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

Highlights

• 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.

Abstract

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, $1.14{\mathrm{cm}}^2$ 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.

Introduction

The growth of livestock production in the world has been increasing on a large scale, especially in emerging countries, such as Brazil and China (Salter, 2017). Livestock production is one of the most important productive sectors in Brazil with more than 210 million heads of cattle (de Souza et al., 2019). Only in 2020, 14,740 million heads of cattle were slaughtered under some type of health inspection service in Brazil, and this sector occupies about 20% of the Brazilian Gross Domestic Product (GDP) (IBGE, 2021).

The carcasses show variability in several traits, as Intramuscular Fat (IMF), marbling and rib-eye area. This variability in traits, such as juiciness and flavor, are related to quality and boning yields, which influence marketing and economic results (Felício, 2010, Kvam and Kongsro, 2017). Ultrasound imaging is commonly used to estimate the size of various cuts of meat or quality traits in live animals (Booth et al., 2006). Unfortunately, ultrasound images are known for having large amount of noise, which can make it difficult to define the exact boundaries or shapes of the regions in these images (Arias et al., 2007). In this regard, new strategies related to the digital image processing field are required to improve the process of extracting information from these groups of images.

In (Booth et al., 2006), the authors proposed a method for the segmentation of pork loins using ultrasound images based on the Active Contour (AC) framework. The method relies on an external measure method based on global pixel intensity. The algorithm presented 83.26% of Intersection over Union (IoU) score. Arias et al. (2007) proposed a method that combines shape priors and image information to achieve automatic estimation of rib-eye area in ultrasound images. The method evolves a curve corresponding to the shape and location of the correct rib-eye area. In the result analysis is pointed that 98.69% of the images had IoU error range smaller than 20%, and 84% was smaller than 15% considering the rigorous measure. In (Kvam and Kongsro, 2017), a custom Deep Learning to predict intramuscular fat in breeding pigs was proposed. The final correlation obtained was of $\bar{R}=0.74$ and Root Mean Squared Error (RMSE) of $1.8$. The proposed Convolutional Neural Network (CNN) is simple but the results were encouraging.

Deep Learning is applied in different contexts like medical images analysis (Shen et al., 2017), estimation of leaf area index and defoliation (Albughdadi et al., 2021, da Silva et al., 2019), segmentation and classification of trees (Liu et al., 2019), detection of tree species (Freudenberg et al., 2019), and etc. However, deep learning applied to automatic segmentation of cattle ultrasound images is a relatively new area and consequently the related literature is very limited. In addition, rib-eye segmentation presents an extra challenge which is the absence of part of the boundaries of the region of interest. Therefore, the segmentation methods need to estimate the limits without visual characteristics appearing in the image due to noise or limitations of the capture device.

Currently, the Deep Learning (Osco et al., 2021) literature has presented several well-established and state-of-the-art semantic segmentation methods to support the automatic segmentation of images. Nonetheless, up to the writing moment, there is no evidence about the usage of deep networks to segment the cattle rib-eye areas in ultrasound images. To fulfill this gap, we evaluated the performance of different state-of-the-art architectures, like the Unet++, FCN, U-Net, SegNet, and Deeplab v3+, to segment cattle rib-eye area in ultrasound imagery. To guarantee the generalizability capacity of the applied neural networks, we also adopted images with intense visual noise in which part of the boundaries is not visible, and the results remain promising. Moreover we reported results with input images at different resolutions, showing that even low resolution images (128x128 pixels) provide satisfactory results. These results are important for the inclusion of such methods in embedded systems. As an additional contribution, we are freely made available¹ the labeled rib-eye area dataset for that future research can use it in other investigations.

This paper is organized as follows. Section 2 describes the compared CNN methods, its parameters and training schema. The models are evaluated and discussed in Section 3. Finally, in Section 4, the conclusion and future work are presented.