Background: Pronuclear assessment appears to have the ability to distinguish good and bad embryos in the zygote stage,but paradoxical results were obtained in clinical studies.This situation might be caused by the robust qualitative detection of the development of dynamic pronuclei. Here,we aim to establish a quantitative pronuclear measurement method by applying expert experience deep learning from large annotated datasets. Methods: Convinced handle-annotated 2PN images(13419) were used for deep learning then corresponded errors were recorded through handle check for subsequent parameters adjusting. We used 790 embryos with 52479 PN images from 155 patients for analysis the area of pronuclei and the preimplantation genetic test results.Establishment of the exponential fitting equation and the key coefficient β1 was extracted from the model for quantitative analysis for pronuclear(PN) annotation and automatic recognition. Findings: Based on the female original PN coefficient β1,the chromosome normal rate in the blastocyst with biggest PN area is much higher than that of the blastocyst with smallest PN area(58.06% vs.45.16%, OR=1.68[1.07-2.64];P=0.031).After adjusting coefficient β1 by the first three frames which high variance of outlier PN areas was removed, coefficient β1 at 12 hours and at 14 hours post-insemination,similar but stronger evidence was obtained. All these discrepancies resulted from the female propositus in the PGT(SR) subgroup and smaller chromosomal errors. Conclusion(s): The results suggest that detailed analysis of the images of embryos could improve our understanding of developmental biology. Funding: None
Research on Automatic Recognition of Casting Defects Based on Deep Learning
