The animal trials described below were approved by the Animal Experiment Committee of SRUC and were conducted in accordance with the requirements of the UK Animals (Scientific Procedures) Act 1986.

The 3D cameras used were Basler Time-of-Flight near infra-red cameras (Basler Inc., Exton, PA). The camera specifications are as follows: 640 × 480 pixels, 20 frames per second, 57° horizontal × 43° vertical angular field of view, accuracy of +/– 1 cm. Eighteen measurements (5 widths, 6 lengths, 5 heights, and 2 diagonals, Figure 1) were extracted from each 3D image and 20 ratios, 11 areas, and 11 volumes were calculated, giving a total of 60 potential predictor variables available for evaluation. Measurements were extracted in real time from 3D images using algorithms developed by Innovent Technology Ltd. using Halcon software (MVTec Software GmbH, München, Germany).

Live animal data was gathered from a range of sources: including both commercial and research farms and from an abattoir lairage.

Cattle were stunned by captive bolt, exsanguinated and their hides were removed. Carcasses were split down the midline and dressed as per normal abattoir practice. Conformation class and fatness class were visually assessed for each carcass by trained abattoir staff (according to the abbreviated EUROP grid commonly used in UK abattoirs). VIA technology (VBS 2000, E+V GmbH, Germany) was operated on-line to predict fat and conformation grades on both the 15 point scale and the EUROP grid (7 fat and 8 conformation grades). Cold carcass weight, saleable meat yield (SMY) estimated by VIA along with visually assessed EUROP fat and conformation grades were provided by the abattoir. Carcass characteristics data for a total of 1649 carcases from both the abattoir and on-farm trial datasets were matched to clear pre-slaughter 3D images, see Table 2 for a breakdown of breeds and sexes.

Data from all abattoir and on-farm sources were combined into one dataset. For the LW predictions the abattoir data consisted of a single LW per animal taken immediately pre-slaughter. The commercial and SRUC on-farm trial data consisted of multiple weights per animal across the finishing period. For the fat grade, conformation grade, CCW and SMY predictions, only the final LW recorded in the beef monitor crates on farms was used alongside the LWs collected in the abattoir trial. No 5L or 5H fat grades and no E and insufficient U+ conformation grades (n = 5) were recorded and so these grades could not be included in the prediction model. A summary of the breeds, sexes, fat grades and conformation grades are shown in Table 2.

Sex was included as a factor in the model. Cattle were categorised as either native type (smaller, quick finishing breeds such as Aberdeen Angus) or continental type (larger breeds such as Charolais) (see Supplementary Table 1 for categorisation of breeds), and this was also included as a factor in the model. From the commercial farm trials, the final measured LW from the weigh crate was included as a predictor variable for carcass characteristics.

Artificial neural networks (ANNs) were selected for this study as they can be used for both regression and classification problems and are capable of handling complex non-linear relationships between large numbers of variables. ANNs comprise a framework of “neurons” which are connected by weighted links (Agatonovic-Kustrin and Beresford, 2000). ANNs can be used for regression and classification problems and have many applications in financial forecasting, machine vision, game theory, medicine and ecology to name only a few. ANNs were developed using the caret package in R (version 3.4.1, R Core Team, 2017). To optimise neural network training, continuous input variables were standardised using a Gaussian transformation (subtracting the mean and dividing by one standard deviation) and min-max scaling between −0.9 and 0.9. The data was then randomly split into training (70%) and validation (30%) subsets.

In this study ANNs were developed through supervised training by backward propagation. The model was presented with the training set and known target values. Weights and biases were automatically randomly initialised to non-zero values (between 1 and −1) by the ANN software and during the training phase the model adjusted the weighted connections by feeding back the error and optimising the weights to decrease the difference between target and output values. Repeated training iterations (three repeats of 10-fold repeated cross validation) further reduced the model error. Models were regularised to prevent overfitting to the training data subset by applying a penalty (a weight decay value) to weights which became relatively much larger than others in the model. Parameter estimation (model size and weight decay values) were optimised after testing 100 potential models (10 possible values per parameter). Several topographies (number of hidden layers and nodes in each layer) were tested for each ANN. The topography which produced the best performance results without overfitting to the training data sub-sets was selected for each ANN. All of the ANNs had one hidden layer with five nodes, except the fat grade classification ANN which only had one node in the hidden layer. The model was then tested on the validation data subset. Model performance was assessed by R² and RMSE for regression (LW, CCW, and SMY). Classification accuracy for fat and conformation grades were assessed by way of confusion matrices. A confusion matrix is a table summarising the number of validation sub-set data points in each class and the predicted classes, and the sensitivity (Equation 1) and specificity (Equation 2) for each class.

Where for any class (x), a true positive is a data point that is correctly predicted to be within class x, a false negative is a data point incorrectly predicted to not be in class x, a true negative is a data point which is correctly predicted to not be in class x and a false positive is a data point which is incorrectly predicted to be in class x.

Stepwise linear regression models were also created for the continuous variables (LW, CCW, and SMY) using the same training and validation data subsets as were used to create the ANNs and were cross validated using the same method. Summary results (R² and RMSE) are reported alongside the ANN results.

Finally, the importance of each predictor variable to the overall ANN was assessed using the VarImp function in R. This function calculates the influence each input variable has on the output by using the connection weight between the input and each hidden neuron and apportioning the connection weights between each hidden neuron and the output between each input variable (based on the method described in Gevrey et al., 2003). Connection weights are analogous to coefficients in a linear model (although the number of connection weights in an ANN is excessive compared to coefficients in a linear model) and so dictate the influence any variable has on the hidden nodes and ultimately on the output e.g., variables with low weights are suppressed and so have little importance and those with large weights are influential and have high importance. Variable importance was scaled from 100 to 0 with 100 being the predictor variable with the highest calculated influence and 0 being redundant. The following calculations below are quoted from Gevrey et al. (2003).

For h = 1 to nh, and for i = 1 to ni

For i = 1 to ni

Where Q is the proportional influence an input neuron has on a hidden neuron, h is a hidden neuron, i is an input neuron, W is a weight and RI is the relative influence of an input neuron (%).

A total of 18,134 3D images were collected during this trial. Of the 16,100 3D images collected on commercial and research farms 1,292 (8%) of images were removed due to a poor outline being obtained. From the abattoir trial 550 of 2,034 3D images (27%) were removed from the analysis. The more stressful environment in the abattoir lairage led to a higher proportion of 3D images being removed from the analysis. Animals were more likely to be agitated and so a good quality 3D image was difficult to obtain. Removal of images from the on-farm data sets is not deemed to be a concern for commercial implementation as multiple images are collected per animal per day; therefore not all images of each individual animal are required to provide a prediction to the end user.

Pre-slaughter LW's ranged from 341 to 774 kg and the mean weight at slaughter was 608 ± 57 kg. The mean CCW was 339 ± 39 kg and mean SMY was 223 ± 32 kg.

In this study LW was predicted for a wide variety of breeds, both steers and heifers, with an R² of 0.70 (RMSE = 42, n = 4443, Figure 2). The performance of the stepwise linear regression for LW was much poorer than the ANN (R² = 0.54, RMSE = 51). Ozkaya et al. (2016) used multiple linear regression of measurements extracted from lateral 2D digital images of Limousin cattle to predict LW with an R² of 0.89. Although sex and breed type had low importance (3 and 0, respectively, Table 3), to investigate the performance of sex and breed specific models the ANN was trained only using the Aberdeen Angus steers data subset (n = 441, Table 1). The model performance increased to R² = 0.77 (RMSE = 37), suggesting that the further development of this system may benefit from breed and sex specific models. As LW had the highest importance (100) for the prediction of CCW, SMY and fat grade, and the importance of sex (CCW: 51, SMY: 29, conformation grade: 1, fat grade: 11) and breed type (CCW: 32, SMY: 15, conformation grade: 18, fat grade: 32) are generally of higher importance for prediction of carcass characteristics (Table 3), breed and sex specific LW models should also improve prediction of these carcass characteristics.

Carcasses which are over a defined weight face a penalty at the abattoir. Being able to predict CCW in the live animal would allow producers to ensure that animals are sent to slaughter before they grow beyond the weight limit. The ANN predicted CCW with R² = 0.88 (RMSE = 14, n = 449, Figure 3) and SMY with R² = 72 (RMSE = 14, n = 448, Figure 4). The stepwise linear regression models predicted CCW with R² = 0.83 (RMSE = 16) and SMY with R² of 0.63 (RMSE = 16). LW was of most importance in the ANNs for CCW and SMY in this study (Table 3). LW has previously been shown to have a strong linear relationship with CCW (Minchin et al., 2009), hot carcass weight (Pogorzelska-Przybylek et al., 2014), and SMY (Realini et al., 2001; Greiner et al., 2003). However, predictor variables extracted from the 3D images still had significant influence over the ANN model outputs (Table 3), and the ANNs had improved performance over the stepwise linear regression models. Greiner et al. (2003) found that when LW was used as a single predictor for SMY their regression model had an R² of 0.66 for a more limited range of animals (534 cross-bred steers) than used in the present study, demonstrating the potential of 3D imaging to provide more accurate predictions of carcass characteristics.

Farmers in the UK are currently paid for their animals on both carcass weight and fat and conformation grades. ANNs were developed for fat and conformation grade using the abbreviated EUROP scale in operation at the abattoir. The accuracy of the classification ANNs for the validation data subset were 54.2% for fat grade and 55.1% for conformation grade. The confusion matrices are shown for the fat (Table 4) and conformation grades (Table 5), along with the sensitivity (ability of the model to correctly classify a data point to that particular grade) and specificity (ability of the model to correctly identify a data point as not belonging to that particular grade) of the model to each grade.

The majority of carcasses were classed as fat grade 2 (47%) or 3 (34%) (Table 2). The fat grade model had a sensitivity of 0.83 for grade 2, but a specificity of 0.46 (Table 4). This low specificity was due to the tendency of the algorithm to classify the grade 3 carcasses as grade 2. The model classified all of the grade 1 carcasses in the validation subset as grade 2 and most of the 4H carcasses as 4L. It did not correctly classify to either grade 1 or 4H (sensitivity equal to 0, Table 4), this was likely due to there being insufficient data points in the training set for these two grades. The specificity of the conformation grade classification ANN model to both P+ and U+ was 1 (Table 5). There were also only a small number of data points collected for carcases of these grades. There was a tendency for the model to classify the O- and R carcases as O+ (O+ had a specificity of 0.56), likely due to the relatively large number of data points in the training data set which were grade O+. It is anticipated that increasing the number of data points in the less desirable grades would improve the predictive performance of these models.

Lambe et al. (2010) used ultrasound measurements of tissue depth in live finishing beef steers and heifers to predict conformation and fat grades using linear regression. The predictions in their study were slightly more accurate (R² = 0.60) for fat grade and similar (R² = 0.56) for conformation class than in the present study, however their models performed poorly on validation data sets (fat class: R² = 0.39–0.46, conformation class: R² = 0.07–0.24). SMY has also been successfully (R² = 0.80) predicted using similar ultrasound measurements (Realini et al., 2001). No literature could be found where a classification model had been used to predict fat and conformation grade of beef carcases. The advantage of a 3D imaging system over manual measurements such as ultrasound are the reduction in stress caused by handling of animals and the automated system can passively provide multiple estimates per animal per day at minimal cost.

In this study LW was found to be the most important predictor of fat grade (weighted importance of 100, Table 3), and was less important, but not redundant (weighted importance of 30) for conformation grade. Minchin et al. (2009) found that LW was not a significant predictor of fat or conformation grade for cull cows from either dairy or beef sired lines. This is likely due to the generally lower body condition and fat cover of cull cows compared to finished beef heifers and steers.

DB was an employee of Innovent Technology Ltd., AE was an employee of Ritchie Ltd., and WT was an employee of Harbro Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.