J. Anim. Sci. 2003. 81:1177-1184
© 2003 American Society of Animal Science
Comparison of different methods to assess the composition of pig bellies in progeny testing
E. Tholen*,1,
U. Baulain
,
M. D. Henning and
K. Schellander*
* Institute of Animal Breeding Science of the University of Bonn, D-53115 Bonn, Germany and
and
Institute for Animal Science Mariensee, Federal Agricultural Research Center, D-31535 Neustadt, Germany
1 Correspondence:
phone: +49-228-733589; fax: +49 228 732284; E-mail:
etholen{at}itz.uni-bonn.de.
 |
Abstract
|
|---|
The objective of the study was to validate methods that assess the belly composition of stationary tested progenies of Piétrain boars. In German performance test stations, there are currently three methods of determining belly compositon in use: 1) a regression equation that contains different carcass characteristics, such as fat thickness and muscle area; 2) planimetric analysis of video or digital images acquired at the cut between the 13th and 14th ribs; and 3) estimation of the belly composition using ultrasound data from a three-dimensional ultrasound image produced an online carcass-grading system. Validation of these techniques was performed on 400 carcasses of stationary-tested Piétrain and Piétrain-sired crossbred pigs, which were slaughtered at a mean carcass weight of 85 and 97 kg. Magnetic resonance imaging (MRI) served as a reference to determine the lean content of the bellies. The correlation to MRI lean content ranged from 0.71 to 0.81, and corresponding correlation values were 0.62 to 0.64 for the digital imaging technique, and 0.53 to 0.59 for the AutoFOM online carcass-grading system. An increase in precision was achieved when information from digital imaging and linear carcass measures were included in the regression equation. Accuracy of the AutoFOM system does not seem to be sufficient to assess the belly composition for the special breeds in performance testing. However, extracting and combining 127 AutoFOM-base recordings into modified equations using partial least squares techniques yielded an improvement in the prediction accuracy for all tested breed and/or weight groups.
Key Words: Belly • Image Processing • Magnetic Resonance Imaging • Performance Testing • Pigs • Ultrasound
|
Introduction
|
|---|
The market value of a pig carcass traditionally has been determined by its overall lean content. In recent years, marketing of carcass cuts has been growing at the expense of marketing entire sides of pork (J. Beuck, personal communication). Thus, to accurately discern primal cut value, pork packers must have detailed information about the lean composition of these pork cuts.
With respect to market value of the entire carcass, the belly cut has a specific importance, particularly in Germany. During the summer months, price differences between processing and barbecue bellies range from $0.50 to $0.75/kg (ZMP-Bilanz Vieh und Fleisch, 2000). Moreover, Evans and Kempster (1978), Fewson et al. (1990), and Pfuhl and Glodek (1996) reported close relationships between the lean content of the whole carcass and the valuable ham, loin, and shoulder cuts, whereas less distinct correlations were observed between the lean content of the belly and other carcass cuts (Fewson et al., 1990; Hulsegge et al., 1994). This slight relationship might have a relevant impact on the added value of pig carcasses and provides the justification to consider the estimated belly cut composition in the payment system of some major slaughterers in Germany (Beuck, 1999) and in the breeding goal of the Bavarian pig-breeding herdbook organizations (Götz, 1999).
Currently, three different belly composition-determining methods are used in the progeny testing schemes of German pig breeding organizations: 1) breed specific regression equations, which contain different carcass characteristics; 2) planimetric analysis of muscle and fat area from digital images of the belly; or 3) estimation of the belly composition within the scope of the AutoFOM online carcass grading routine (SFK Technology, Herlev, Denmark). Therefore, the objective of the study was to validate these techniques in relation to the accurate belly composition that was determined by magnetic resonance imaging.
|
Material and Methods
|
|---|
Because of its dominant position in German crossbreed programs, the investigation used carcasses of Piétrain and Piétrain-sired pigs. Pork bellies from purebred Piétrain (Pi; n = 200) and Piétrain x [German Large White ( German Landrace] (PiF1, n = 198) gilts originating from sibling and progeny performance testing in two stations were taken to evaluate the aforementioned different belly composition-determining methods. Pigs were housed in groups of two female full sibs given ad libitum access to feed and water. The wheat-barley-soybean meal diets contained 16% CP, 1% lysine, and 13.0 MJ/kg of ME. The fattening period started at 35 kg and ended for all Pi and 98 PiF1 gilts at a live weight of 105 kg (carcass weight of approximate 85 kg) or, for 100 PiF1 pigs, at a live weight of 125 kg (carcass weight of approximately 97 kg).
All pigs were slaughtered at a commercial abattoir where an AutoFOM device was used in the online carcass-grading routine. A conveyer pulled carcasses across an array of ultrasound transducers. The U-shaped transducer array consisted of 16 transducers designed to anatomically fit the back of the carcass. Each of the 16 transducers performed 200 A-scans perpendicular to each carcass at a distance of 5 mm. Overall, approximately 3,200 A-scan measurements were taken for each carcass. Of these, 127 of the most informative single measures were extracted and used in our analysis. Besides technical ultrasound signals, the information comprised various fat and muscle depths used to estimate the carcass composition (Brøndum et al., 1998). After online grading, linear fat measurements were taken according to the rules of station testing of pigs in Germany (ALZ, 2001). At the 13th-/14th-rib interface, longissimus muscle (LM) and fat areas, as well as lean:fat proportions of an 8-cm-wide belly cross-section, were measured via digital image on a 1:1 scale and at a resolution of 640 x 512 pixels. Skala, the semiautomatic program package (Skala, O. Wegner, Kiel, Germany), was used to differentiate between fat and muscle areas on the digital image.
Lean content of the entire carcass was estimated by the "Bonner formula," comprising four linear carcass measurements (Table 1
, Schmitten et al. 1986). Breed-specific regression equations (Gruber formula; Table 1) were used to estimate the lean content of the bellies. These formulas were evaluated by Tholen et al. (1998), using information about the belly lean content, which was determined by dissection and linear carcass measurements from Pi gilts and Pi x German Landrace barrows. Correlation coefficients (R2 values) of the "Gruber formula" for belly lean content were 58.5 and 61.6% for Pi gilts and Pi x German Landrace barrows, respectively (Tholen et al., 1998).
Bellies were trimmed at the ends and sides according to the "Deutsche Landwirtschaftsgesellschaft" dissection scheme (Figure 1; Scheper and Scholz, 1985). Deboned and trimmed bellies were investigated by magnetic resonance imaging (MRI) at the Institute for Animal Science Mariensee (Neustadt, Germany; Figure 2). Magnetic resonance imaging provides insight into the body’s interior and is one of the most powerful diagnostic tools in medicine. A detailed description of MRI and its application in animal science was given by Baulain (1997) and Mitchell et al. (2001). Baulain et al. (1998) have demonstrated that MRI is a convenient way to measure muscle and fat volume in the pig belly with high accuracy, and MRI can replace total dissection as a reference technique to determine lean content.
View larger version (50K):
[in this window]
[in a new window]
|
Figure 2. Magnetic resonance images; longitudinal view of a pig belly and transverse images of a belly between the 7th and 14th ribs. Each trimmed belly was virtually divided into 22 to 27 slices, which had a thickness of 8 mm and a slice-to-slice distance of 16 mm. In a semiautomatic image-processing procedure, pixel surface information of the slices was classified as muscle or fat tissue. These tissue areas were summed over the entire belly and multiplied by slice distance to calculate tissue volumes.
|
|
To compute muscle and fat volume, each belly was virtually divided into 22 to 27 slices with a thickness of 8 mm and a slice-to-slice distance of 16 mm (Figure 2
). The image matrix was 30 cm x 15 cm and consisted of 256 x 128 pixels. In a semi-automatic image processing procedure, each pixel was classified into muscle or fat tissue applying a cluster analysis (Scholz et al., 1993) and muscle and fat areas were calculated for each section. To compute muscle and fat volumes, these tissue areas were summed up over the entire belly and multiplied by slice-to-slice distance. This volume calculation is derived from the Cavalieri method, which is a well-known technique in stereology (Roberts et al., 1993). To enable a comparison of percentage of lean estimated by the different methods in progeny testing to the lean content by MRI, muscle weight was predicted from muscle volume assuming a specific gravity of muscle tissue of 1.06 g/cm3.
Statistical Analysis
Differences and linear relationships between MRI lean content of the belly and the other composition-determining methods were calculated throughout and within breed and/or weight groups. Additionally, differences between all methods were calculated within quartile intervals of different belly composition classes arranged according to MRI lean content.
The Stepwise procedure of SAS (Version 8.2, SAS Inst., Inc., Cary, NC) was used to construct modified estimation functions of the MRI lean content. Eleven linear stationary carcass measurements were used as potential independent variables. The Stepwise process was stopped when only variables with an F-statistic of P < 0.15 were included in the prediction equations.
Alternatively, an attempt was made to construct an estimation function of the MRI lean content of the belly using partial least squares (PLS) techniques (de Jong, 1993) implemented in the SAS procedure PLS. Ordinary least squares regression has the single goal of minimizing sample response prediction error, seeking linear functions of the predictors (linear carcass measurements) that explain as much variation in each response (lean content of the belly) as possible. The techniques implemented in the PLS procedure have the additional goal of accounting for variation in the predictors, under the assumption that directions in the predictor space that are well sampled should provide better prediction for new observations when the predictors are highly correlated. The applied PLS method consists of three basic steps. The first step is the extraction of linear combinations of latent vectors (factors which balance the two objectives, seeking for factors that explain both responses and prediction variation). The second step focuses on the isolation of a few underlying factors that provide a good predictive model by cross validation (split sample). Within that framework, test and training datasets were arbitrarily chosen in order to fit the model (test data) and to measure how well models with different numbers of extracted factors fit the other data (training data) set. The "predicted residual sum of squares" statistic is based on the residuals generated by this repeatedly performed process. The isolation of a sufficient number of factors is based on the minimal "predicted residual sum of squares" statistic and a significance test proposed by Van Voet (1994). The third step requires the quantification of the importance of each variable by Wold’s "variable importance for projection" (VIP; Wold, 1994). The regression coefficients represent the importance that each predictor has in the prediction of the response, whereas the VIP represents the value of each predictor in fitting the PLS model for both predictors and response. If a predictor has a relatively small coefficient (in absolute value) and a small VIP value, then it is a prime candidate for deletion. Wold (1994) considered a value less than 0.8 to be "small" for the VIP.
According to the PLS steps described above, two sets of PLS estimation functions were constructed by means of 11 linear stationary carcass measurements or the 127 AutoFOM-base recordings available. After the third step, all traits that had a VIP of less than 0.8 were deleted, and in the case of the PLS analysis of the 127 AutoFOM-base recordings, only 20 of the most informative measurements were selected. After these performed reductions, the first and second steps were repeated in order to evaluate the final regression equations.
|
Results and Discussion
|
|---|
Weights and percentages of lean from pork bellies estimated by the different techniques are shown in Table 2. Mean values of the "Gruber formula" were very close to the MRI reference values in all breed and/or weight groups. The observed marginal deviations could be explained by differences between the data set used by Tholen et al. (1998) in the evaluation of the Gruber formula and our investigations. Performance test records of Bavarian herdbook Pi gilts were used in the study of Tholen et al. (1998) to evaluate the validity of the Gruber formula for Pi, which had a 1.5% higher carcass lean content relative to the bellies of Pi gilts used in the present study. In addition, the belly lean content, as estimated by the Gruber formula of commercial slaughter pigs, was also validated on carcasses of Pi x German Landrace castrates (Tholen et al., 1998), whereas the current investigation was based on the performance records of Pi x F1 gilts.
Relative to the MRI lean content, distinctly lower values were observed for lean content of the belly estimated by AutoFOM (Table 2). The lean content of the belly projected by digital imaging of the belly considerably overestimated the MRI lean content of all breed and/or weight groups by 6.9 to 9.5%. This bias can be explained by the specific position (13th/14th rib) and definition (8-cm-wide belly cross-section) of the digital image interface of the belly.
Differences between MRI reference values and lean content of the belly estimated by the Gruber formula, digital imaging, and AutoFOM on difference composition classes of bellies are presented in Figure 3. The Gruber formula overestimated lean content of fat bellies and underestimated lean content of lean bellies. Belly lean content was considerably underestimated by the AutoFOM system, especially as MRI lean content increased, whereas the digital imaging device consistently overestimated belly lean content regardless of compositional class.
As demonstrated in Figure 4, the Gruber formula provided the highest accuracy in the estimation of MRI lean content for all breed and/or weight groups. The root mean square error (RMSE) was 0.5 to 0.9% lower, and the coefficient of determination was 15 to 31% higher, for the Gruber formula than for the digital imaging and AutoFOM methods. The accuracy of the Gruber formula was similar to that reported by Tholen et al. (1998), as ascertained by dissection of bellies into lean, fat, and bone components. Combining linear carcass measurements led to a RMSE of 2.58 and 2.55% for Pi and Pi x German Landrace pigs, respectively (Tholen et al., 1998).
View larger version (22K):
[in this window]
[in a new window]
|
Figure 4. Prediction accuracies of the different methods of estimating lean content compared to magnetic resonance imaging of bellies from Piétrain and Piétrain x [German Large White x German Landrace] (PiF1) gilts slaughtered at a mean carcass weight (CW) of 85 kg or 97 kg. I) Estimated carcass lean content: Bonner formula; II) AutoFOM lean content of the belly; III) digital imaging lean content of the belly, 13th-/14th-rib cut; IV) lean content of the belly estimated by the Gruber formula; V) adjusted estimation function using linear carcass measurements; VI) adjusted estimation function using linear carcass measurements and lean content of the belly estimated by digital imaging 13th-/14th-rib cut.
|
|
The relatively low accuracy in the estimation of MRI lean content by AutoFOM can be explained by the fact that: 1) almost none of the ultrasound measurements were taken directly at the belly region (carcasses passed the U-shaped transducer array lying on their backs, so the lean content of the belly could only be estimated by ultrasound measurements of backfat and loin thickness indirectly); 2) trimmed bellies in the present study comprised ends and sides, including teats; and 3) the investigations included only female pigs in a predefined weight range. Also, in the evaluation and verification of the AutoFOM regression equations by Brøndum et al. (1998) and Branscheid and Dobrowolski (1999), commercial pigs of both sexes were used, which had lower means and higher standard deviations in all relevant carcass composition traits. These explanations might account for the low correlations between the directly recorded and estimated AutoFOM belly weights (r = 0.27, 0.24, and 0.32 for Pi and PiF1 slaughtered at a mean carcass weight of 85 or 97 kg, respectively) as well.
In relation to the Gruber formula, a combination of different sources of information (linear measurements, digital imaging, and AutoFOM) led to a slight increase in prediction accuracy (Figure 4). Using only linear measurements decreased the RMSE by 0.05 (Pi) to 0.13% (PiF1, carcass weight = 97 kg). Adding digital image information led to a further reduction of 0.23 (PiF1, carcass weight = 97 kg) to 0.32% (Pi), whereas improvements were achieved by the additional integration of the AutoFOM lean content of the belly, but only in the Pi-breed.
Similar accuracies were accomplished by the PLS evaluation of the final regression equations; however, in contrast to the applied stepwise method, all traits were included with the exception of carcass weight and carcass length (Table 3). According to the VIP estimates (Figure 5), the carcass traits "thinnest fat measurement above loin muscle area" and "fat area, 13th-/14th-rib interface" were most important. The VIP for the traits "lumbar region backfat," "digital imaging lean content of the belly, 13th-/14th-rib interface," and "AutoFOM lean content of the belly" exceeded 1.0 for all groups. Hence, compared to the Stepwise procedure, the AutoFOM lean content of the belly played a more important role in the accuracy and the robustness of the PLS equations. However, the final equations include regression coefficients of some backfat measurements with biologically unexpected positive signs, which could limit the acceptance of the PLS formulas by practical breeders.
View this table:
[in this window]
[in a new window]
|
Table 3. Regression coefficients for estimating magnetic resonance imaging lean content of the belly using linear carcass measurements depending on different breed and/or carcass weight groups (85 or 97 kg)
|
|
View larger version (51K):
[in this window]
[in a new window]
|
Figure 5. Variable of importance (VIP) of carcass traits used in the partial least squares estimation of lean content by magnetic resonance imaging of bellies from Piétrain and Piétrain x (German Large White x German Landrace) (PiF1) gilts slaughtered at a mean carcass weight (CW) of 85 kg or 97 kg.
|
|
In order to improve the accuracy of the regression formula of the AutoFOM to predict belly lean content, adjusted equations were constructed using 127 basic AutoFOM-base recordings and PLS techniques. The calculated predicted residual sums of squares, depending on a different number of extracted factors for the analyzed different data sets, are exhibited in Figure 6. For all breed and/or weight groups, a minimum was achieved by the extraction of two factors.
View larger version (16K):
[in this window]
[in a new window]
|
Figure 6. Breed and weight specific "predicted residual sum of squares" statistics for the prediction of magnetic resonance imaging lean content of pork belies using the 127 basic AutoFOM recordings, partial least squares regression principles, and repeatedly performed cross validations using test and training datasets are affected by breed (Piétrain vs. Piétrain x [German Large White x German Landrace] [PiF1]) and mean carcass weight (CW; 85 kg or 97 kg).
|
|
In relation to the original formulas, the accuracy of the newly constructed functions to estimate belly lean content using the basic AutoFOM measurements was considerably improved (Table 4). The coefficient of determination was increased from 17.2 (PiF1, carcass weight = 97 kg) to 24.8% (PiF1, carcass weight = 85 kg), and the RMSE was reduced from 0.57 (Pi) to 1.07% (PiF1, carcass weight = 85 kg). Similar to the PLS analysis using linear carcass traits recorded on station, some of the relationships among basic AutoFOM fat and muscle depth recordings included in the final regression equation were unexpected; however, from an animal breeder’s perspective, the distinct improvements in accuracy demonstrate the potential of the AutoFOM device to verify the true carcass value of specific breeds.
View this table:
[in this window]
[in a new window]
|
Table 4. Accuracy to predict the magnetic resonance imaging lean content of the belly using original and adjusted regression equation depending on breed and/or carcass weight groups
|
|
|
Implications
|
|---|
Detailed information about the composition of carcass cuts is an essential prerequisite for efficient marketing of pork. Due to its complex muscle structure, it is particularly difficult to accurately estimate the composition of pork bellies. The results of this study suggest that linear carcass measures from stationary tested progenies provide sufficient information to improve the belly composition within the Piétrain breed, whereas a rather low level of accuracy can be expected when estimating the belly composition by fat and muscle areas obtained by digital imaging at a single belly interface or by the noninvasive, online AutoFOM carcass grading system. However, the accuracy of AutoFOM can be substantially increased by adapting the regression equation to include the 127 ultrasonic basic recordings. Besides the already affirmed advantages for the swine packing industry, these results indicate the potential benefit of AutoFOM for the performance testing regimens of pig-breeding companies.
Received for publication September 4, 2002.
Accepted for publication January 21, 2003.
|
Literature Cited
|
|---|
ALZ. 1999. Stationspruefung auf Mastleistung, Schlachtkoerperwert und Fleischbeschaffenheit beim Schwein. Mimeo. Zentralverband der Deutschen Schweineproduktion, Ausschuss fuer Leistungspruefung und Zuchtwertschaetzung, Bonn, Germany.
Baulain, U. 1997. Magnetic resonance imaging for the in vivo determination of body composition in animal science. Comput. Electron. Agric. 17:189–203.
Baulain, U., M. Henning, E. Tholen, W. Wittmann, and W. Peschke. 1998. Objektive Erfassung des Fleischanteils im Schweinebauch. 2. Mitteilung: Verwendung von Bildinformationen aus dem MR-Imaging. Zuechtungskunde 70:202–212.
Beuck, J. 1999. Abrechnungsmodell nach Handelswert-AutoFOM macht es möglich. Schweinezucht und Schweinemast 5:44–47.
Branscheid, W., and A. Dobrowolski. 1999. Evaluation of market value: comparison between different techniques applied on pork carcasses. Arch. Anim. Breeding 43:131–137.
Brøndum, J., M. Egebo, C. Agerskov, and H. Busk. 1998. Online pork carcass grading with the AutoFOM ultrasound system. J. Anim. Sci. 76:1859–1868.[Abstract/Free Full Text]
Evans, D. G., and A. J. Kempster. 1979. A comparison of different predictors of the lean content of pig carcasses. Anim. Prod. 28:97–108.
De Jong, S. 1993. SIMPLS: An alternative approach to Partial Least Squares Regression. Chemom. Intell. Lab. Syst. 18:251–263.
Fewson, D., W. Branscheid, and E. Sack. 1990. Untersuchungen ueber den Fleisch- und Fettanteil einzelner Teilstuecke und der Schlachthaelfte beim Schwein. Zuechtungskunde 62:38–51.
Götz, K.-U. 1999. BLUP-Tiermodellzuchtwertschaetzung beim Schwein in Bayern. Zucht-ziele. Available: h t t p : / / w w w .s t m l f . b a y e r n . d e / b l t / i n f o s/ s c h w e i n e d o k u / s c h w e i n e d o k u . h t m l . Accessed Dec. 9, 1999.
Hulsegge, B., P. Sterrenburg, and G. S. M. Merkus. 1994. Prediction of lean meat content in pig carcasses and in the major cuts from multiple measurements made with the Hennessy Grading Probe. Anim. Prod. 59:119–123.
Mitchell, A. D., A. M. Scholz, P. C. Wang, and H. Song. 2001. Body composition analysis of the pig by magnetic resonance imaging. J. Anim. Sci. 79:1800–1813.[Abstract/Free Full Text]
Pfuhl, K., and P. Glodek. 1996. Die Bestimmung des Fettgehaltes von Schweinebaeuchen mittels NIR und dessen Beziehung zu anderen Verfettungskriterien an der Schlachthaelfte. Zuechtungskunde 68:48–64.
Roberts, N., L. M. Cruz-Orive, N. M. K. Reid, D. A. Brodie, M. Bourne, and R. H. T. Edwards. 1993. Unbiased estimation of human body composition by the Cavalieri method using magnetic resonance imaging. J. Microsc. 171:239–253.[Medline]
Scheper, J., and W. Scholz. 1985. DLG-Schnittfuehrung fuer die Zerlegung der Schlachtkoerper von Rind, Kalb, Schwein und Schaf. DLG-Verlag, Frankfurt, D 85:22–26.
Schmitten, F., W. Trappmann, and H. Jüngst. 1986. Zuchtziel bessere Fleischqualität. Deutsche Geflgelwirtschaft und Schweineproduktion 30:896–897.
Scholz, A., U. Baulain, and E. Kallweit. 1993. Quantitative Analyse von Schnittbildern lebender Schweine aus der Magnet-Resonanz-Tomographie. Zuechtungskunde 65:206–215.
Tholen, E., W. Peschke, U. Baulain, and K. Schellander. 1998. Objektive Erfassung des Fleischanteils im Schweinebauch. 1. Mitteilung: Entwicklung von Schaetzgleichungen aus Schlachtkoerpermaßen. Zuechtungskunde 70:196–204.
Van der Voet, H. 1994. Comparing the predictive accuracy of models using a simple Randomization test. J. Chemom. Intell. Lab. Syst. 25:313–323.
Wold, S. 1995. PLS for Multivariate Linear Modeling. Page 195 in Chemometric Methods in Molecular Design. Vol. 2. H. van de Waterbeemd, ed. Verlag Chemie, Weinheim, Germany.
Top
Abstract
Introduction
), AutoFOM (shaded box), and digital image (
) of the interface at 13th/14th rib cut across different belly lean composition classes. The breed and/or weight groups include Piétrain and Piétrain x (German Large White x German Landrace) (PiF1) gilts slaughtered at a mean carcass weight (CW) of 85 or 97 kg.