This work investigates whether automatically quantified pronuclear (PN) dynamics on day 0 of development are associated with embryo ploidy and can provide predictive information beyond maternal age. We focus on the timing and spatial behavior of the two pronuclei (PN1 and PN2) in time-lapse microscopy (TLM) sequences and their relationship with preimplantation genetic testing for aneuploidy (PGT-A) outcomes. We developed a fully automated, two-stage computer vision pipeline for PN segmentation and tracking in bright-field TLM images. First, a U-Net model was trained on manually annotated frames to detect the combined PN region. Second, a geometric circle-fitting procedure was applied to these joint PN masks to synthetically separate overlapping pronuclei and generate a large weakly labeled dataset of individual PN masks. A MiT-based U-Net was then trained on this expanded dataset to directly segment PN1, PN2, and the embryo in each frame. The final model, trained exclusively on data from a single clinic (Cohort 1), was applied to time-lapse sequences from two independent IVF clinics with linked PGT-A results. From the resulting PN masks, we constructed PN time series and extracted quantitative features describing (i) PN appearance and disappearance times, (ii) asynchrony of disappearance, (iii) PN size, (iv) inter-PN distance, and (v) radial PN position relative to the embryo center. PGT-A outcomes were harmonized to a binary euploid/aneuploid label per embryo. Gradient-boosted tree models were trained with patient-level grouped cross-validation under three predictor sets: maternal age only, PN features only, and age plus PN features. PN-only models achieved performance above chance but inferior to age-only models, indicating that day-0 PN dynamics contain real but modest information about ploidy. Combining age with PN features consistently improved discrimination compared to age alone, although the gain in AUC was small. The most influential PN-derived predictors involved PN timing, size, and radial position. However, PN-based models trained in one clinic transferred poorly to the other, highlighting sensitivity to cross-site domain shift.
Tsakalis et al. (Fri,) studied this question.