An automated method to quantify the composition of live pigs based on computed tomography segmentation using deep neural networks

Abstract

Knowledge of the body composition of growing pigs is of interest to breeding companies and producers, as it can be used to optimize production. As an emerging noninvasive technology computed tomography (CT) has been extensively applied in animal production studies. Recently, deep learning has generated new insights in medical image segmentation. In this paper, we describe a noninvasive method to automatically divide and quantify the composition of live pigs based on CT imaging and the application of deep neural network. The method consists of torso segmentation, visceral tissue removal, and identification and quantification of bone, lean meat, and fat using CT scans. The challenge addressed by this method is to identify the internal organs with complex structures and variable densities, the heart, lung, liver, stomach, spleen, kidney, colon, cecum, esophagus, jejunum, uterine horn, bladder, ureter, and rectum. We developed a bidirectional convolutional residual (BCR) framework to automatically and precisely segment the internal organs in a set of CT scans, and all internal organs share one label to remove these organs at once. To facilitate comparison with previous models, we tested BCR framework and found that it outperforms the classical approaches U-Net, residual U-Net (Res-U-Net) and dense U-Net (Dense-U-Net). We experimented on 40 pigs and compared our method with manual dissection. The results correlated well, and demonstrate that our method can accurately estimate the body composition proportion of fat, lean meat and bone—of live pigs, so the method will be valuable for livestock production.