**4. Discussion**

Canadian mature cow grades (D grades) are assigned to one of the D1, D2, D3, or D4 grades depending on variables such as muscling (excellent, medium, or deficient), fat color (white or yellow), and fat measure (lower than, equal to, or higher than 15 mm) [2]. In the present study, similar numbers of carcasses for each grade were considered (26.2%, 25.0%, 23.8%, and 25.0%, respectively for D1, D2, D3, and D4). The ranges of the HCW, CCW, grade fat and fat thickness, LMY and RCY, REA, and marbling scores of the research carcass population used in the present study were representative of those found in the Canadian beef market [1].

The technologies used in the present study provided estimation values of the total amount of tissue and an overall description of the composition of the whole carcass and primal cuts without requiring the destructive procedure of dissection.

In the literature, most of the studies considering the use of CVS in beef carcass classification have focused on the quantification of LMY, RCY, and/or the total amount of fat, lean, and bone using CCC systems. Among others, Farrow et al. [31], Lu and Tan [32], McEvers et al. [10], and Shackelford et al. [33] used several variables obtained from the analysis of rib-eye images to define different regression equations to improve the accuracy, precision, and robustness of total tissue amount, LMY, or RCY estimations. The results reported by these authors (R<sup>2</sup> = 0.43–0.91) are within the range of those observed in the present study. All the authors agreed that CVS-related equations were an improvement on current prediction systems.

In agreemen<sup>t</sup> with the present study, Borggaard et al. [34] described similar R<sup>2</sup> values for total fat and RCY (%) using a BCC-2 camera and a HCC system, but carried out the statistical analysis by means of principal component analysis (PCA) and neural networks. Likewise, Pabiou et al. [35] used the VBS 2000 carcass grading unit (HCC) to predict carcass cut yields in cattle. Diverging from our study, HCC and CCW variables were used in the estimation models, and it was stated that stepwise regression showed slightly better R<sup>2</sup> values than the PLSR procedure, thus explaining 71%, 72%, and 75% of the variance for RCY (%), total fat (%), and total bone (%), respectively.

Vote et al. [36] compared BCSys (HCC) and CVS BeefCam (CCC) to study their potential as grading systems for Uruguayan beef carcasses. They reported higher RCY R<sup>2</sup> values for CVS estimations than for values from the USDA equation (values coming from graders). In agreemen<sup>t</sup> with the present study, for total fat and bone estimations, higher R<sup>2</sup> values were shown when using HCC or HCC + CCC technologies than when the CCC system was considered. In addition, in Vote et al. [36], bone amount estimations resulted in lower R<sup>2</sup> values than fat amount estimations, and HCW was also included in the models.

RCY and LMY values are commonly obtained from equations in which rib-eye and fat thickness measurements are considered [25]. Because the equations are built from these variables, it is not surprising that the CCC predicts RCY and LMY better than HCC, as the linear measures of rib-eye and rib-eye area along with the fat thickness obtained with the CCC are likely improving these estimates. Nevertheless, in the case of cull cows, it is possible that the industry could be more interested in lean to fat ratios, in which case, HCC predictions outperformed CCC. Hence, cows could be graded accurately in terms of lean/fat ratio using a camera system that does not require knife-ribbing.

The literature regarding the estimation of cattle primal composition is scarce. Using a dual CVS system (HCC + CCC), Cannell et al. [37] tested a total of 296 carcasses: 158 steers (103 light (HCW ≤ 339 kg) and 55 heavy (HCW > 340 kg)) and 138 heifers (51 light and 87 heavy), and described, in agreemen<sup>t</sup> with our results, R<sup>2</sup> > 0.65 for primal fabrication yields on average using a selection of HCC and CCC variables (higher coefficient of correlation), HCW, and stepwise regression (higher R2). Using a similar statistical approach, the HCC system, including the CCW variable (VBS 2000), Pabiou et al. [35] defined four cut-out groups according to their retail value (low, medium, high, and very high value) and obtained higher R<sup>2</sup> values (0.84, 0.65, and 0.87) than in the present study for wholesale primal weights for steers, heifers, and bulls, respectively. In turn, Craigie et al. [11] used VBS 2000 technology (HCC system) and described R<sup>2</sup> > 0.80 for the estimation of saleable (retail cut) sirloin weight, considering HCW in the regression models.

In other species, Rius-Vilarrasa et al. [38], using VSS 2000 (HCC system for lambs) and PLSR statistical analysis, reported R<sup>2</sup> values of 0.86 for breast and 0.96 for leg primals. Lorenzo et al. [39] reported R<sup>2</sup> values between 0.53 and 0.89 for the prediction of foal carcass composition and wholesale cut yields using HCC. Nevertheless, CCW was also considered as a describing variable in the prediction models, whereas HCW was used in the present study. The CCW has been described as a good estimator in the case of lamb carcasses [40]; however, its suitability has been questioned for cattle [35].

Kipper et al. [17] assessed the accuracy of the methodology using the concepts of trueness, defined in our case as the degree of agreemen<sup>t</sup> between the dissection and the instrumental estimation values, and precision, as indicative of the degree of internal agreemen<sup>t</sup> (dispersion). In addition, the trueness was considered to be the sum of ECT and ER; precision was associated with ED and overall accuracy was related to MSPE [17]. Paying attention to error parameters, higher ECT values in CCC and HCC + CCC than in HCC were detected in fat estimates, whereas the opposite behavior was observed for ED. Therefore, the similar values of R<sup>2</sup> and high values of ED imply that the three instrumental approaches could be considered highly accurate and precise, the PLSR analysis being suitable for estimation. However, CCC and HCC + CCC fat estimations showed lower trueness than HCC estimations (Table 2).

In agreemen<sup>t</sup> with the present results, the feasibility of DXA technology in assessing carcass composition has been stated for broiler chickens [13], pigs [15], and sheep [14], and good R<sup>2</sup> values have also been described for calves [16,18]. Aligning with our results, in all these studies, higher R<sup>2</sup> values were described for total fat and total lean estimations than for total bone estimations.

<sup>L</sup>ópez-Campos et al. [18] described similar results for the estimation of fat, lean, and bone mass of primals using DXA with youthful cattle. The basis of the DXA technology lies in the different absorption ratios from a low and a high energy X-ray beams when interacting with the tissues. The software estimates the mass of two different tissues at each scanned voxel; therefore, it is possible to differentiate between fat and lean when no bone is present but, where the sample matrix contains bone, fat, and lean, the mass fraction can only be established as bone and soft tissue, with the individual measurements of fat and lean obtained from other regions of the scan. Therefore, the higher the amount of bone detected, the more difficult the differentiation between fat and lean. In addition, a medical DXA unit has been used in the present study, thus being calibrated for the measurement of human bone mineral content and bone mineral density, but not for the whole bone content of livestock.

Again, the fact that ED explained more than 99% of the MSPE values would imply that the differences among R<sup>2</sup> values were highly related to the dispersion (precision) and poorly related to the trueness. Therefore, the external factors such as calibration method, software analysis, the defining variables considered for the estimation models (HCW, CCW, marbling, color, gender, etc.), and the meat cutters' decision making (tissue differentiation and cutting) would be the defining variables of dispersion. Accordingly, the low R<sup>2</sup> value of flank bone might be a consequence of the low amount of bone included in this primal, thus implying a high variability in both DXA estimations and weight measurements. The foreshank showed the highest bone to soft tissue ratio, although the different contributions

to lean or fat estimates remain unclear: higher and lower accuracy of lean and fat estimates of foreshank, respectively.

Regarding bone prediction, similar results to those from this study were described in pigs [41] and chickens [42]. Kipper et al. [41] and Schallier et al. [42] described better R<sup>2</sup> values when the predicted amount of bone was correlated with ash content. This implies that the presence of small pieces of lean or fat that were adhered to the bone decreases the accuracy and precision of the analysis, whereas it increases the error between actual and predicted values.

Similarly to DXA, computed tomography (CT) is a technology also based on X-ray attenuation. Prieto et al. [43] described lower R<sup>2</sup> values for IM and total fat than for SQ and total lean predictions (0.77, 0.89, 0.94, and 0.99, respectively) using spiral CT. Concurring with the present study, Navajas et al. [44] described lower R<sup>2</sup> values for carcass total bone than for fat and/or lean estimates when using CT technology (R<sup>2</sup> = 077, 0.92, and 0.96). In addition, Navajas et al. [45] described R<sup>2</sup> values of 0.92, 0.99, and 0.97, respectively, for fat, lean, and bone for the primal estimations.

To date, DXA has been limited by practical constraints for deployment in the industry (horizontal table scans, operation at room temperature, and rate of scan in minutes rather than seconds). However, Scott Technologies Ltd. (New Zealand) has developed an upright DXA scanner, capable of scanning at a rate of 540 lamb carcasses per hour while maintaining performance accuracy. This technology adaptation was originally used to mark anatomical features to program robotic cutting. The technology is now being envisioned as a means of lean yield prediction in beef and lamb plants in Australia and New Zealand.

May et al. [46] reported that estimated yield differences could be attributed partially to differences in seam fat deposition (different fat deposition along the carcass). Likewise, in practice, the fabrication of the boneless, closely trimmed round, loin, rib, and chuck retail cuts is performed manually by meat cutters, thus implying another subjective source of variability. Although these factors might introduce variations in the cutability, the present results sugges<sup>t</sup> that both CVS and DXA technologies have the potential to estimate beef carcass traits such as total or retail cut yield performance.

Finally, it is worth mentioning that, based on its performance, DXA might be seen as the gold standard candidate technology for carcass composition estimation. Currently, DXA technology is still under development and it is also being used as a means of envisioning bone location for robotic carcass fabrication. The costs and other operational factors are limiting its industrial implementation. However, if a facility had the capabilities to set up both camera systems, and knife rib cows at the 12/13th, combining the HCC and CCC data, could result in prediction accuracies very similar to DXA. This approach would be of benefit to the plants in determining which carcasses would be profitable for specific fabrication lines.
