Non-destructive and efficient prediction of intramuscular fat in live pigs based on ultrasound images and machine learning

Intramuscular fat (IMF) content plays an essential role in the evaluation of meat quality. To select pig breeds with different IMF content, developing a method to predict the IMF content in live pigs was greatly significant to reduce the cost and time of breeding. In the current study, real-time ultrasound images 5 cm off-midline across the third and fourth last thoracic ribs of 336 live pigs were collected using the B-model technique, and image feature parameters were extracted by computer image processing techniques. Furthermore, multiple linear regression (MLR) and two machine learning algorithms, support vector machine (SVM) and back-propagation artificial neural network (BPANN), were used to develop the prediction models of IMF content. The experimental pigs were divided into a training dataset (n = 266) for developing the prediction models and a validation dataset (n = 70) for estimating the accuracy of the models, and a test set (n = 67) for additional model performance evaluation. The results reveal that the coefficient of determination (R²) of models ranges from 0.65 to 0.80 with a root-mean-square error (RMSE) range of 0.50 %–0.65 % in the training dataset. By contrast, the correlation coefficients (R) between the predicted IMF (PIMF) and the chemically measured IMF (CIMF) range from 0.72 to 0.82 with an RMSE ≤ 0.69 % for all the models in the validation and test dataset. Moreover, the results indicate that the individual ratio of absolute difference (ADIF) within 1 % between PIMF and CIMF is > 86.57 % for all the models. In addition, classification accuracy shows that the BPANN1 model has superior classification ability in both low and high IMF content groups compared to the other two types of models in the validation dataset, but not in the test dataset. The MLR models are superior to other models in the medium IMF content group. Overall, our research demonstrates that it is feasible to predict IMF content based on ultrasound images in live pigs and provides several alternative models for accurate determination of IMF content, which could accelerate the genetic improvement of IMF content, thereby improving the pork quality in pig breeding programs.