closely related to the amount of protein. The first PC of the T2 distributions explained 77% of the variation, and Figure 4 shows that the scores for this component separated quite well between normal and WB fillets.
Linear Discriminant Analysis In this section, we first used LDA to evaluate different approaches for classification of WB versus normal fillets on the calibration set of 50 samples. The main aim was to find a good classifier that can be used online. It was also interesting to investigate if T2 relaxation distributions could separate well between the 2 groups. We then used the discriminant functions based on NIR data from the calibration set to classify the test set obtained 4 mo later. All the methods performed well on the calibration set; however, the best discrimination was obtained when either the BWI or the full NIR spectra were used for classification (Table 1a). The BWI separated well between normal and WB fillets, which could also be seen from Figure 4. PC1 from the NIR spectra performed equally well, and that is reasonable because these 2 parameters were highly correlated. They both expressed the shift variation in the water peak. The approach of making a protein calibration based on NIR, and then use this as a classifier, was slightly less successful for this data set. The predicted protein values of WB overlapped those of the normal fillets. The fact that the T2 distributions from NMR could distinguish almost perfectly between normal and WB fillets indicates again that water binding properties are distinctly different for the 2 groups. We did also investigate if the different measurements could distinguish between moderate and severe WB fillets (Table 1b). This proved to be more difficult, where NMR and NIR spectra obtained best results with about 76% correct classifications. It should also be noted that the 2 parameters BWI and predicted protein together obtained 76.9% correct classification. The classification results for the test set differed slightly from the calibration set (Table 2). The BWI performed less well with 7 miss-classifications, whereas
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
8
WOLD ET AL. Table 1. Classification results for different data for the calibration set (cross validation) Data type
# PC1
#Correct
#Non-correct
a) Classification of WB versus normal fillets BWI 1 NIR spectra 1 NIR estimated protein 1 Protein 1 NMR 2
50 50 47 44 49
0 0 3 6 1
b) Classification of moderate versus severe WB BWI 1 NIR spectra 2 NIR estimated protein 1 Protein 1 NMR 2
14 20 15 18 20
12 6 11 8 6
% Correct classification 100 100 94 88 98 53.8 76.9 57.7 69.2 76.9
#Correct and #Non-correct indicate the number of correctly and non-correctly classified samples. 1 #PC—Number of principal components used in the model.
Table 2. Classification results for different data for the test set. Data type
# PC1
# Correct
# Non-correct
% correct classification
Classification of WB versus normal fillets BWI 1 NIR spectra 4 NIR estimated protein 1
45 49 50
7 3 2
86.5 94.2 96.1
b) Classification of moderate versus severe WB BWI 1 NIR spectra 1 NIR estimated protein 1
15 22 22
17 10 10
46.8 68.7 68.7
#Correct and #Non-correct indicate the number of correctly and non-correctly classified samples. 1 #PC—Number of principal components used in the model.
NIR predicted protein and NIR spectra performed better with only 3 and 2 miss classifications, respectively. Note that for the NIR spectra, the discriminant function was based on 4 PCs, not only the first PC shown in Figure 4. It means that there was more information in the NIR spectrum than only the water peak shift that was used for the classification. The high number of correct classifications on a dataset obtained 4 mo later indicates that the method is robust, as was also demonstrated by Wold et al. (2017). The WB myopathy appears in different degrees, from moderate to severe. The variation is both in degree within an affected part of the muscle, and in extension of affected areas. There will therefore be a gradual change from normal to severe cases. In such a system there will always be borderline cases between the groups and impossible to obtain 100% correct classification for a number of samples. From the rather limited set of samples in this study, it is clear that both the water binding property and the protein content are good discriminators with regard to WB. Both features can be derived from the NIR spectrum. Therefore, the NIR spectrum is a good discriminator and the spectrum alone can be used successfully as input in a classification model. With respect to an industrial application, there are certainly some advantages of using just the NIR spectrum for classification rather than a protein calibration. Chemical reference
measurements for protein can be avoided, which means reduced costs. On the other hand, for system operators it might be more informative and meaningful to use a system that produces protein and BWI values, rather than a black box classifier of the NIR spectra. In Figure 6, the NIR predicted protein is plotted versus BWI for all samples included in both calibration and test set. Most WB fillets had low scores for both parameters, but there were samples manually classified as WB with high protein and low BWI, or high BWI and low protein. There was also one normal breast with high protein content, but low BWI. The manual classification is based mainly on the hardness of the breasts, and also the share of hard tissue. For samples in the gray zone between normal and WB, the developed hardness does not necessarily correlate well with protein content or BWI. Tasoniero et al. (2017) found rather low correlations in the range 0.41 to 0.48 between NMR T2 parameters and measured hardness for WB fillets. Figure 6 also shows that protein content and BWI do not discriminate well between moderate and severe WB. There can be at least 3 reasons for this. (1) As stated above, the hardness does not necessarily correlate well with the 2 parameters, (2) the NIR spectra were obtained from a limited region of the fillet and did not include distributional information about hardness, and (3) the heterogeneity in depth of the fillet will
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
9
NIR SPECTROSCOPY DETECTS WOODY BREAST 25
24
Predicted protein %
23
22
21
20
19
18 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Bound water index Figure 6. Predicted protein content versus BWI for all samples (calibration and test set). Normal (white), moderate WB (green), severe WB (red). Calibration set: circles, test set: triangles.
affect the optical measurements and might obscure the relation.
Effects of Sample Heterogeneity It is obvious that the morphological changes in the muscle tissue associated with the WB syndrome are unevenly distributed in the fillets. The distribution follows some typical patterns, it varies both across the fillet surface and in depth depending on the severity of the syndrome. This means that the protein content and water binding properties will also vary accordingly. Figure 7 shows average estimated BWI and protein for the 3 groups in the test set. All the fillets were measured with NIR at the points 1 to 4 (Figure 1). There were significant differences between the normal fillets and the WB fillets. However, there were also significant differences within the fillets, in particular for the WB fillets. The BWI and protein contents were significantly lower in the rostral thicker part of the WB fillets (points 1 and 2 vs. point 4), whereas the variation in the normal fillets was less pronounced. This heterogeneity indicates that it is important where on the fillet the optical measurements are performed. When NIR spectra from the points 1 to 4 were used as separate test sets for the classification model based on the NIR spectra, the number of misclassified samples was 4, 3, 8, and 12, respectively. At the points 3 and 4, we obtained measurements that were less representative for the overall classification of the fillets.
Figure 8 shows that the water binding properties for WB fillets in the upper 1 cm layer were different from the layer at 1 to 2 cm depth. The average T2 relaxation distributions for moderate and severe WB in the upper layer were very similar, illustrating why it is difficult to separate the 2 groups by NMR. Curves from the different samples illustrate the large variation within the groups. Deeper down in these fillets, the water was more tightly bound and resembled the NMR T2 distribution from normal fillets. The average water contents in the upper layer for the 3 groups were 75.5, 78.6, and 79.4% for normal, moderate, and severe WB, respectively, significantly higher in the WB tissue, whereas at 1 to 2 cm depth there were no significant difference in water content between the groups (average values of 75.3, 75.5, and 74.9%, respectively). These findings confirm the heterogeneity of the WB fillets that has been described by others (Soglia et al., 2017; Tasoniero et al., 2017; Wold et al., 2017). They also illustrate the challenge of obtaining a representative measurement of the severity of WB by the use of optical spectroscopy. It is probably important to measure to a certain depth in the fillet in order to discriminate well between normal and affected fillets. Our results suggest that about 1 cm should be sufficient. However, if measurements penetrate too deep, the signal difference between normal and WB tissue would be reduced, because a larger share of normal tissue below the affected tissue will contribute to the measured signal.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
10
WOLD ET AL. 0.8
24.5 24
0.75
23.5
Predicted protein (%)
0.7
BWI
0.65
0.6
23 22.5 22 21.5
0.55 21 0.5
0.45
20.5
1
2
3
4
Measurement point
20
1
2
3
4
Measurement point
Figure 7. Group average NIR estimated BWI and protein content from 4 points of measurement on chicken fillets in test set. Normal fillets (blue), moderate WB (green), severe WB (red).
Upper 1 cm layer 0.06
(A)
Arbitrary unit
0.05 0.04 0.03 0.02 0.01 0 10
10 Relaxation time, sec
0
10
Underlying 1-2 cm layer 0.06
(B)
has been demonstrated in various industrial applications (Wold et al., 2010, 2011). The classification of muscle quality can then be conducted at pixel level, and affected parts can be trimmed away. In this study, we used manual inspection and palpation by an experienced veterinary to sort the fillets into the 3 classes: normal, moderate, and severe WB. NIR and NMR can successfully separate between normal and WB. None of the methods can pick up the degree of severity very well. This might partly be explained by the heterogeneous nature of the syndrome. Another reason might be that the variation in hardness is not related to either protein content or water binding properties, but to some other features, such as for instance the amount of collagen.
Arbitrary unit
0.05 0.04
CONCLUSION
0.03 0.02 0.01 0 10
10 Relaxation time, sec
10
Figure 8. NMR T2 distributions from (A) upper 1 cm layer, and (B) underlying 1 to 2 cm layer. Normal (blue), moderate WB (green), severe WB (red). Thick lines are average curve per group, thin curves are from single samples.
Because some fillets can be affected locally, it can be an advantage to be able to detect affected and less affected regions on the same fillet. This is possible to do by the use of online NIR spectral imaging, which
NIR is a well-suited method for online and nondestructive detection of the syndrome WB in chicken fillets. The main reason for success is that NIR can quantify both the protein content and how loosely the water is bound to the muscle matrix. These are 2 established markers for this particular muscle syndrome. In practice, a classification model can be based on NIR spectra only, that is, additional measurements of protein or NMR are not needed. It is, however, important to include fillets with all relevant variation in such a classification model: variation in degree of WB, in color, remaining blood, temperature, and so on. The heterogeneous distribution of the syndrome implies that it is important where on the breasts the NIR measurement is performed. Results indicate that if the NIR system can measure to a depth of about 1 cm into
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
NIR SPECTROSCOPY DETECTS WOODY BREAST
the breast muscle, good separation between normal and WB can be obtained. Better discrimination is obtained for measurements on the upper part of the breast compared to the lower part. An industrial hyperspectral imaging system would be a good practical solution to handle this heterogeneity.
ACKNOWLEDGMENTS This work was funded by the Norwegian Agricultural Food Research Foundation (Oslo, Norway) through the project FoodSMaCK—Spectroscopy, Modelling & Consumer Knowledge (project number 262308/F40).
REFERENCES Barnes, R. J., M. S. Dhanoa, and S. J. Lister. 1989. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Appl. Spectrosc. 43:772–777. Bertram, H. C., S. Dønstrup, A. H. Karlsson, and H. J. Andersen. 2002a. Continuous distribution analysis of T2 relaxation in meat—an approach in the determination of water-holding capacity. Meat Sci. 60:279–285. Bertram, H. C., P. P. Purslow, and H. J. Andersen. 2002b. Relationship between meat structure, water mobility, and distribution: a low-field nuclear magnetic resonance study. J. Agric. Food Chem. 50:824–829. Carr, H. Y., and E. M. Purcell. 1954. Effects of diffusion on free precession in nuclear magnetic resonance experiments. Am. J. Physiol. 94:630–638. Chung, S. H., A. E. Cerussi, C. Klifa, H. M. Baek, O. Birgul, and G. Gulsen. 2008. In vivo water state measurements in breast cancer using broadband diffuse optical spectroscopy. Phys. Med. Biol. 53:6713–6727. Chung, S. H., H. Yu, M. Y. Su, A. E. Cerussi, and B. J. Tromberg. 2012. Molecular imaging of water binding state and diffusion in breast cancer using diffuse optical spectroscopy and diffusion weighted MRI. J. Biomed. Opt. 17:1–7. Cozzolino, D., I. Murray, and R. Paterson. 1996. Visible and near infrared reflectance spectroscopy for the determination of moisture, fat and protein in chicken breast and thigh muscle. J. Near Infrared Spectrosc. 4:213–223. Duda, R. O., and P. E. Hart. 1973. Pattern Classification and Scene Analysis. New York: Wiley Interscience. Folkestad, A., K. Kolstad, J. P. Wold, K. A. Rørvik, J. Tschudi, and K. H. Haugholt. 2008. Rapid and non-invasive measurements of fat and pigment concentrations in live and slaughtered Atlantic salmon (Salmo salar L.). Aquaculture 280:129–135. Gee, K. 2016. Poultry’s tough new problem: ‘Woody Breast’. Wall Street Journal. Sect. Business and Technology, Mar 29. CCLXVII: B1. Jacques, S. L. 2013. Optical properties of biological tissues: a review. Phys. Med. Biol. 58:R37–R61. Kuttappan, V. A., B. M. Hargis, and C. M. Owens. 2016. White striping and woody breast myopathies in the modern poultry industry: a review. Poult. Sci. 95:2724–2733. Li, M., H. Wang, G. Zhao, M. Qiao, M. Li, L. Sun, X. Gao, and J. Zhang. 2014. Determining the drying degree and quality of chicken jerky by LF-NMR. J. Food Eng. 139:43–49. Libnau, F. O., O. M. Kvalheim, A. A. Christy, and J. Toft. 1994. Spectra of water in the near- and mid-infrared region. Vib. Spectrosc. 7:243–254. Martens, H., and T. Næs. 1993. Multivariate Calibration. New York: Wiley.
11
Meiboom, S., and D. Gill. 1958. Modified spin-echo method for measuring nuclear times. Rev. Sci. Instrum. 29:688–691. Mudalal, S., M. Lorenzi, F. Soglia, C. Cavani, and M. Petracci. 2015. Implications of white striping and wooden breast abnormalities on quality traits of raw and marinated chicken meat. Animal. 9:728–734. NMKL. 1989. Determination of fat according to SBR (SchmidBondzynaski-Ratslaff) in meat and meat products. Nordic committee on food analysis. No. 131:1–6. NMKL. 1991. Determination of water content in meat and meat products. Nordic committee on food analysis. No. 23, 3rd ed.:1– 4. NMKL. 2003. Determination of nitrogen in foods and feeds according to Kjeldahl. Nordic committee on food analysis. No. 6, 4th ed.: 1:9. O’Farrell, M., K. A. H. Bakke, J. Tchudi, and J. P. Wold. 2011. NIR Interactance system for non-contact monitoring of the temperature profiles of baked liver pate. Appl. Spectrosc. 65:1372–1379. Porep, J. U., D. R. Kammerer, and R. Carle. 2015. On-line application of near infrared (NIR) spectroscopy in food production. Trends Food Sci. Technol. 46:211–230. Prieto, N., R. Roehe, P. Lav´ın, G. Batten, and S. Andr´es. 2009. Application of near infrared reflectance spectroscopy to predict meat and meat products quality: a review. Meat Sci. 83:175–186. Sihvo, H. K., K. Immonen, and E. Puolanne. 2014. Myodegeneration with fibrosis and regeneration in the Pectoralis Major muscle of broilers. Vet. Pathol. 51:619–623. Soglia, F., J. Gao, M. Mazzoni, E. Puolanne, C. Cavani, M. Petracci, and P. Ertbjerg. 2017. Superficial and deep changes of histology, texture and particle size distribution in broiler wooden breast muscle during refrigerated storage. Poult. Sci. 96:3465–3472. Soglia, F., L. Laghi, L. Canonico, C. Cavani, and M. Petracci. 2016a. Functional property issues in broiler breast meat related to emerging muscle abnormalities. Food Res. Int. 89:1071–1076. Soglia, F., S. Mudalal, E. Babini, M. Di Nunzio, M. Mazzoni, F. Sirri, C. Cavani, and M. Petracci. 2016b. Histology, composition, and quality traits of chicken Pectoralis major muscle affected by wooden breast abnormality. Poult. Sci. 95:651–659. Tasoniero, G., H. C. Bertram, J. F. Young, A. D. Zotte, and E. Puolanne. 2017. Relationship between hardness and myowater properties in Wooden Breast affected chicken meat: A nuclear magnetic resonance study. LWT – Food Sci. Technol. 86:20– 24. Tijare, V. V., F. L. Yang, V. A. Kuttappan, C. Z. Alvarado, C. N. Coon, and C. M. Owens. 2016. Meat quality of broiler breast fillets with white striping and woody breast muscle myopathies. Poult. Sci. 95:2167–2173. Wold, J.P. 2016. On-line and non-destructive measurement of core temperature in heat treated fish cakes by NIR hyperspectral imaging. Innov. Food Sci. Emerg. Technol. 33: 431– 437. Wold, J. P., M. O’Farrell, M. Høy, and J. Tschudi. 2011. Online determination and control of fat content in batches of beef trimmings by NIR imaging spectroscopy. Meat Sci. 89:317– 324. Wold, J. P., M. Kermit, and A. Woll. 2010. Rapid nondestructive determination of edible meat content in crabs (Cancer pagurus) by near-infrared imaging spectroscopy. Appl. Spectrosc. 64:691– 699 Wold, J. P., E. Veiseth-Kent, V. Høst, and A. Løvland. 2017. Rapid on-line detection and grading of wooden breast myopathy in chicken fillets by near-infrared spectroscopy. PLoS One. 12:e0173384. Zanetti, M. A., D. C. Tedesco, T. Schneider, S. T. F. Teixeira, L. Daroit, F. Pilotto, E. L., Dickel, S. P. Santos, and L. Ruschel dos Santos. 2018. Economic Losses Associated with Wooden Breast and White Striping in Broilers. Semina: Ciˆencias Agr´arias, Londrina, 39:887–892.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
Nofima AS, Norwegian Institute for Food and Fisheries Research, Muninbakken 9-13, Breivika, NO-9291 Tromsø, Norway; and † Nortura SA, Lørenveien 37, NO-0513 Oslo, Norway sification of normal versus WB on the training set, and 96% correct classification on a test set of 52 fillets. The main reason for why NIRS can successfully discriminate between WB and normal fillets is the methods sensitivity to both protein content and degree of water binding in the muscle, both established markers for WB. The classification model can be based on NIR spectra only, calibration against protein is not needed. The affected muscle tissue associated with the WB syndrome is unevenly distributed in the fillets, and this heterogeneity was characterized by NIRS and NMR. Clear differences in water binding properties were found between the superficial 1 cm layer and the deeper layer at 1 to 2 cm depth. Significant differences in protein estimates by NIRS at different measurement points along the chicken fillets were obtained for WB fillets. The findings suggest how to obtain optimal sampling with NIRS for best possible discrimination between WB and normal breast fillets.
ABSTRACT The muscle syndrome woody breast (WB) impairs quality of chicken fillets and is a challenge to the poultry meat industry. There is a need for online detection of affected fillets for automatic quality sorting in process. Near-infrared spectroscopy (NIRS) is a promising method, and in this study we elucidate the spectral properties of WB versus normal fillets. On a training set of 50 chicken fillets (20 normal, 30 WB), we measured NIR, nuclear magnetic resonance (NMR) T2 relaxation distributions, and crude chemical composition. NIRS could estimate protein in the fillets with an accuracy of ±0.64 percentage points. T2 distributions showed that there was a larger share of free water in WB fillets. This difference in water binding generated a shift and narrowing of the water absorption peak in NIR around 980 nm, quantified by a bound water index (BWI). The correlation between BWI and T2 distributions was 0.78, indicating that NIRS contains information about degree of water binding. Discriminant analysis showed that NIRS obtained 100% correct clas-
Key words: woody breast, chicken fillet, NIR spectroscopy, NMR relaxation, classification 2018 Poultry Science 0:1–11 http://dx.doi.org/10.3382/ps/pey351
INTRODUCTION
broiler chickens (Kuttapan et al., 2016). The situation calls for a rapid method for detection, so that affected fillets can be automatically sorted out from the production line. Such a method can also be used to map incidences in large numbers of fillets, in order to trace the main causes responsible for the syndrome, and to work with improvement strategies. Near-infrared spectroscopy (NIRS) has been demonstrated to work well for rapid and non-destructive detection of WB fillets (Wold et al., 2017). The method was applied online in an industrial process and could classify fillets passing by on a conveyor belt at high speed according to protein content. Woody breast fillets have, on average, lower protein content compared to normal breast muscle (Soglia et al., 2016a). During the study, we got indications that NIR could measure not only protein, but probably also water binding properties in the fillets. It has been reported that WB tissue contains more loosely bound water than normal breast muscle (Soglia et al., 2016a). It is important to know if the spectra contain information about both protein
During the last years, the muscle syndrome woody breast (WB) has become a challenge to the poultry meat industry. Woody breast is a myopathy affecting the muscle tissue in the chicken breast (Pectoralis major), making the muscle appear as pale, hard, and outbulging (Sihvo et al., 2014). Because the appearance of this meat is unpleasant and the functional properties are impaired, WB fillets have to be downgraded and used for less valuable products. With typical incidences of 5 to 10% of WB in markets with heavy broilers, this represents significant economic losses (Gee, 2016; Zanetti et al., 2018). The causes for WB are still not clear, but are most likely multifactorial where an important part is related to the fast growth of modern
C 2018 Poultry Science Association Inc. Received January 31, 2018. Accepted July 16, 2018. 1 Corresponding author: [email protected]
1 Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
2
WOLD ET AL.
and water binding properties, in order to understand and apply the method for WB detection in the best possible way. Soglia et al. (2016a) reported mean crude protein content for WB fillets of 22.6% compared to 24.65% in normal muscle tissue, whereas Wold et al. (2017) found average protein values of 18.4, 18.9, and 23.5% in severe WB, moderate WB, and normal fillets, respectively, in the upper 1 cm surface layer of the fillets. The lacking muscle protein in WB muscle is replaced by mainly moisture, but also higher levels of fat and collagen are found (Soglia et al., 2016b, 2017). Note that the protein content in WB fillets is not evenly distributed. The surface part of the fillet tends to be more affected with low protein content, whereas deeper into the fillet, the protein content can be as in normal fillets. Also muscle texture varies unevenly; the superficial layer is harder than the deeper layer, which again does not differ from normal muscle (Soglia et al., 2017). NIRS is widely used for rapid, online and non-destructive determination of typically fat, protein, and moisture in a variety of food products (Porep et al., 2015), and the method is able to estimate the amount of protein in intact chicken fillets (Cozzolino et al., 1996; Wold et al., 2017). Wold et al. (2017) used a protein calibration based on NIRS to successfully detect WB fillets. Fillets with estimated protein below about 21.9% were classified as WB. It has been reported that the water holding capacity in WB fillets is poorer than in normal fillets, shown by increased drip loss and cooking loss (Mudalal et al., 2015; Tijare et al., 2016). Soglia et al. (2016a) measured water mobility by the use of low field nuclear magnetic resonance (NMR) and reported that there are significantly higher shares of loosely bound water in WB tissue, probably due to muscle fiber degeneration. Water binding properties affect the spectral features of the water absorption peak at 980 nm in the NIR spectrum. Hydrogen bonding between water and other molecules such as proteins causes a spectral shift toward longer wavelengths as well as a peak broadening (Chung et al., 2008). This shift, as well as other spectral properties, has been suggested as a basis for measuring water holding capacity and drip loss in meat by NIRS, but it has been difficult to establish solid calibrations with causal explanations for these applications (Prieto et al., 2009). Within medical diagnostics, the spectral water band shift has been used to detect human breast cancer, because cancerous tissue contains more loosely bound water compared to normal tissue (Chung et al., 2012). Chung et al. (2008) introduced a bound water index (BWI) that quantifies the shift observed in tissue water absorption spectra measured by diffuse NIR spectroscopy, and this index correlated well with the occurrence of malignant breast tissue. The index can also be calculated in NIR images of breast tissue, highlighting regions with loosely bound water indicating tumor areas.
Low-field NMR gives well-founded estimates of water distribution and mobility in meat (Bertram et al., 2002a,b). The spin-spin relaxation time, (T2 ), indicates shares of bound water located within highly organized myofibrillar protein matrix, and more loosely bound extramyofibrillar water. Tasoniero et al. (2017) found that there is a rather low correlation (R = 0.41–0.48) between T2 signals and hardness measured on WB fillets. The objective of this paper is to better understand the spectral properties of chicken muscle tissue affected by the WB syndrome. We focus on the NIR wavelength region 760 to 1,040 nm, because this region has been proven well suited to distinguish WB from normal fillets. NIR spectra in interaction mode was recorded from 50 chicken fillets, normal and WB. The NIR spectra were interpreted based on the T2 relaxation distributions obtained by low-field NMR as well as the crude chemical composition. Different approaches for classification of WB fillets based on NIR spectroscopy were tested, including the BWI and prediction of protein content, with the aim of finding the most reliable method. The different approaches were then evaluated on a test set of 52 new fillets. The effects of sample heterogeneity and spectroscopic sampling are discussed.
MATERIAL AND METHODS Chicken Fillets Calibration Set A total of 50 skin and boneless breast fillets (Musculus pectoralis major) were sampled directly from a Norwegian processing line approximately 3 h after CO2 stunning, bleeding, and slaughter of the birds. The birds were as hatched, 32 to 34 d old, and of the strain Ross 308. Average standard broiler live weight at the processing plant is 2.1 kg. The samples were classified by an experienced veterinarian based on visual inspection and palpation of consistency (normal, hard, and very hard). Breast fillets with hard consistency and limited distribution of very hard parts were classified as moderate WB, whereas fillets with extensive areas of very hard consistency were considered severe WB. Woody breast fillets were pale in color, but no chicken breasts had significant amounts of serous fluid at the surface described from some studies in markets with higher slaughter weights. A total of 24 normal fillets and 26 WB fillets were taken out. The WB fillets were classified in 13 moderate and 13 severe WB. Woody breast fillets were from 2 different flocks, whereas the normal fillets were from the same flock. The fillets were packed in plastic bags and stored overnight at 4◦ C, before they were measured with NIR and NMR, and the fat and protein contents were determined. NIR and NMR were measured about 24 h after slaughter. Tasoniero et al. (2017) reported that there were no significant changes in NMR T2 parameters, nor in measured hardness, in normal and WB fillets from 10 to 24 h post mortem.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
NIR SPECTROSCOPY DETECTS WOODY BREAST
Figure 1. Sampling regime of chicken fillets. Circular region 2 was used for NIR and NMR measurements for calibration set. For the test set, region 2 was used for NMR, and NIR was measured in regions number 1 to 4. The excised cylinder was used for NMR measurements. The fillet part above the black horizontal line was used for determination of crude chemical composition of the samples in calibration set. The 2 yellow rectangular regions indicate size and shape of the illuminated area during an NIR measurement. The absorption spectrum was measured between the 2 illuminated regions, in this case in the center of the circular region 2.
Test Set A total of 52 skin and boneless breast fillets were taken from the production line in the same way as the calibration set, but 4 mo later. The WB fillets were from several flocks, whereas the normal fillets were from 2 different flocks. Twenty normal fillets and 32 WB fillets were taken out. The WB fillets were classified in 20 moderate and 12 severe WB. The fillets were packed in plastic bags and stored overnight at 4◦ C before measurements with NIR. NMR was measured on 5 samples from each group to determine T2 relaxation distributions in superficial and deep layers of the fillets. Water content was determined in the same samples.
NIR Measurements Each fillet was measured by an NIR instrument collecting spectra in interaction mode. It was originally designed to measure fat in the muscle of live salmon (Folkestad et al., 2008). Two halogen light sources of 50 W each illuminate the sample in 2 rectangular regions of 5 mm × 20 mm size (indicated in Figure 1). Distance between the 2 illuminated regions is 10 mm. The system collects the light that has penetrated down into the sample and comes up again in a small area of 4 × 4 mm between the 2 illuminated rectangles. The collected light has then traveled from an approximate square area of typically 2 cm × 2 cm, and a depth in the muscle of approximately 10 to 15 mm is probed (Wold,
3
2015). The receiving part of the instrument is in physical contact with the sample in order to reduce stray light. The mode of measurement is called interaction and makes it possible to measure through the skin of a salmon to sample the muscle below. It is therefore also possible to measure rather deep into chicken fillet to probe chemical properties. The system measures at 15 wavelengths in the region 760 to 1,040 nm, and spectral bandwidth is 20 nm. The system does about 70 measurements per second and these were averaged over the total measurement time of 1 s per sample. The system is thoroughly described in O’Farrell et al. (2011). Each intact fillet in the calibration set was measured on the skin side at the location indicated in Figure 1. Temperature of fillets was 4◦ C. A spectrum of pure water at 4◦ C was also measured. This was done by filling a square glass beaker with 1 L of water. At the bottom of the beaker, there was a polished aluminium plate working as a mirror for the NIR radiation, resulting in a transflection mode spectrum. The measurement principle, interaction, is the same as the one used for detection of WB fillets by Wold et al. (2017). That was however an imaging system, capturing the entire width of a 80 cm wide conveyor belt, measuring on fillets in motion without physical contact, whereas the system used in this study measured only a smaller square-like sample area of samples in the steady state. This point system captures spectra with higher signal to noise ratio and penetrates a few mm deeper into the samples.
NMR Relaxation Measurements For the calibration set, a cylindrical sample of diameter 8 mm and height approximately 20 mm was excised from each fillet at the center of where the NIR measurement was performed (Figure 1). The entire cylinder was weighted and placed in a sealed Teflon container, which was inserted in the NMR probe. Before measurement, the samples were thermostated to 25◦ C for 30 min. For the calibration set, the relaxation measurements were performed with a Maran Ultra Resonance 0.5 tesla (Oxford Instruments, Oxfordshire, UK). T2 was measured using the Carr-Purcell-Meiboom-Gill pulse sequence (CPMG) (Carr and Purcell, 1954; Meiboom and Gill, 1958). The T2 measurements were done with a tau-value of 150 μs. Data from 8,000 echoes were collected during 16 scans. The CPMG data were analyzed using WinDXP software (Resonance instruments, Oxfordshire, UK). From the test set, a cylindrical sample of diameter 8 mm and height 20 mm was excised from 5 fillets from each class (normal, moderate, and severe WB) from the corresponding site as for the calibration set (indicated region 2 in Figure 1). The cylinders were cut in 2 equally sized parts, an upper and a lower sub-sample, representing the superficial (0 to 1 cm) and deeper layer (1 to 2 cm) of the fillets. The pieces were weighted and
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
4
WOLD ET AL.
placed in a sealed Teflon container, which was inserted in the NMR probe. Before measurement, the samples were thermostated to 25◦ C. For each subsample, we determined the T2 relaxations distribution. The CPMG analysis of the test set was performed by a 20 MHz NMR instrument (Antek Instruments Inc., Houston, TX). The measurements were done with a tau-value of 150 μs. Data from 8,000 echoes were collected during 4 scans. Another NMR system was used for the test set samples due to technical problems with the one used for the calibrations set.
Crude Chemical Composition Calibrations Set The upper part of the fillet, where the NIR and NMR measurements were performed, was used for determination of crude protein and fat (indicated in Figure 1). The inner fillet (Pectoralis minor) was removed, and the remaining outer fillet (Pectoralis major) was used for analyses. This sample was homogenized and sent to external lab for analysis. Fat was determined based on extraction with diethyl ether and petroleum ether according to the Schmid-BondzynskiRatslaff method (NMKL 131, 1989). Crude proteins were determined according to the Kjeldahl method by the use of copper sulfate as catalyst (NMKL 6, 2003). Test set Water content in upper and lower part of excised subsamples (cylinders) from test set was determined by weighing, drying at 105◦ C for 24 h, and then weighting again (NMKL 23, 1991).
Data Analysis Spectral Pre-processing The NIR spectra were linearized using the logarithm of the inverse of the interactance spectrum (T), log10(1/T). To reduce light scattering effects, the log10(1/T) spectra were normalized by standard normal variate (SNV) (Barnes et al., 1989). That is, for each spectrum the mean value was subtracted and the spectrum was then divided by the standard deviation of the spectrum. Bound Water Index BWI for each sample was calculated based on the NIR spectrum and the spectrum of pure water, both SNV corrected. The wavelength region used for this calculation was 960 nm to 1,040 nm because it covers the water absorption peak. BWI was calculated as BWI =
i
|Sfillet (λi ) − Swater (λi )|
where Sfillet (λi ) is the spectral value at wavelength λi, and Swater (λi ) is the corresponding spectral value in the water spectrum. A large spectral shift from pure water results in a high BWI and vice versa. Protein Calibration An NIR calibration for protein was made by partial least squares regression (PLSR) (Martens and Næs, 1993). The calibration model was
based on one spectrum from each fillet. Performance of the model was estimated by the root mean squared error (RMSE) and the correlation (R) between measured and estimated protein, where the estimated values were obtained by full cross-validation. The cross-validated protein estimates were then used for classification of WB fillets. This calibration was also used to estimate protein values in the test set. Linear Discriminant Analysis Linear discriminant analysis (LDA) (Duda and Hart, 1973) was used to test how well different measures could discriminate WB from normal fillets. Whole NIR spectra, the BWI, and the NIR estimated protein were tested separately. NMR T2 relaxation distributions were also tested as a discriminator. Because the variables in both the NIR spectra and NMR profiles were highly correlated, we used the score values from a principal component analysis (PCA) (Marten and Næs, 1993) of these data sets. These score values are orthogonal to each other and well suited as input variables in LDA. We used the Mahalanobis distance as metric in the discriminant function. The functions were validated by full cross validation on the training set and then evaluated on the independent test set. One-way analysis of variance was performed to analyze group differences between normal and WB. If the P-value was less than 0.05 the differences were considered significant. The software The Unscrambler ver. 9.8 (CAMO Software AS, Oslo, Norway) was used for regression analysis. Calculation of BWI, LDA, spectral pre-processing and significance testing were carried out by the use of MATLAB version 7.10 (The MathWorks Inc., Natick, MA).
RESULTS AND DISCUSSION Crude Chemical Composition Mean values of fat and protein content for the 3 groups, normal, moderate WB, and severe WB, in the calibration set are shown in Figure 2. There were no significant differences in fat content between the 3 groups, whereas the means for protein were significantly different for all groups. The mean values for protein were comparable with those obtained by for instance Soglia et al., (2016a), who obtained 24.65 and 21.60% protein for normal and WB, respectively. The protein values for moderate and severe WB were higher than those obtained for moderate and severe WB by Wold et al. (2017) (18.9 and 18.4%, respectively), which is reasonable, because those values were measured in the 1 cm superficial layer of the fillets. Wold et al. (2017) found that a protein content of 21.9%, estimated by NIR, was a suitable limit to separate between normal and WB fillets. In this study, there were 6 fillets from the moderate WB group with protein content above this threshold. This result can be explained by the sampling. Because the upper 2 cm
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
NIR SPECTROSCOPY DETECTS WOODY BREAST
5
Figure 2. Sample distribution of chemically measured fat and protein in the calibration set. The horizontal line represents the mean, whereas the colored areas correspond to ±1 standard deviation (light area) and ±1.96 standard deviations (= 95% confidence interval, dark area).
layer was measured, higher protein values will be obtained compared to the previous study when the upper 1 cm sample layer was used. When only upper 1 cm is used, a more clear difference between normal and WB can be expected. This is further discussed below.
Spectral Properties The main visible difference in the NIR spectra was a subtle shift in the absorption peak at about 980 nm (Figure 3). At about 1,000 nm there was an apparent isosbestic point, whereas systematic differences occurred at shorter and longer wavelengths. The main peak at 980 nm originates mainly from water absorption. Relatively small contributions are expected from fat, protein, and myoglobin (Jacques, 2013) for chicken breast muscle in this spectral region. For simplicity, the peak is referred to as the “water peak”. The water peak in spectra from WB fillets was slightly blueshifted toward shorter wavelengths compared to normal fillets. The peak was also more narrow. It was close to the peak of pure liquid water, suggesting that much of the water in WB fillets is free, and very weakly bound to other macromolecules. The BWI quantifies this shift, and it is clear that it varied systematically with the quality of the muscle tissue (Figure 4). Normal fillets had high BWI values (above 0.6), whereas WB fillets had lower values (below 0.6). Low values indicate that the spectral difference from pure water is small. There was no difference in BWI between moderate and severe WB fillets. When compared with the distribution of protein in the same samples (Figure 2), it is clear that the BWI was a better discriminator of normal and WB fillets than the protein content for this particular data set. So although the shift observed in Figure 3 looks subtle to the naked eye, it is rather pronounced for a stable NIR spectroscopic system of high quality. The observed variation is much larger than the noise level in such systems, which gives a
good basis for discrimination between normal and WB fillets. A PCA of the NIR spectra (760 to 1,040 nm) resulted in a first principal component (PC1), which explained 97% of the spectral variation. The score values of PC1 correlated highly with BWI (R = 0.997), which means that PC1 expressed more or less the same shift variation. It should be noted that the calculation of BWI in this case is simplified compared to the original procedure proposed by Chung et al. (2008). Before calculating BWI they corrected the spectra by removing contributions from lipids and hemoglobin. These contributions are probably substantial and can vary a lot in and between human breasts. The spectral variation due to fat and myoglobin in the chicken fillets is probably small enough to be ignored, but this should be investigated further before using BWI alone to classify WB. Fat content is very low and would most likely not affect the spectra much, but the amount of remaining blood in the fillets might vary and could affect the spectral shape and also the calculated BWI. Protein has an absorption peak at about 1,020 nm (not discernable in the spectra). A lower protein content and a corresponding increase in moisture, as the case is for WB muscle, could also affect the observed spectral shift. Another important parameter to consider is sample temperature, because temperature also induces a shift in the water peak. As temperature increases, the fraction of hydrogen-bound water molecules is reduced, causing the peak to increase in intensity, narrow in bandwidth, and blueshift to shorter wavelengths (Libnau et al., 1994). Actually, NIR spectroscopy as measured in this work can be used to determine core temperature in heat treated or chilled food products (Wold, 2015). Temperature changes will therefore directly affect the value of BWI. In a poultry processing plant, the temperature of the fillets should therefore be stable within a few degrees. If this is not the case,
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
6
WOLD ET AL. 2
2
1.5
1.5
Arbitrary unit
Arbitrary unit
1
0.5
1
0.5
0 0
-0.5
-1
800
950 900 850 Wavelength, nm
1000
-0.5 960
980
1000 Wavelength, nm
1020
1040
Figure 3. Left panel: SNV corrected NIR spectra from one selected filet of each class normal (blue), moderate WB (green), severe WB (red). Right panel: zoom-in on water absorption peak with spectrum of liquid water (black) added for comparison.
Figure 4. Sample distributions of derived properties from NIR spectra and NMR T2 relaxation distributions. The horizontal line represents the mean, whereas the colored areas correspond to ±1 standard deviation (light area) and ±1.96 standard deviations (= 95% confidence interval, dark area).
then it should also be possible to correct the spectra for temperature effects before calculation of BWI.
Protein Calibration based on NIR The protein calibration based on NIR spectra from the chicken fillets made by PLSR required 5 latent variables. By cross validation we estimated the root mean square error of cross validation (RMSECV) to 0.64 percentage points and a correlation, R, between estimated and measured protein of 0.89. This result is close to what we previously obtained (RMSECV = 0.57, R = 0.76) (Wold et al., 2017). The higher R in this study was due to a greater protein content range. For classification of WB fillets, Wold et al. (2017) found an optimal decision limit of 21.9%. Fillets with protein estimated under this limit were classified as WB. When the same limit was used in this study, 5
of the 26 WB samples were erroneously classified as normal (Figure 4). These samples were all from the moderate WB group. All normal samples were correctly classified.
NMR Relaxation Figure 5 shows T2 relaxation distributions measured in chicken breast muscle from 1 normal and 2 WB samples from the calibration set, the same samples as used to illustrate NIR spectra in Figure 3. The distributions consist of 2 major components, which are characteristic for meat. The peak between 0.02 and 0.1 s is often referred to as T21 , and the smaller peak in the region 0.18 and 0.50 s is called T22 (Bertram et al, 2002b). T21 indicates the share of water bound in myofibrils, whereas T22 indicates share of more loosely bound water outside the muscle fibers.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
7
NIR SPECTROSCOPY DETECTS WOODY BREAST 0.08 0.07
Amplitude
0.06 0.05 0.04 0.03 0.02 0.01 0 10
10
10
T2 relaxation time, sec Figure 5. NMR T2 relaxation distributions from normal fillet (blue), moderate WB (green), and severe WB (red).
The T2 distribution for the normal chicken muscle in this study resembles similar measurements reported for pork and chicken muscle, a large T21 component and a much smaller T22 component separated by a time region with no signals (Bertram et al., 2002a,b; Li et al., 2014). T2 distributions from WB muscles had less contribution from the T21 component and more from the T22 component. The 2 components were also broader and connected to one continuous curve, indicating that the water binding properties in WB muscle span a range of different states. The distributions clearly illustrate that there is a much larger share of loosely bound water with higher mobility in WB muscle. This corresponds with earlier observations of water relaxation times in WB muscle (Soglia et al., 2016a). Bertram et al. (2002a) observed a third population of water (T2b ) at very short relaxation times (0.001 to 0.005 s), and proposed that the signal originated from water tightly bound to macromolecules. We did also observe signals in this region; however, they were of very low intensity and did not vary systematically with the degree of WB. These signals could possibly vary with for instance changes in protein secondary structure, but we did not study this structure, and it was therefore difficult to interpret the variation in T2b . Because T2 relaxation expresses the water binding properties of water in the meat, it is of interest to elucidate how the NIR spectra and NMR profiles related to each other. This was done by PLSR analysis and showed that there was a correlation (R) between T2 relaxation distributions and BWI of 0.78. Note also that there was a corresponding correlation (R) between NMR T2 and protein content of 0.92. The water binding properties were hence closely related to the amount of protein. The first PC of the T2 distributions explained 77% of the variation, and Figure 4 shows that the scores for this component separated quite well between normal and WB fillets.
Linear Discriminant Analysis In this section, we first used LDA to evaluate different approaches for classification of WB versus normal fillets on the calibration set of 50 samples. The main aim was to find a good classifier that can be used online. It was also interesting to investigate if T2 relaxation distributions could separate well between the 2 groups. We then used the discriminant functions based on NIR data from the calibration set to classify the test set obtained 4 mo later. All the methods performed well on the calibration set; however, the best discrimination was obtained when either the BWI or the full NIR spectra were used for classification (Table 1a). The BWI separated well between normal and WB fillets, which could also be seen from Figure 4. PC1 from the NIR spectra performed equally well, and that is reasonable because these 2 parameters were highly correlated. They both expressed the shift variation in the water peak. The approach of making a protein calibration based on NIR, and then use this as a classifier, was slightly less successful for this data set. The predicted protein values of WB overlapped those of the normal fillets. The fact that the T2 distributions from NMR could distinguish almost perfectly between normal and WB fillets indicates again that water binding properties are distinctly different for the 2 groups. We did also investigate if the different measurements could distinguish between moderate and severe WB fillets (Table 1b). This proved to be more difficult, where NMR and NIR spectra obtained best results with about 76% correct classifications. It should also be noted that the 2 parameters BWI and predicted protein together obtained 76.9% correct classification. The classification results for the test set differed slightly from the calibration set (Table 2). The BWI performed less well with 7 miss-classifications, whereas
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
8
WOLD ET AL. Table 1. Classification results for different data for the calibration set (cross validation) Data type
# PC1
#Correct
#Non-correct
a) Classification of WB versus normal fillets BWI 1 NIR spectra 1 NIR estimated protein 1 Protein 1 NMR 2
50 50 47 44 49
0 0 3 6 1
b) Classification of moderate versus severe WB BWI 1 NIR spectra 2 NIR estimated protein 1 Protein 1 NMR 2
14 20 15 18 20
12 6 11 8 6
% Correct classification 100 100 94 88 98 53.8 76.9 57.7 69.2 76.9
#Correct and #Non-correct indicate the number of correctly and non-correctly classified samples. 1 #PC—Number of principal components used in the model.
Table 2. Classification results for different data for the test set. Data type
# PC1
# Correct
# Non-correct
% correct classification
Classification of WB versus normal fillets BWI 1 NIR spectra 4 NIR estimated protein 1
45 49 50
7 3 2
86.5 94.2 96.1
b) Classification of moderate versus severe WB BWI 1 NIR spectra 1 NIR estimated protein 1
15 22 22
17 10 10
46.8 68.7 68.7
#Correct and #Non-correct indicate the number of correctly and non-correctly classified samples. 1 #PC—Number of principal components used in the model.
NIR predicted protein and NIR spectra performed better with only 3 and 2 miss classifications, respectively. Note that for the NIR spectra, the discriminant function was based on 4 PCs, not only the first PC shown in Figure 4. It means that there was more information in the NIR spectrum than only the water peak shift that was used for the classification. The high number of correct classifications on a dataset obtained 4 mo later indicates that the method is robust, as was also demonstrated by Wold et al. (2017). The WB myopathy appears in different degrees, from moderate to severe. The variation is both in degree within an affected part of the muscle, and in extension of affected areas. There will therefore be a gradual change from normal to severe cases. In such a system there will always be borderline cases between the groups and impossible to obtain 100% correct classification for a number of samples. From the rather limited set of samples in this study, it is clear that both the water binding property and the protein content are good discriminators with regard to WB. Both features can be derived from the NIR spectrum. Therefore, the NIR spectrum is a good discriminator and the spectrum alone can be used successfully as input in a classification model. With respect to an industrial application, there are certainly some advantages of using just the NIR spectrum for classification rather than a protein calibration. Chemical reference
measurements for protein can be avoided, which means reduced costs. On the other hand, for system operators it might be more informative and meaningful to use a system that produces protein and BWI values, rather than a black box classifier of the NIR spectra. In Figure 6, the NIR predicted protein is plotted versus BWI for all samples included in both calibration and test set. Most WB fillets had low scores for both parameters, but there were samples manually classified as WB with high protein and low BWI, or high BWI and low protein. There was also one normal breast with high protein content, but low BWI. The manual classification is based mainly on the hardness of the breasts, and also the share of hard tissue. For samples in the gray zone between normal and WB, the developed hardness does not necessarily correlate well with protein content or BWI. Tasoniero et al. (2017) found rather low correlations in the range 0.41 to 0.48 between NMR T2 parameters and measured hardness for WB fillets. Figure 6 also shows that protein content and BWI do not discriminate well between moderate and severe WB. There can be at least 3 reasons for this. (1) As stated above, the hardness does not necessarily correlate well with the 2 parameters, (2) the NIR spectra were obtained from a limited region of the fillet and did not include distributional information about hardness, and (3) the heterogeneity in depth of the fillet will
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
9
NIR SPECTROSCOPY DETECTS WOODY BREAST 25
24
Predicted protein %
23
22
21
20
19
18 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Bound water index Figure 6. Predicted protein content versus BWI for all samples (calibration and test set). Normal (white), moderate WB (green), severe WB (red). Calibration set: circles, test set: triangles.
affect the optical measurements and might obscure the relation.
Effects of Sample Heterogeneity It is obvious that the morphological changes in the muscle tissue associated with the WB syndrome are unevenly distributed in the fillets. The distribution follows some typical patterns, it varies both across the fillet surface and in depth depending on the severity of the syndrome. This means that the protein content and water binding properties will also vary accordingly. Figure 7 shows average estimated BWI and protein for the 3 groups in the test set. All the fillets were measured with NIR at the points 1 to 4 (Figure 1). There were significant differences between the normal fillets and the WB fillets. However, there were also significant differences within the fillets, in particular for the WB fillets. The BWI and protein contents were significantly lower in the rostral thicker part of the WB fillets (points 1 and 2 vs. point 4), whereas the variation in the normal fillets was less pronounced. This heterogeneity indicates that it is important where on the fillet the optical measurements are performed. When NIR spectra from the points 1 to 4 were used as separate test sets for the classification model based on the NIR spectra, the number of misclassified samples was 4, 3, 8, and 12, respectively. At the points 3 and 4, we obtained measurements that were less representative for the overall classification of the fillets.
Figure 8 shows that the water binding properties for WB fillets in the upper 1 cm layer were different from the layer at 1 to 2 cm depth. The average T2 relaxation distributions for moderate and severe WB in the upper layer were very similar, illustrating why it is difficult to separate the 2 groups by NMR. Curves from the different samples illustrate the large variation within the groups. Deeper down in these fillets, the water was more tightly bound and resembled the NMR T2 distribution from normal fillets. The average water contents in the upper layer for the 3 groups were 75.5, 78.6, and 79.4% for normal, moderate, and severe WB, respectively, significantly higher in the WB tissue, whereas at 1 to 2 cm depth there were no significant difference in water content between the groups (average values of 75.3, 75.5, and 74.9%, respectively). These findings confirm the heterogeneity of the WB fillets that has been described by others (Soglia et al., 2017; Tasoniero et al., 2017; Wold et al., 2017). They also illustrate the challenge of obtaining a representative measurement of the severity of WB by the use of optical spectroscopy. It is probably important to measure to a certain depth in the fillet in order to discriminate well between normal and affected fillets. Our results suggest that about 1 cm should be sufficient. However, if measurements penetrate too deep, the signal difference between normal and WB tissue would be reduced, because a larger share of normal tissue below the affected tissue will contribute to the measured signal.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
10
WOLD ET AL. 0.8
24.5 24
0.75
23.5
Predicted protein (%)
0.7
BWI
0.65
0.6
23 22.5 22 21.5
0.55 21 0.5
0.45
20.5
1
2
3
4
Measurement point
20
1
2
3
4
Measurement point
Figure 7. Group average NIR estimated BWI and protein content from 4 points of measurement on chicken fillets in test set. Normal fillets (blue), moderate WB (green), severe WB (red).
Upper 1 cm layer 0.06
(A)
Arbitrary unit
0.05 0.04 0.03 0.02 0.01 0 10
10 Relaxation time, sec
0
10
Underlying 1-2 cm layer 0.06
(B)
has been demonstrated in various industrial applications (Wold et al., 2010, 2011). The classification of muscle quality can then be conducted at pixel level, and affected parts can be trimmed away. In this study, we used manual inspection and palpation by an experienced veterinary to sort the fillets into the 3 classes: normal, moderate, and severe WB. NIR and NMR can successfully separate between normal and WB. None of the methods can pick up the degree of severity very well. This might partly be explained by the heterogeneous nature of the syndrome. Another reason might be that the variation in hardness is not related to either protein content or water binding properties, but to some other features, such as for instance the amount of collagen.
Arbitrary unit
0.05 0.04
CONCLUSION
0.03 0.02 0.01 0 10
10 Relaxation time, sec
10
Figure 8. NMR T2 distributions from (A) upper 1 cm layer, and (B) underlying 1 to 2 cm layer. Normal (blue), moderate WB (green), severe WB (red). Thick lines are average curve per group, thin curves are from single samples.
Because some fillets can be affected locally, it can be an advantage to be able to detect affected and less affected regions on the same fillet. This is possible to do by the use of online NIR spectral imaging, which
NIR is a well-suited method for online and nondestructive detection of the syndrome WB in chicken fillets. The main reason for success is that NIR can quantify both the protein content and how loosely the water is bound to the muscle matrix. These are 2 established markers for this particular muscle syndrome. In practice, a classification model can be based on NIR spectra only, that is, additional measurements of protein or NMR are not needed. It is, however, important to include fillets with all relevant variation in such a classification model: variation in degree of WB, in color, remaining blood, temperature, and so on. The heterogeneous distribution of the syndrome implies that it is important where on the breasts the NIR measurement is performed. Results indicate that if the NIR system can measure to a depth of about 1 cm into
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018
NIR SPECTROSCOPY DETECTS WOODY BREAST
the breast muscle, good separation between normal and WB can be obtained. Better discrimination is obtained for measurements on the upper part of the breast compared to the lower part. An industrial hyperspectral imaging system would be a good practical solution to handle this heterogeneity.
ACKNOWLEDGMENTS This work was funded by the Norwegian Agricultural Food Research Foundation (Oslo, Norway) through the project FoodSMaCK—Spectroscopy, Modelling & Consumer Knowledge (project number 262308/F40).
REFERENCES Barnes, R. J., M. S. Dhanoa, and S. J. Lister. 1989. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Appl. Spectrosc. 43:772–777. Bertram, H. C., S. Dønstrup, A. H. Karlsson, and H. J. Andersen. 2002a. Continuous distribution analysis of T2 relaxation in meat—an approach in the determination of water-holding capacity. Meat Sci. 60:279–285. Bertram, H. C., P. P. Purslow, and H. J. Andersen. 2002b. Relationship between meat structure, water mobility, and distribution: a low-field nuclear magnetic resonance study. J. Agric. Food Chem. 50:824–829. Carr, H. Y., and E. M. Purcell. 1954. Effects of diffusion on free precession in nuclear magnetic resonance experiments. Am. J. Physiol. 94:630–638. Chung, S. H., A. E. Cerussi, C. Klifa, H. M. Baek, O. Birgul, and G. Gulsen. 2008. In vivo water state measurements in breast cancer using broadband diffuse optical spectroscopy. Phys. Med. Biol. 53:6713–6727. Chung, S. H., H. Yu, M. Y. Su, A. E. Cerussi, and B. J. Tromberg. 2012. Molecular imaging of water binding state and diffusion in breast cancer using diffuse optical spectroscopy and diffusion weighted MRI. J. Biomed. Opt. 17:1–7. Cozzolino, D., I. Murray, and R. Paterson. 1996. Visible and near infrared reflectance spectroscopy for the determination of moisture, fat and protein in chicken breast and thigh muscle. J. Near Infrared Spectrosc. 4:213–223. Duda, R. O., and P. E. Hart. 1973. Pattern Classification and Scene Analysis. New York: Wiley Interscience. Folkestad, A., K. Kolstad, J. P. Wold, K. A. Rørvik, J. Tschudi, and K. H. Haugholt. 2008. Rapid and non-invasive measurements of fat and pigment concentrations in live and slaughtered Atlantic salmon (Salmo salar L.). Aquaculture 280:129–135. Gee, K. 2016. Poultry’s tough new problem: ‘Woody Breast’. Wall Street Journal. Sect. Business and Technology, Mar 29. CCLXVII: B1. Jacques, S. L. 2013. Optical properties of biological tissues: a review. Phys. Med. Biol. 58:R37–R61. Kuttappan, V. A., B. M. Hargis, and C. M. Owens. 2016. White striping and woody breast myopathies in the modern poultry industry: a review. Poult. Sci. 95:2724–2733. Li, M., H. Wang, G. Zhao, M. Qiao, M. Li, L. Sun, X. Gao, and J. Zhang. 2014. Determining the drying degree and quality of chicken jerky by LF-NMR. J. Food Eng. 139:43–49. Libnau, F. O., O. M. Kvalheim, A. A. Christy, and J. Toft. 1994. Spectra of water in the near- and mid-infrared region. Vib. Spectrosc. 7:243–254. Martens, H., and T. Næs. 1993. Multivariate Calibration. New York: Wiley.
11
Meiboom, S., and D. Gill. 1958. Modified spin-echo method for measuring nuclear times. Rev. Sci. Instrum. 29:688–691. Mudalal, S., M. Lorenzi, F. Soglia, C. Cavani, and M. Petracci. 2015. Implications of white striping and wooden breast abnormalities on quality traits of raw and marinated chicken meat. Animal. 9:728–734. NMKL. 1989. Determination of fat according to SBR (SchmidBondzynaski-Ratslaff) in meat and meat products. Nordic committee on food analysis. No. 131:1–6. NMKL. 1991. Determination of water content in meat and meat products. Nordic committee on food analysis. No. 23, 3rd ed.:1– 4. NMKL. 2003. Determination of nitrogen in foods and feeds according to Kjeldahl. Nordic committee on food analysis. No. 6, 4th ed.: 1:9. O’Farrell, M., K. A. H. Bakke, J. Tchudi, and J. P. Wold. 2011. NIR Interactance system for non-contact monitoring of the temperature profiles of baked liver pate. Appl. Spectrosc. 65:1372–1379. Porep, J. U., D. R. Kammerer, and R. Carle. 2015. On-line application of near infrared (NIR) spectroscopy in food production. Trends Food Sci. Technol. 46:211–230. Prieto, N., R. Roehe, P. Lav´ın, G. Batten, and S. Andr´es. 2009. Application of near infrared reflectance spectroscopy to predict meat and meat products quality: a review. Meat Sci. 83:175–186. Sihvo, H. K., K. Immonen, and E. Puolanne. 2014. Myodegeneration with fibrosis and regeneration in the Pectoralis Major muscle of broilers. Vet. Pathol. 51:619–623. Soglia, F., J. Gao, M. Mazzoni, E. Puolanne, C. Cavani, M. Petracci, and P. Ertbjerg. 2017. Superficial and deep changes of histology, texture and particle size distribution in broiler wooden breast muscle during refrigerated storage. Poult. Sci. 96:3465–3472. Soglia, F., L. Laghi, L. Canonico, C. Cavani, and M. Petracci. 2016a. Functional property issues in broiler breast meat related to emerging muscle abnormalities. Food Res. Int. 89:1071–1076. Soglia, F., S. Mudalal, E. Babini, M. Di Nunzio, M. Mazzoni, F. Sirri, C. Cavani, and M. Petracci. 2016b. Histology, composition, and quality traits of chicken Pectoralis major muscle affected by wooden breast abnormality. Poult. Sci. 95:651–659. Tasoniero, G., H. C. Bertram, J. F. Young, A. D. Zotte, and E. Puolanne. 2017. Relationship between hardness and myowater properties in Wooden Breast affected chicken meat: A nuclear magnetic resonance study. LWT – Food Sci. Technol. 86:20– 24. Tijare, V. V., F. L. Yang, V. A. Kuttappan, C. Z. Alvarado, C. N. Coon, and C. M. Owens. 2016. Meat quality of broiler breast fillets with white striping and woody breast muscle myopathies. Poult. Sci. 95:2167–2173. Wold, J.P. 2016. On-line and non-destructive measurement of core temperature in heat treated fish cakes by NIR hyperspectral imaging. Innov. Food Sci. Emerg. Technol. 33: 431– 437. Wold, J. P., M. O’Farrell, M. Høy, and J. Tschudi. 2011. Online determination and control of fat content in batches of beef trimmings by NIR imaging spectroscopy. Meat Sci. 89:317– 324. Wold, J. P., M. Kermit, and A. Woll. 2010. Rapid nondestructive determination of edible meat content in crabs (Cancer pagurus) by near-infrared imaging spectroscopy. Appl. Spectrosc. 64:691– 699 Wold, J. P., E. Veiseth-Kent, V. Høst, and A. Løvland. 2017. Rapid on-line detection and grading of wooden breast myopathy in chicken fillets by near-infrared spectroscopy. PLoS One. 12:e0173384. Zanetti, M. A., D. C. Tedesco, T. Schneider, S. T. F. Teixeira, L. Daroit, F. Pilotto, E. L., Dickel, S. P. Santos, and L. Ruschel dos Santos. 2018. Economic Losses Associated with Wooden Breast and White Striping in Broilers. Semina: Ciˆencias Agr´arias, Londrina, 39:887–892.
Downloaded from https://academic.oup.com/ps/advance-article-abstract/doi/10.3382/ps/pey351/5079646 by Bukkyo University Library user on 27 August 2018