Automated luteal blood perfusion assessment in cows using image processing and deep learning algorithms

The aim of this study was to compare luteal blood perfusion (BP) estimated by subjective human evaluation, ImageJ pixel analysis, and a deep learning (DL)-based system. Color Doppler ultrasound (CDU) examinations utilized herein were from three independent studies where CDU videos were collected at embryo transfer (Day 7; early luteal phase; ELP; n = 129), or on Days 15 (mid luteal phase; MLP; n = 48) and 20 (late luteal phase; LLP; n = 47) of pregnancy. Each corpus luteum (CL) was classified as cavitated (CV; cavity ≥10% of luteal area) or non-cavitated (NC; <10%). Luteal BP was estimated using three methods: Human evaluation, ImageJ pixel counting, and a DL pipeline for frame selection, CL segmentation, and BP quantification. Positive linear relationships were observed among BP estimation methods in all datasets (r ≥ 0.52; P < 0.001) and within CL cavity classifications (r ≥ 0.62; P < 0.001). Strong relationships were consistently observed in the LLP dataset (r ≥ 0.79; P < 0.001). No relationship was detected between circulating progesterone (P4) and BP at ELP or MLP (P ≥ 0.14), regardless of the estimation method. Positive linear relationships (P ≤ 0.001) with P4 were observed for Human, ImageJ, and DL estimates at LLP. In conclusion, DL-derived BP estimates showed strong linear relationships, satisfactory agreement, and biological validity comparable to Human and ImageJ methods, supporting its use for automated assessment of luteal BP in cattle.