Automatic segmentation of cattle rib-eye area in ultrasound images using the UNet++ deep neural network

Abstract

• Deep Neural Networks perform well the semantic segmentation task in ultrasound images. • UNet ++ leads out better results inside the group of the Deep Neural Networks tested. • Cattle Rib-Eye Area can be mapped using the UNet++ neural network with high accuracy. Ultrasound imaging is commonly used to estimate the size of various cuts of meat or quality traits in live animals. Unfortunately, ultrasound images are known for having large amount of visual noise, which can make it difficult to define the exact boundaries or shapes of the regions of the interest in these images. Therefore, new strategies related to the digital image processing field are required to improve the process of obtaining information from these groups of images. In this context, artificial intelligence, through deep learning methods particularly, has proved to be an optimized and efficient strategy, but that has not yet been investigated in the cattle rib-eye area. This paper aims to investigate the feasibility of applying the Unet++ deep neural network to automatic segmentation of cattle rib-eye area in ultrasound images. Additionally, several well established deep learning semantic segmentation models are compared with Unet++ performance. These architectures are FCN, U-Net, SegNet, and Deeplab v3+. The models were tested on a dataset composed of gray scale images of cattle ultrasound. All models showed excellent results in both location and boundaries. Best results showed 97.37 % in IoU, 1.14 cm 2 in MAE and coefficient of determination ( R 2 ) of 0.999 . The labeled rib-eye area dataset used in this study is available for future research.