Simulação preliminar da predição do volume globular bovino: um método não invasivo por meio de imagens RGB e redes neurais convolucionais.

Abstract

Among the various sectors of agricultural production in Brazil, livestock farming stands out due to the substantial number of animals and the animal and the products derived from animal production, serving both domestic consumption and export markets. The maintenance of the productive status of livestock depends on various factors, including nutrition, sanitary management, genetic and technological practices, all of which require constant investment and expansion. Numerous diseases affect herds, however, bovine parasitic meningoencephalitis stands out for causing significant losses in the production chain, ranging from health problems in animals to economic losses for the producer. Therefore, there is a need to monitor the production impact generated by this disease. Currently, the precise diagnosis of bovine parasitic meningoencephalitis is done invasively through blood collection for subsequent laboratory tests that accurately provide values for packed cell volume for the application of the correct treatment. However, no current method allows for the measurement of cell volume without contact with the animal. Therefore, the development of an approach capable of estimating packed cell volume without the need for contact with the animals can assist veterinarians in managing appropriate treatments. The use of convolutional neural networks enables a deep level of image analysis for pattern recognition. For this experiment, bovine images collected during handling at the Empresa Brasileira de Agropecuária (Embrapa) were used. In parallel with image collection, blood samples were taken for laboratory analysis of cell volume values in each animal. Image preprocessing and data augmentation techniques were employed to assist in the model training task. The best results obtained for the regression performed with the created model yielded values for the coefficient of determination (R²) and correlation coefficient with values of -0.06 and 0.36%, respectively. While the current study falls short of offering a conclusive solution for predicting the packed cell volume of cattle based on the ocular region as input data, it underscores the need for further investigations into more refined methodologies and/or alternative accuracy.