256 Characterize the predictability of high-quality semen boars from B-Ultrasound Imaging.

Abstract This study investigates the factors influencing the predictive power of B-ultrasound imaging in assessing semen quality in young boars. Semen quality is a key factor in animal breeding, directly influencing reproductive success and herd productivity. While advances in molecular evaluation and cryopreservation have improved breeding outcomes, traditional fertility assessments remain resource-intensive. Deep learning (DL) has revolutionized medical imaging and presents new opportunities for veterinary applications, including reproductive assessments via ultrasound. However, its application to B-ultrasound imaging in boars remains underexplored. This study addresses this gap by using DL models to analyze testicular ultrasound features for predicting semen quality more efficiently. A total of 107 boars born in 2021 were examined using an IMV Technologies ExaPad Mini ultrasound system under standardized imaging settings. Each boar was scanned to collect 9–17 images, resulting in a dataset of 1,417 images. Semen quality data, including sperm concentration, motility, morphology, and volume, were collected biweekly to monthly throughout 2022. Boars were classified based on industry standards, requiring a total sperm count of ≥ 20 × 10⁹, motility ≥ 70%, and morphology ≥ 70% from collections between 200 and 540 days of age. Those consistently meeting these criteria were categorized as high-quality semen boars (n = 82), while those with at least 10% of collections failing to meet the standards or flagged for semen quality issues were classified as underperforming (n = 25). DL models (YOLOv11n and YOLOv11m) were trained and validated using a random image data-splitting method. Image preprocessing involved cropping to focus on seminiferous tubules and applying brightness augmentation. Model performance, measured by precision and recall, was analyzed using ANOVA. Results indicate that full images, the vertical angle, and avoiding brightness augmentation optimize model performance. Brightness augmentation significantly reduced precision (p < 0.001), leading to a 3.34% decline. Cropping negatively impacted sensitivity (p < 0.001, coefficient = -0.2049), while the vertical angle improved it (p < 0.001, coefficient = 0.1302). The YOLOv11m model had a slightly higher Recall than YOLOv11n, but the difference was not statistically significant (p = 0.548). In conclusion, the strong performance of full images suggests that areas beyond the seminiferous tubules contribute to semen quality prediction. The vertical angle likely captures more relevant details than the horizontal angle, as it provides a clearer view of the rete testis structure. Additionally, the precision drop with brightness augmentation highlights the importance of preserving natural brightness levels. This study shows that optimizing image selection and acquisition enhances ultrasound-based semen quality prediction in boars.