As the U.S. beef industry continues to advance and evolve, areas of efficiency are under increasing scrutiny and regulatory pressures. A large area of focus is on measures of beef output on an individual animal basis. Yield Grade is the standard currently utilized for estimating the expected yield of boneless, closely trimmed retail cuts, but investigation of technologies or tools that predict carcass yield better than Yield Grade is warranted given the changes within industry. Visual human appraisal of livestock composition has long been used, trusted, and valued in the U.S. livestock industry, but it will always be a subjective measure, even among the most experienced evaluators. Very few studies have evaluated the ability of whole-body measurement or imaging technology to mimic human appraisal methods in predicting carcass composition in live beef cattle. The “gold standard” methods traditionally used to measure carcass composition are considerably expensive and laborious, especially when conducted with enough sample size to be commercially meaningful. Computed tomography (CT) has been shown highly accurate at determining carcass composition of sheep and hogs, although less work with CT has been conducted in cattle. The objectives of the study were to evaluate computed tomography (CT) as a method to measure carcass composition and to use millimeter radar measurements of live cattle to predict CT-determined carcass composition in real-time.
Cattle (N = 98) of varying of varying sex, live weight, fatness, and muscularity were scanned using a millimeter radar linear body measurement system immediately before harvest at the WSU Meat Laboratory. Carcasses were chilled, and standard USDA Yield Grade and Quality Grade data were obtained. Carcass left sides were separated using reasonably representative industry breaks into 9 primal and were CT scanned so images obtained were at a cross-sectional thickness of 8 mm. Volumes were calculated from CT images and computed to masses within each primal, and Hounsfield unit (HU) thresholds were established between muscle, fat, and bone.
A subset of rounds (N = 42) was dissected into soft tissue and bone. The mass of fat and fat-free lean (muscle) was determined from a chloroform-methanol extraction method. Chemically determined fat and muscle masses and dissectible bone mass were used to test the performance of CT in measuring tissue composition of the dissected rounds. Machine learning techniques were used to develop a model predicting carcass muscle percentage from body measurements obtained from the radar system. A series of 70 pre- 38 determined body dimensions were calculated from millimeter radar measurements. A collective subset of body dimensions was identified as being most contributive to percent muscle using a Boruta algorithm. Data were split in to 50% training and 50% testing datasets, and a repeated via cross-validation.
Chemically determined and CT-measured percentages were highly related for fat, muscle, and bone of the dissected rounds, indicating a high level of precision in the CT measurement of carcass tissues. Chemically determined and CT-measured percentages exhibited a high level of accuracy for each tissue type. All relationships appeared to be linear in a nearly 1-to-1 ratio across the range of values. Although all carcass regions were not dissected, CT-determined and actual primal weights provided the ability to assess the performance of CT in carcass regions other than the round. The mean percentage difference in CT-determined and actual weights in each of 9 primals was not more than 1%, and the RMSE of the percentage difference was not more than 1.5% for any primal, indicating a high level of precision as well as accuracy. Since tissue proportions were included to calculate the mass of each primal, it could be deduced that CT is also reasonably accurate at measuring tissue masses within primals, although this was not evaluated in all primals.
Body dimensions selected for their collective relationship to the target variable of percent muscle category generally included body widths, at high and low locations, on the anterior (heart girth) and posterior (spring of rib) ends of the rib cage. On the training dataset, the model classified 63% of samples into their correct percent muscle category and 76% of samples within one category (or 2 muscle percentage units) of their correct category. The application of the model to data in the testing dataset essentially replicated the application of the model to “blind” or “unseen” data, just as would happen if the radar system were implemented in commercial practice for carcass composition determination. In the testing dataset, the model classified 24% of samples into their correct percent muscle category, and 49% of samples within one category (or 2 muscle percentage units) of their correct classification. Differences in radar-predicted and CT-measured muscle percentage values across the entire dataset (only n = 34 cattle used in training the model; n = 62 “unseen” cattle) showed that 64% of all cattle could be predicted within 3 percentage units of their CT-measured muscle percentage. Additionally, the relationships between radar-predicted and CT-measured muscle percent indicated that the model had reasonable predictive precision.
This study demonstrated that computed tomography can be used to measure composition of fat, muscle, and bone in beef carcasses and cuts with a high degree of accuracy and precision, such that it might be evaluated as a gold standard method for carcass composition determination. Body dimensions of live cattle rapidly measured pre-slaughter using millimeter radar could be used to develop a relatively robust model capable of predicting muscle percentage in carcasses at a commercially meaningful accuracy. Further exploration and development of such a model for predicting carcass composition with equal or better accuracy and precision would help the beef industry more adequately quantify and manage its outputs to optimal targets
ARMS#120525-22