Pork carcass composition derived from a neural network model of electromagnetic scans

We used an advanced computer logic system (NETS 3.0) to decipher electromagnetic (EM) scans in lieu of traditional linear regression for estimation of pork carcass composition. Fifty EM scans of pork carcasses were obtained on-line (prerigor) at a swine slaughter facility. Right sides were cut into wholesale parts and dissected into fat, lean, and bone to obtain total dissected carcass and primal cut lean. In this study, the input layer consisted of 81 nodes (80-point EM scan curve and warm carcass weight), one hidden layer of 42 nodes, and an output layer consisting of one node, which were run separately for outputs of ham, loin, or shoulder lean. The hidden layer connected to the output of total lean contained 50 nodes. Thirty-five scans were used for training of the network. The new network was then tested with 15 previously unseen input/output pairs. Separate neural networks were developed for the estimation of dissected total carcass, ham, loin, and shoulder lean. The NETS configuration improved on linear regression equations for estimation of total carcass lean by .31 kg, ham lean by .284 kg, and shoulder lean by .148 kg. Our results show that advanced computer logic systems have the capacity to improve upon traditional linear regression equations for prediction of pork carcass composition.