A highly accurate framework for estimating eye muscle area and backfat thickness of pigs in vivo using deep learning.

Eye muscle area (EMA) and backfat thickness (BFT) are key determinants of pig carcass value. While ultrasound imaging allows non-invasive estimation of these traits in live animals, measurement accuracy is often compromised by image noise. Manual segmentation of muscle and fat boundaries is not only labor-intensive but also prone to inter-operator variability. To overcome these limitations, this study developed a deep learning-based framework for automated segmentation of pig ultrasound images to accurately estimate EMA and BFT. We constructed a large-scale dataset comprising 10,088 pig ultrasound images and evaluated six neural network architectures. Among them, ReAMS-UNet achieved the highest segmentation performance for the eye muscle region, with an Intersection over Union (IoU) of 0.9788 and a Dice coefficient of 0.9893. The model's EMA estimates showed highly consistent with manual analyses, yielding a mean absolute error (MAE) of 0.359 cm2 and a coefficient of determination (R2) of 0.9964. For BFT estimation, an integrated approach combining precise image binarization with eye muscle segmentation resulted in an MAE of 0.562 mm and an R2 of 0.9910 compared to manual measurements. Furthermore, correlation analysis with actual carcass data revealed Pearson correlation coefficients exceeding 0.9 for both traits, demonstrating that the framework performs on par with experienced technicians. These results underscore the potential of the proposed method to enhance efficiency and objectivity in the pig industry.