Y. Bozkurt, S. Aktan and S. Ozkaya

Y. Bozkurt, S. Aktan and S. Ozkaya, 2008. Digital Image Analysis to Predict Carcass Weight and Some Carcass Characteristics of Beef Cattle. Asian Journal of Animal and Veterinary Advances, 3: 129-137.

DOI: 10.3923/ajava.2008.129.137

URL: https://scialert.net/abstract/?doi=ajava.2008.129.137

MATERIALS AND METHODS

---

Selection and Measurements of Carcasses and Carcass Traits
The carcasses used in this study were from a group of Holstein, Brown Swiss and their crosses slaughtered at a commercial beef packing plant (Burdur Province, Turkey). A total of 55 digital images and carcass measurements were taken such as Hot Carcass Weight (HCW), Carcass Area (CA), Carcass Length (CL), Carcass Depth (CD) and Longissimus Muscle Area (LMA) from different breeds of beef cattle. HCW was taken as carcasses hanging on the rail by a weighing scale with a digital display (kg). Rib eye area or Longissimus Muscle Area (LMA) images were taken twenty-four hours following slaughter after being cut between 11 and 12th ribs over the cross section of the ribbed surface with a camera. Each independent determination of a total of 29 LMA via plastic grid (clear plastic sheet with small dots arranged at the centre of a square 1.56 cm² in area) was recorded. In addition, independent tracings of each exposed LMA were obtained using acetate paper according to the procedures of Naumann (1951). After completion of data collection, three acetate paper tracings for each LMA were measured using a compensating polar planimeter to determine acetate/planimeter-traced LMA.

Digital Images and Image Analysis
Images were captured using a Canon digital camera MV850i. The camera was set on a standard quality (640x512 pixel resolution). Illumination conditions, location of camera and camera settings were tried to be the same and constant for all samples. Whole carcass images were taken by placing the reference card over each carcass (Fig. 1) and obtaining two sequential but separate images without moving the camera head unit in a fixed position perpendicular to the long axis of the carcass. Images collected by the instrument were obtained while carcasses were in a stationary position on the rail. In digital carcass images, CA was measured from the left side as the area around the whole carcass in cm²; CL was the distance from the point of the shoulders to the ischium (cm); in other words, from the sternum (manubrium) to the aitchbone (tuber ischiadicum); CD, from sternum area immediately caudal to the forelimbs to the top of the thoracic vertebra (cm). In digital LMA images (Fig. 2), LMA was measured as the area around the lean meat excluding the fat cover (cm). Digital images were downloaded from the camera to a computer file and processed using Image Pro Plus 5 software to obtain carcass and LMA measurements from the images. In order to calibrate the software a reference card with a known ruler (20 cm) was positioned next to the object such that the same distance and focus were kept when images were captured.

Fig. 1:	Digital Image of hot carcass together with reference card and processed by Image pro software

Fig. 2:	LMA measurements using acetate paper grid and processed by Image pro software

Statistical Analysis
Regression models were developed and assessed for prediction of Hot Carcass Weight (HCW) and some carcass traits using some carcass measurements as predictors. Descriptive statistics and regression analysis of HCW and LMA on each of the independent variables were performed using the GLM (General Linear Model) procedure of MINITAB, v.13 Inc. (2001). Pearson`s correlation coefficients were calculated between actual and predicted values obtained by image analysis. Linear, quadratic and cubic effects of independent variables on HCW were included in the following model:

Yi = b0+b1xi+b2xi2+b3xi3+ei

where:
Yi	=	HCW observation of an ith animal
b0	=	intercept
b1,b2, b3	=	corresponding linear, quadratic and cubic regression coefficients
xi	=	carcass measurements (CA, CL, CD)
ei	=	residual error term.

The difference between actual and predicted LMAs was examined by Students`t test using the statistical package program MINITAB v.13 for windows. The observed and predicted LMAs were also compared using the Mean-Square Prediction Error (MSPE):

where, n is the number of pairs of observed and predicted values being compared.

I = (1, 2, 3,......, n)

Oi is the observed LMAs with ith variable.
Pi is the predicted LMAs with ith variable.
The MSPE can be considered as the sum of three components described by Rook et al. (1990).

Where, S_O² and S_P² are the variances of the observed and predicted LMAs, respectively. and are the means of the observed and predicted LMAs, b is the slope of the regression of observed values on predicted and r is the correlation coefficient between O and P.

Apart from common regression analysis, MSPE has been used to determine the prediction ability of regression models and sources of error components in many studies by Smoler et al. (1998), Bozkurt and Ap Dewi, (2001), Yan et al. (2003) and Bozkurt (2006).

RESULTS AND DISCUSSION

---

The hot carcass and Longismuss muscle area digital images processed by Image Pro Plus 5 software (Fig. 1 and 2).

Descriptive statistics of HCW, CA, CL and CD obtained by image analysis used for prediction of hot carcass weight (Table 1), together with their means, standard errors and coefficient of variations.

The great variability shown by the HCW and CA values were expected on the account of variation of animal`s breed type, weight and size.

Results of multiple regression of hot carcass weight on digital measurements of CA, CL and CD used as predictors for prediction of HCW (Table 2), together with coefficient of determination (R²%).

The highest R² values (Table 2) were observed from the equation that contained all carcass traits (R² = 86.4%), the equation that included CA and CL measurements except CD (R² = 86.3%) and those equations that included CA and CD (R² = 86.3%), but the equation that included CL and CD gave R²=63.9%. Among the equations that included only one predictor; CA gave the highest R² = 85.9%, followed by CL and CD, respectively. The lowest R² value was obtained by CL (R² = 56.9%). Regression coefficients of CD in all equations were found non-significant except when CD used as single predictor, but still gave a very low R² value (Table 2).

Table 1:	Descriptive statistics for hot carcass weight and other carcass measurements obtained by digital images
SE: Standard errors; CV: Coefficient of variation

Table 2:	Multiple Linear regression equations to predict hot carcass weight using CA, CL and CD as predictors#
ns: statistically non-significant (p>0.05); # Only non significant regression coefficients had superscripts (ns), the rest were significant at p<0.05

Table 3:	Correlation coefficients of variables between HCW and other carcass measurements

Table 4:	Regressions of hot carcass weight on the linear, quadratic and cubic effects of each carcass measurements obtained by digital image analysis#
ns = statistically non-significant (p>0.05); # Only non significant regression coefficients had superscripts (ns), the rest were significant at p<0.05

In contrast to these findings, Cannell et al. (2002) reported that the VIA (video-image analysis) scan system predicted 72% of the variation in adjusted preliminary yield grade. Steiner et al. (2003b) reported that the CVS and VIA scan systems predicted 44 and 76% of the variation in adjusted preliminary yield grade, respectively.

Pearson`s correlation coefficients of variables between HCW and other carcass measurements, CA, CL and CD, obtained by the digital image analysis (Table 3).

All correlation coefficient values were found statistically significant. The highest correlation coefficient value (r = 0.93) was obtained between HCW and CA, followed by CL and CD (r = 0.78 and 0.75), respectively in relation to HCW (Table 3). Therefore, HCW was highly correlated with CA measurements. Results of regressions of hot carcass weight on the linear, quadratic and cubic effects of other carcass measurements obtained by the digital image analysis (Table 4).

The R² values from the regressions indicate that carcass area was the most highly related to hot carcass weight considering all linear, quadratic and cubic coefficient terms, 85.9, 86 and 91.3%, respectively (Table 4). Quadratic coefficients of all measurements were not significant and although for all measurements addition of cubic terms increased R² values considerably cubic coefficients of the measurements were very low, almost zero. Therefore, there was a highly linear relation between HCW and other measurements, CA, CL and CD.

Table 5:	Comparison between actual and predicted LMA values
SD: Standard Deviation

Table 6:	Mean Square Prediction Error and Proportions of MSPE (%)
ns = statistically non-significant (p>0.05)

In respect to linear equations, it was observed that 1 cm² change in CA resulted in 0.029 kg change in HCW. Similarly, 1 cm change in CL and CD resulted in 3.38 and approximately 6.5 kg respectively. It was obvious that a 1 cm² change in CA resulted in lesser hot carcass weight change compared to the rest of carcass measurements.

The acetate/planimeter measurements of LMA of a carcass side obtained by experts and the average of three independent tracings with a compensating polar planimeter were defined as "Aactual" measures of LMA and image-based LMA as "predicted" LMA. Image-based predicted LMA (99.3 cm²) was 0.2% lower than the actual LMA (99.5 cm²).

Coefficient of determination (R²), r and MPE were obtained as 85.6, 0.925 and 77%, respectively (Table 5). Mean bias (predicted minus actual) was negative (0.15 ± 3.28) and the difference between actual and predicted LMA was not statistically significant. There was a highly significant correlation between two values (r= 0.93). In this study, coefficient of determination value (R² = 85.6) found for LMA prediction is slightly lower than that of values found in some studies as Shackelford et al. (1998) found (R² = 88). This indicated that LMA of a carcass can be predicted with a high accuracy by digital image analysis.

The accuracy of the equations can be evaluated according to components of Mean Square Prediction Error (MSPE) as a proportion of the mean actual LMA. MSPE and its proportions (Table 6).

The MSPE of the predictions was 74.2 cm² in this study. In terms of contribution of components to MSPE; the values of bias, line and random error were 0.00, 0.2 and 99.7%, respectively (Table 6). The prediction model had a greater proportion of error derived from random than other components. A small proportion of bias as a component of MSPE showed that the error derived from bias was substantially low and there was a minimal variation between predicted and actual LMAs. The MSPE can be divided into three components due to mean bias (or mean deviation (-) of the prediction), Line bias (or deviation of the slope (b) of the regression of O on P from unity) and the random variation about this regression line. A positive mean bias indicates that the equations are generally overestimating relative to observed values while negative mean bias indicating underestimation. Therefore, there was a minimal underestimation of LMA values obtained by digital image analysis.

In regression equation to predict HCW using image based predicted LMA (HCW = 197+1.33LMA) R² value was obtained as 19.9% which was similar to the value (R² = 19%) obtained by Lu and Tan (2004). The correlation between predicted and actual LMA (r = 0.92) was similar to Schutte et al. (1998) r = 0.90; Cannell et al. (2002) r = 0.93; Gardner et al. (1995) r = 0.95 but higher than those of Steiner et al. (2003b) r = 0.83; Lu and Tan (2004) r = 0.60. Determination coefficient obtained in this study (R² = 85.6%) (Table 5) was slightly lower than those obtained by Shackelford et al. (2003) R² = 87%; Shackelford et al. (1998) R² = 88% and Steiner et al. (2003a) R² = 91%. However, it was higher than those obtained by ultrasound method by Perkins et al. (1992a) r = 0.60; Perkins et al. (1992b) r = 0.82; Greiner et al. (2003) r = 0.86, Suguisawa et al. (2003) r = 0.68; Waldner et al. (1992) r = 0.73; Jorge et al. (2004) r = 0.78. Therefore, LMA can be predicted by using digital image analysis with high accuracy and confidence.

CONCLUSIONS

---

It can be concluded that CA obtained by digital image analysis as only one parameter can provide a considerably reliable prediction of hot carcass weight. It is unavoidable that some images may not be clear enough for processing due to the lighting conditions inside the plant and improper position of reference cards placed on carcasses can make it difficult to measure correctly especially carcass areas on digital images. Prediction ability of the equations may also be affected by the variation of the slaughtered animal`s breed type and size.

Therefore, hot carcass weights and rib eye area (LMA) can be predicted by the digital analysis system with confidence and flexibility because the acceptable agreement and the close relationship between actual and predicted LMA gives general support to provide predictions of hot carcass weights of the slaughtered animals. However, there is still a need for further investigations for different breeds of cattle, taking in to account of their size as well under better controlled experimental conditions.