**3. Results**

### *3.1. Cow Carcass Population*

All the carcasses used in the present study showed ossification processes at the caps of the thoracic vertebrae ranging from 50% to 100% ossified, resulting in carcasses graded as Canada D mature type grades [2]. Values of HCW (277.3–410.2 kg), CCW (271.3–401.9 kg), grade fat (0.0–29.0 mm), fat thickness (0.0–27.9 mm), REA (60–120 cm2), LMY (49.0–61.0%), RCY (42.9–54.5%), and marbling scores (100–733, USDA marbling score), of the carcass population (*n* = 111) used were within the actual range (Table 1) of the Canadian beef carcass market [1].

**Table 1.** Descriptive statistics of carcass characteristics of the population used to obtain the prediction equations between the camera vision system values and whole carcass and primal composition (fat, lean, and bone).


1 SD: standard deviation; 2 HCW: Hot carcass weight; 3 CCW: Cold carcass weight; 4 Rib-eye width and length, and muscle score in agreemen<sup>t</sup> with Jones [24] and Segura et al. [25]; 5 LMY: estimated total lean meat yield [25]; 6 RCY: retail cut yield [25]; 7 Marbling scores: Official United States Standards for Grades of Beef Carcasses (marbling scores: 0 = Devoid, 100 = Practically Devoid, 200 = Traces, 300 = Slight, 400 = Small, 500 = Modest, 600 = Moderate, 700 = Slightly Abundant, 800 = Moderately Abundant, 900 = Abundant) [23]; 8 Ossification (%): Ossification processes of the carcasses assessed on the caps of the spinal processes and ribs (i.e., >50% ossification, a carcass receives a D grade) according to <sup>L</sup>ópez-Campos et al. [22] and the Canadian beef Grading Agency [2].

### *3.2. Primal Weight Estimation*

Overall, CVS lean and fat predictions (Table 2) showed high R<sup>2</sup> values for most of the primal cuts, while R<sup>2</sup> values for bone were much lower. The HCC had similar performance to the CCC for fat predictions in the primal cuts, ranging from R<sup>2</sup> = 0.47 to 0.88 compared to R<sup>2</sup> = 0.51–0.92, respectively. More specifically, fat foreshank regression models showed the lowest R<sup>2</sup> values, while plate, rib and round showed a 14.4% improvement with the HCC compared to CCC. Likewise, lean weight predictions in the primal cuts were superior using the HCC compared to the CCC, except for the rib with R<sup>2</sup> ranging from a low of 0.53 in the foreshank to a high of 0.90 in the round for the HCC, and 0.32 in the foreshank to 0.69 in the rib for the CCC. Additionally, the HCC outperformed the CCC in the prediction of bone weight, showing R<sup>2</sup> values as high as 0.79 in the round, while the highest R<sup>2</sup> for the CCC bone weight was 0.38 in the chuck. Neither of the camera systems studied was able to accurately predict flank bone weight, and the CCC could not either accurately predict bone weight in the brisket, loin, rib, plate, or foreshank (R<sup>2</sup> < 0.10, LV = 1 related variable was considered for prediction equation development).

When considering the error and/or variance partitioning, no remarkable differences were found for bone estimations. Nevertheless, CCC showed ECT values 97.4% higher and ED values 5.8% lower than HCC for fat primal estimations, although no difference was observed for MSPE. In the case of lean primal estimations, MSPE value for CCC resulted 63.2% higher than for HCC, although ED values for HCC resulted only 0.7% higher than CCC, and 81.7% of the difference was due to ER instead of ECT (24.4%) when compared to CCC (Table 2).

The combination of both CVS technologies did not improve fat estimations for most of the primals; only the round predictions (R<sup>2</sup> = 0.88) showed some improvements compared to the HCC or CCC estimations; 3.4% and 18.2%, respectively. Conversely, lean estimations of brisket, chuck, plate, and rib showed, respectively, a 10.8%, 3.3%, 10.1%, and 16.2% higher R<sup>2</sup> values for HCC + CCC than for HCC. Additionally, HCC + CCC improved bone estimations in the case of brisket (R<sup>2</sup> = 0.42), chuck (R<sup>2</sup> = 0.71), and loin (R<sup>2</sup> = 0.76) compared to the individual CVS.

In the HCC+CCC, the contribution of ED to MSPE value was again much higher than the inputs coming from ER and ECT values (Table 2). For fat estimations, HCC + CCC showed lower ED and ER values and higher ECT values than HCC, and lower ECT and ER but higher ED values than CCC. In the case of lean estimations, HCC + CCC showed MSPE similar values to HCC, but these were lower than CCC. The ED values were lower than HCC and similar to CCC. The ER values were again higher than ECT, as observed in the HCC to CCC comparison. In addition, no remarkable differences were found for bone estimations. Interestingly, for fat estimations, the LV number for HCC + CCC was lower than for HCC or CCC.

In contrast, DXA primal estimations (Table 3), on average, had R<sup>2</sup> values for fat (0.95), lean (0.97), and bone (0.82) higher than those for CVS, and even outperformed the prediction equations utilizing all camera variables (HCC + CCC; Table 2). Except for the foreshank fat weight (R<sup>2</sup> = 0.74), DXA lean and fat weight predictions for the rest of the primals showed R<sup>2</sup> values between 0.94 and 0.99 and 0.96 and 0.99, respectively. Similar to the CVS, lower values of R<sup>2</sup> were observed for bone than for fat and/or lean variables; however, even flank bone weight (R<sup>2</sup> = 0.31) was predicted more accurately using DXA than by using both camera systems combined. For the other primal bone weight predictions, DXA R<sup>2</sup> ranged from 0.85 to 0.94, whereas the combined camera R<sup>2</sup> values ranged from 0.36 to 0.76. Overall, there were improvements in most tissue primal predictions using DXA when compared to the camera systems. On average, for all the primals, there was an overall proportional improvement in DXA R<sup>2</sup> values of 26.0%, 48.9%, and 24.8% compared to HCC, CCC, or HCC + CCC, respectively, as well as an increase in the DXA R<sup>2</sup> values of 16.0%, 29.0%, and 54.7% for fat, lean, and bone estimations, respectively. The MSPE showed relatively low values and was defined by ED in a percentage higher than 98.7% for fat estimations, and higher than 99.8% for lean and bone tissue estimations (Table 3).



[26]. 5 No statistically significant regression model (*p* > 0.05) was obtained. LV = 1 was considered to establish a prediction equation.

**Table 3.** Partial least square regression models estimating fat, lean, and bone for individual primal cuts from dual-energy X-ray absorptiometry (DXA) values (n = 111). Coefficient of determination (R2), mean square prediction error (MSPE), error in central tendency (ECT), error due to regression (ER), error due to disturbances (ED), and the number of latent variables (LV) are presented for each model.


1 Primals according to Institutional Meat Purchase Specifications (IMPS) for Fresh Beef Products, Series 100 [26].

### *3.3. Overall Carcass Tissue Composition and Yield Estimations*

Overall, relatively high R<sup>2</sup> values (>0.75) were obtained between the estimations with the different technologies and the actual dissection values and yield equation estimates of LMY and RCY. Particularly, high relationships (R<sup>2</sup> > 0.80) were observed between the estimations with DXA and HCC and the actual dissection values (Table 4). With the exception of LMY (R<sup>2</sup> = 0.66 vs. 0.85), the HCC had similar or higher predictions for overall total carcass composition than the CCC (Table 4). In particular, the HCC predicted fat weights similar to (R<sup>2</sup> = 0.92 vs. 0.93) and lean weights (R<sup>2</sup> = 0.89 vs. 0.67) and bone weights (R<sup>2</sup> = 0.82 vs. 0.31) better than the CCC camera. In fact, the HCC performed similar to DXA for all the total carcass composition estimates (R<sup>2</sup> > 0.80), and only dropped in prediction accuracy for the LMY and RCY (R<sup>2</sup> = 0.66 and 0.68 for the HCC and R<sup>2</sup> = 0.81 and 0.86 for the DXA). Adding the CCC variables to the prediction (HCC + CCC) resulted in very similar prediction accuracies to those of DXA for all overall total carcass composition, including the estimates of LMY and RCY.

**Table 4.** Partial least square regression models estimating total fat, lean, and bone amounts, and total subcutaneous (SQ), body cavity (BC), and intermuscular (IM) fat amounts for whole carcass sides and total lean meat yield (LMY) and retail cut yield (RCY) from dual-energy X-ray absorptiometry (DXA) and computer vision system (CVS) values. Coefficient of determination (R2), mean square prediction error (MSPE), error in central tendency (ECT), error due to regression (ER), error due to disturbances (ED), and the number of latent variables (LV) are presented for each model.


HCC + CCC = regression models obtained using the variables from both CVS systems.

In general, DXA estimations showed lower MSPE values than any CVS (2.0% vs. 14.6% on average, respectively), and, particularly, higher MSPE values were observed for the CCC procedure than for the HCC or HCC + CCC ones (21.4%, 11.8%, and 10.5% on average, respectively). Besides, IM, total fat, and total lean estimates showed the highest MSPE values, whereas BC fat showed the lowest. Implicating MSPE components, 88% for ED input was observed for the estimation of total fat using CCC variables, while values higher than 90% were observed for the others.
