Beef carcass yield and composition prediction: Integrating 3D imaging, dual X-ray absorptiometry, computerized tomography, and machine learning

This study evaluated multiple methodologies to predict salable yield and estimate bone weight in beef rounds using traditional regression models, machine learning, and novel machine learning approaches. Traditional models, including multiple linear regression (MLR), explained only 17% of the variation in subprimal yield (SY) using hot carcass weight (HCW), ribeye area (REA), and backfat thickness (BF) as predictors. Including cattle type improved model performance (R² = 0.61). Similarly, a neural network model incorporating cattle type achieved an R² of 0.61, indicating that within cattle type the relationship between these predictors and SY is essentially linear. Deep learning models using data from DXA-derived pixel matrices failed to achieve meaningful predictions, with a negative R² in the validation set. While an XGBoost model trained with DXA data achieved an R² of 0.67, but its performance was inconsistent with negative R² values when evaluated at the primal or carcass side level. 3D imaging incorporating geometric features from carcass cross-sections into XGBoost models (R² = 0.07). slightly outperformed the same model trained with current yield grade predictors (negative R2). However, a novel MeshGNN model using 3D mesh data achieved an R² of 0.60 despite the limited sample size, matching the best-performing traditional model adjusted for cattle type. These findings demonstrate that 3D imaging combined with machine learning has strong potential for improving yield predictions, as MeshGNN achieved comparable performance to traditional methods adjusted for cattle type despite a significantly smaller dataset. While DXA shows promise as a predictive tool, larger sample sizes are needed to enhance model performance.