Mensuração automática de espessura de gordura subcutânea a partir de imagens ultrassonográficas de bovinos utilizando deep learning

Abstract

Currently, livestock is one of the main economic activities in Brazil and has a large share of the country’s exports. Due to the demands of international markets, the search for increased meat quality has become a preponderant concern in the whole production chain. In this sense, different strategies of genetic improvement have been used, among them, is a search for cattle with better body fat finishing. The evaluation of body fat is performed at different moments in the production chain, most of the time noninvasively, with live animals and with the use of ultrasound. From the obtained ultrasound images, a specialist manually performs the fat finishing classification, generating a delay in the process, that ends up harming the veracity of the results obtained. This study aims to develop an automatic approach to evaluate the finished fat in cattle, based on convolutional neural networks, from ultrasonographic images. The images for construction of the regression models were provided by the Brazilian Association of Hereford and Braford and were previously classified by a specialist. Of the 7,951 images used in the project, 6,758 were separated for network training and 1,193 for testing. The generalizability of the classifiers was evaluated by the correlation of the results obtained with the previously classified values that were provided. According to a specialist, the correlation of the classifications and the predictions should be at 0.85 or greater. The correlation of the results of the network with the actual results was 0.97. This correlation of the final results prove the effectiveness of the use of Convolutional Neural Networks to automate the process of subcutaneous fat classification in ultrasound images of bovine.