121 Relationship between human and computer vision-based assessments of luteal blood perfusion in embryo transfer recipients

Luteal blood perfusion based on color Doppler ultrasonography examinations at the time of embryo transfer has been shown to be positively associated with pregnancy rates in beef cattle. However, interpreting color Doppler ultrasound images requires highly trained personnel and may present variability between observers, challenging large-scale applications of this technology in embryo transfer programs. This study aimed to evaluate the relationship between human and deep learning-based estimates of luteal area and blood perfusion. We hypothesized that luteal area and blood perfusion estimates generated by computer vision are correlated with human estimates generated by highly trained individuals. Postpartum beef embryo recipient cows (n = 750) were exposed to an industry-standard estrus synchronization protocol. Brightness (B)-mode and color Doppler ultrasonography examinations of the corpus luteum were performed immediately before fixed-time embryo transfer, and 200–300 frame videos were recorded for each recipient cow. Luteal area was determined using the B-mode tracing function and luteal blood perfusion was determined based on subjective estimations of the percentage of luteal area with color Doppler signals. Cows were categorized based on human-based luteal blood perfusion score: I (<25%), II (25%–45%), or III (>45%). A deep learning-based approach used a multi-step pipeline in which frames were extracted from ultrasound videos, cropped, and resized. A convolutional neural network identified high-quality frames, and a U-net model segmented the corpus luteum region. Luteal area was calculated based on the pixel count in the predicted masks. Luteal blood perfusion color signals were isolated to compute blood perfusion, and dark regions were used to compute cavity area, which was subtracted from final estimates of area and blood perfusion. A strong positive correlation existed between human and DL estimates for luteal area (r = 0.63; P < 0.001) and blood perfusion (r = 0.74; P < 0.001) estimation. Embryo recipients with human-based luteal blood perfusion score III had greater (P < 0.001) luteal blood perfusion based on the DL method compared with embryo recipients with scores I and II. Moreover, recipients with human-based luteal blood perfusion score II had greater luteal blood perfusion based on the DL method compared with score I recipients. In summary, these results indicate that computer vision-based analysis of color Doppler examinations yields estimates that are comparable to highly trained human individuals. Therefore, upon further improvements, computer vision methods of color Doppler ultrasound images might be used to optimize embryo recipient selection based on luteal blood perfusion.