Successful classification of pregnant and empty ultrasound imagery of sheep using convolutional neural network

Point-of-care diagnostic imaging is crucial in various veterinary and agricultural settings. In sheep, pregnancy diagnosis by ultrasound is routinely undertaken to determine pregnancy status, fetal number, age and health. However, diagnostic accuracy remains challenging, even for experienced scanners. Deep learning holds potential to address such challenges and improve accuracy and efficiency of ultrasound interpretation. This study evaluated a deep learning model to identify pregnant and empty (non-pregnant) ewes from continuous free-hand ultrasound video frames, reflecting real world conditions. Transcutaneous ultrasound (OviScan 6) imagery collected from 951 ewes, 44 days post 5-week joining. From these recordings, frames were extracted and labelled with diagnosis class (pregnant or empty). Two datasets were constructed: an unbalanced subset reflecting field conditions (884 pregnant; 67 empty), and a class-balanced subset for comparison (67 pregnant; 67 empty). A binary Convolutional Neural Network with two convolutional layers and one fully connected layer was trained on each. During internal testing, both models achieved 100% ewe-level classification accuracy, with only minor frame-level misclassification. Threshold optimization indicated the balanced model maintained perfect performance across thresholds, while the unbalanced model required a higher decision threshold (0.80) to achieve optimal performance. External evaluation on an independent cohort of 390 ewes demonstrated high diagnostic accuracy for both models, with the balanced model achieving 0.987 accuracy and the unbalanced model achieving 0.995 accuracy at the ewe level. These findings show that deep learning can accurately distinguish pregnant and empty ewes under commercial scanning conditions, with strong generalizability across variation in fetal number and gestational age.