Comparação entre modelos de redes neurais U-Net e U-Net Fuzzy para mensurar regiões de imagens de ultrassonografia bovina

Abstract

Beef cattle production is one of the fundamental pillars of the Brazilian economy. Carcass ultrasonography enables the in vivo measurement of the Longissimus Muscle Area (LMA) and the Backfat Thickness (BFT) and Rump Fat Thickness (RFT), traits associated with mus- cularity and early development in cattle. However, obtaining these measurements still relies predominantly on manual delineation and interpretation performed by specialists, making the process subjective, time-consuming, and susceptible to inter-observer variability. The objec- tive of this study is to develop and compare two models for the automatic segmentation and measurement of these indicators in ultrasound images from two cattle breeds. The first model is a multitask U-Net convolutional neural network, while the second, termed Fuzzy U-Net, in- tegrates a U-Net architecture with an Adaptive Neuro-Fuzzy Inference System (ANFIS). The images were obtained from live animals at the Advanced Center for Beef Cattle Research and Development of the Institute of Animal Science (SAA-SP), located in Sertãozinho, São Paulo, Brazil. A total of 135 images were collected, including 108 from the Nelore breed and 27 from the Caracu breed. The dataset was expanded through data augmentation techniques, such as the inclusion of images with increased and decreased brightness. The segmentation masks used as reference, known as ground truth masks, were manually created with the assistance of a specialist. These masks consisted of binary images used to isolate specific regions within the ultrasound images, serving as references for model training and evaluation. Model performance was assessed using metrics such as accuracy, Mean Absolute Error (MAE), and the Dice coeffi- cient, which quantifies the similarity between the segmentation generated by the model and the reference masks produced by the specialist. The results indicate that the Fuzzy U-Net model outperforms the Multitask U-Net model in measuring the LMA and BFT indicators across all evaluation metrics, achieving accuracies of 95.15% for LMA and 99.95% for BFT. In contrast, both the Multitask U-Net and Fuzzy U-Net models exhibited similar performance for the RFT indicator. Finally, a computational application was developed to automatically detect specific regions and measure the indicators of interest in bovine ultrasound images using both models. This tool has the potential to assist in cattle selection and herd management.