Computer vision analysis of luteal color Doppler ultrasonography for early and automated pregnancy diagnosis in Bos taurus beef cows

Abstract

This study evaluated the suitability of applying supervised deep learning (DL) algorithms for early and real-time pregnancy diagnosis in beef cattle using luteal color Doppler (CD) ultrasonography recorded on days 20 (D20) and 22 (D22) after fixed-time artificial insemination (FTAI). CD ultrasound videos from 390 females were manually evaluated by trained personnel to perform the human-based pregnancy diagnosis (Human). Images were extracted at a rate of 5 frames per second from each video, resulting in 10,533 (D20) and 10,413 (D22) valid frames after applying a frame-filtering pipeline. Three convolutional neural network (CNN) architectures—VGG19, Xception, and ResNet50—along with their averaged inference (Combined), were evaluated using restricted five-fold cross-validation, ensuring that images from the same animal did not appear in both training and validation sets. Final inferences for each animal were determined by averaging the network outputs across all video frames. Pregnancy status was confirmed on day 29 using conventional ultrasonography and treated as ground truth for assessing both Human and DL-based predictions. Accuracy levels were similar across methods, ranging from 0.84 (VGG19) to 0.87 (Human) for D20 and from 0.86 (VGG19) to 0.93 (Human) for D22. Based on Matthew’s correlation coefficient, the Combined and Xception architectures demonstrated the best overall agreement with true pregnancy status among DL models. These architectures performed comparably to human diagnosis, with the Combined model achieving similar F1 scores (0.89 vs. 0.91), higher specificity (0.72 vs. 0.65), and slightly lower sensitivity (0.95 vs. 1.00) on D20. Xception showed similar performance to human diagnosis on D22, with comparable accuracy (0.91 vs. 0.93), specificity (0.79 vs. 0.81), sensitivity (0.99 vs. 1.00), and F1 score (0.93 vs. 0.94). In conclusion, DL algorithms can effectively predict pregnancy status using CD ultrasonography earlier than industry-standard methods, with performance comparable to that of trained personnel.