Abstract: Abstract: An accurate and rapid detection of back-fat thickness was essential to the precise feeding of the breeding pigs. However, the current manual analysis of backfat thickness cannot fully meet the large-scale production. In this study, a more accurate and stable determination of backfat thickness was developed with an automatic tool in pig breeding. The back-fat thickness was measured from the B-scan ultrasonography (B-Scan) image of pigs, where a Fully Convolutional Networks (FCN) model was trained for this task. The best performance of the model was achieved at the 20 literation with the batch size of 32, the learning rate of 0.005, and the optimizer of Adam. A prediction validation was conducted with the testing dataset, slaughtering measured dataset, and a separate human labelled dataset. The testing dataset was then segmented for the basic model performance with the FCN for the back-fat thickness. The measurement between the prediction and expert labelled (ground truth) was not statistically significant (P>0.05), where the correlation coefficient was 0.92 (P<0.01), indicating the reliable measurement of back-fat thickness for the targeted breeds using the FCN model. The slaughtering measured dataset was obtained from five selected pigs, which were the relatively close weight and body characteristics inspected by the expert group. The measuring points were chosen and marked on and near the standard B-scan measurement points, where the 5 points along the head to tail on each side were equally distanced. The B-Scan images of the measurement points were firstly taken for every pig to measure the back-fact thickness using the conventional inspection and the prediction model. Then, the direct measurements were taken with a vernier calliper on the chosen measurement points after slaughtering these selected pigs. Further analysis was made after the conventional inspection, prediction model, and direct measurement after slaughtering. It was found that there were smaller values in the conventional and prediction model than that in the direct measurement, where the minimum was found in the prediction model. The correlation coefficient between the three methods was all above 0.92 (P<0.01). Moreover, the conventional and prediction model was not statistically significant (P>0.05), with a correlation coefficient of 0.97 (P<0.01). It infers that the prediction model was much more accurate, compared with the conventional, indicating a reliable alternative way for the back-fat measurement. Furthermore, the separate human labelling dataset included the measurements on the same set of B-Scan images from four groups of six people, indicating the different levels of professional knowledge and experience. Specifically, the Expert Group (EG) involved experts combined with depth breeding knowledge and at least five years of field experience. The Professional Group (PG) were the licensed professionals on sites with the relatively focused on practice which carried hundreds of measurements. The Student Group (SG) were the undergraduates and postgraduates who majored in animal husbandry with less hands-on experience. The Novice Group (NG) were people not familiar with the industry, who only went through a quick lesson. The results between each group were not statistically significant (P>0.05), the same as every group and the prediction model. However, the standard deviation of in-group data varied notably, where the minimum was the expert group of 0.17 mm and the maximum was the novice group of 1.67 mm. In comparison, the prediction model presented a result free of human factors. Consequently, an FCN model was trained and implemented to semantic segment the B-scan images, with emphasis on some of the popular breeds in the industry, such as Large White, Landrace, and Duroc. The prediction was compared with the conventional inspection and direct measurements after slaughtering, indicating that the measured back-fat thickness from a B-Scan image using the FCN prediction model provided practically accurate, quick, and stable values. This finding can potentially provide production benefits, in order to significantly reduce the knowledge and experience requirements, as well as the workload and the training cost of staff.