Evaluation of a Numerical, Real-Time Ultrasound Imaging Model for the Prediction of Litter Size in Pregnant Sows, with Machine Learning

Abstract

The present study aimed to evaluate the accuracy of a numerical model, quantifying real-time ultrasonographic (RTU) images of pregnant sows, to predict litter size. The time of the test with the least error was also considered. A number of 4165 pregnancies in Farm 1 and 438 in Farm 2 were diagnosed twice, with the quality of the RTU images translated into rated-scale values (RSV1 and RSV2). When a deep neural network (DNN) was trained, the evaluation of the method showed that the prediction of litter size can be performed with little error. Root square mean error (RMSE) for training, validation with data from Farm 1, and testing on the data from Farm 2 were 0.91, 0.97, and 1.05, respectively. Corresponding mean absolute errors (MAE) were 2.27, 2.41, and 2.58. Time appeared to be a critical factor for the accuracy of the model. The smallest MAE was achieved when the RTU was performed at days 20–22. It is concluded that a numerical, RTU imaging model is a prominent predictor of litter size, when a DNN is used. Therefore, early routinely evaluated RTU images of pregnant sows can predict litter size, with machine learning, in an automated manner and provide a useful tool for the efficient management of pregnant sows.