Deep learning-driven automated carcass segmentation and composition quantification in live pigs via large-scale CT imaging and its application in genetic analysis of pig breeding

Abstract

Carcass segmentation and composition (CSC) traits are important indicators for assessing the economic efficiency of pig production. Conventional determination of these traits by slaughter has the drawbacks of high costs and the inability to retain breeding stock. Combining computed tomography (CT) with deep learning enables the non-invasive evaluation of live animal carcass characteristics. In this study, we proposed UPPECT for predicting CSC traits of live pigs based on deep learning. A labeled dataset comprising 300 pigs with a total of 63,708 CT images was constructed for training the nnU-Net model to automatically segment different cuts of pig carcasses. The composition quantification process was optimized using adaptive thresholding and bone filling to achieve accurate prediction of 16 CSC traits. At last, the genetic parameters of CSC traits obtained by UPPECT were estimated for 4,063 pigs. The segmentation model demonstrated excellent performance with a PA of 0.9992, an IoU of 0.9910 and an F1-score of 0.9955. We slaughtered and dissected 50 pigs to obtain real CSC trait values as the validation dataset. The results showed that our method improved the accuracy of composition quantification after optimization, and our predictions for all traits were highly correlated with manual dissection results, with correlation coefficients up to 0.9568. The heritability estimates ranged from 0.52 to 0.85 for all traits. Our study enables non-invasive and precise measurement of CSC traits of live pigs, which makes an important contribution to the breeding practice. A graphical user interface software for UPPECT is freely accessible at https://github.com/StMerce/UPPECT.