Development and validation of deep learning based embryo selection across multiple days of transfer.

Area of Science:

• Reproductive Medicine
• Artificial Intelligence in Healthcare
• Embryology

Background:

• Traditional embryo evaluation in IVF relies on subjective morphological assessment.
• Time-lapse imaging offers objective morphokinetic data but requires expert interpretation.
• Automated analysis tools are needed to improve consistency and efficiency in embryo grading.

Purpose of the Study:

• To develop and validate iDAScore v2.0, a fully automated deep learning model for human embryo evaluation.
• To assess the model's predictive performance for embryo implantation across different developmental stages (Day 2, 3, 5+).
• To compare iDAScore v2.0 performance against existing embryo scoring systems.

Main Methods:

• Development and training of the iDAScore v2.0 model using a large, diverse dataset of 181,428 human embryos from 22 international IVF clinics.
• Validation of the model's predictive accuracy using areas under the receiver operating curve (AUC) for embryo implantation outcomes.
• Comparative analysis of iDAScore v2.0 performance against KIDScore D3 and KIDScore D5 v3 models.

Main Results:

• iDAScore v2.0 demonstrated predictive performance with AUCs ranging from 0.621 to 0.707, depending on the day of embryo transfer.
• Model performance showed improvement over time and strong correlation with key morphokinetic parameters.
• iDAScore v2.0 achieved performance equivalent to KIDScore D3 on Day 3 embryos and significantly surpassed KIDScore D5 v3 on Day 5+ embryos.

Conclusions:

• The iDAScore v2.0 model provides reliable, automated analysis of human embryo time-lapse sequences for implantation likelihood.
• This tool offers objective embryo ranking at both cleavage and blastocyst stages, improving grading consistency.
• iDAScore v2.0 enhances efficiency in IVF workflows by reducing the need for manual user input in embryo evaluation.