Advanced Embryo Ploidy Classification Using Vision Transformers: Integration of Sequential Time-Lapse Imaging and...

Area of Science:

• Assisted Reproductive Technology (ART)
• Computational Biology
• Genetics

Background:

• Accurate embryo ploidy identification is vital for successful assisted reproductive technology (ART) outcomes.
• Conventional deep learning models struggle with class imbalance, especially underrepresenting mosaic embryos.
• Mosaic embryos pose a significant challenge in ploidy classification.

Purpose of the Study:

• To enhance embryo ploidy classification accuracy using Vision Transformers (ViTs).
• To integrate sequential time-lapse imaging with ViTs for improved embryo assessment.
• To address data imbalance, particularly the underrepresentation of mosaic embryos, through random undersampling (RUS).

Main Methods:

• A retrospective analysis of 1020 blastocyst-stage time-lapse videos was conducted.
• Customized deep learning models, specifically ViT-B/16 and ViT-B/32, were fine-tuned.
• Random undersampling (RUS) was applied to create a balanced dataset of 17,000 images per class (Euploid, aneuploid, mosaic).

Main Results:

• The ViT-B/16 model achieved 0.84 accuracy in binary and 0.67 in multiclass classification on the balanced dataset.
• Performance significantly dropped to 0.49 on the imbalanced dataset, highlighting the effectiveness of RUS.
• Random undersampling notably improved the prediction accuracy for minority classes, including mosaic embryos.

Conclusions:

• Integrating ViTs with sequential time-lapse imaging and RUS offers a promising non-invasive method for embryo ploidy classification.
• This approach enhances the accuracy of identifying mosaic embryos.
• The findings support more informed embryo selection in ART, potentially improving success rates.