Abstract
Color Doppler ultrasonography (CDU) can estimate luteal blood perfusion (BP) across different time points in the cattle estrous cycle. However, its interpretation is subjective and limits large-scale application. This study aimed to validate a deep learning (DL) pipeline for automated quantification of luteal BP and compare its estimates with human and image processing–based assessments. We also evaluated the ability of these methods to estimate circulating progesterone (P4) concentrations across stages of the estrous cycle and according to CL cavity status. Beef cows (n = 240) were synchronized for ovulation and examined by CDU on day 7 (early luteal phase, ELP; n = 146), between days 14 and 16 (late luteal phase, LLP; n = 46), or day 20 (luteolysis phase, LP; n = 48). Cavity proportion was determined using the B-mode tracing function, and CLs were classified as no/small (< 10%; n = 202) or large (≥10%; n = 38) cavities. Videos of the luteal-bearing ovary were analyzed using (1) subjective human evaluation, where two trained evaluators estimated BP; (2) ImageJ-based pixel counting, where BP was calculated as the number of color Doppler pixels divided by total pixels in the two most representative frames; and (3) a DL system combining convolutional neural network frame selection and U-net segmentation to automatically identify the CL region, isolate Doppler-positive pixels, and correct for cavities. Strong positive relationships were found among methods within ELP (R > 0.6; P = 0.01), LLP (R > 0.6; P = 0.01), and LP (R > 0.6; P = 0.01), and when CLs were stratified by cavity status: no/small (R > 0.64; P = 0.01) and large (R > 0.66; P = 0.01). Simple linear regressions indicated no relationship between BP and circulating P4 during ELP or LLP (P > 0.14). During LP, ImageJ (R = 0.47; P = 0.01), Human (R = 0.69; P = 0.01), and DL (R = 0.59; P = 0.01) methods showed positive relationships with P4. Among CLs without or with small cavities, ImageJ (R = 0.15; P = 0.03), Human (R = 0.40; P = 0.01), and DL (R = 0.25; P = 0.01) were related to P4, whereas no relationships were detected in CLs with large cavities (P > 0.19). These findings validate the DL pipeline, demonstrating that it produces BP estimates consistent with established approaches. Luteal BP predicted CL functionality based on P4 only during luteolysis, with human evaluation showing the strongest relationship, followed by DL and ImageJ. Although correlations among BP methods persisted regardless of cavity presence, large cavities eliminated the relationship between BP and P4. In conclusion, DL provides valid and objective BP estimates, but the estrous cycle stage and cavity presence influence its ability to estimate luteal function.