Improving Deep Learning-Based Algorithm for Ploidy Status Prediction Through Combined U-NET Blastocyst Segmentation...

Area of Science:

• Reproductive Medicine
• Artificial Intelligence
• Embryology

Background:

• Optimizing artificial intelligence (AI) models for ploidy status assessment is crucial in reproductive medicine.
• Current methods involve algorithm investigation, image segmentation refinement, and pattern analysis during embryonic development.

Purpose of the Study:

• To evaluate the effectiveness of U-NET architecture for embryo segmentation and time-lapse image extraction.
• To improve the accuracy of AI models for predicting embryonic ploidy status using pre-biopsy imaging data.

Main Methods:

• A convolutional neural network (CNN)-based model was developed using 1,020 time-lapse videos of blastocysts with known ploidy status.
• U-NET architecture was applied for blastocyst image segmentation.
• Sequential images were extracted 3 and 10 hours prior to biopsy, generating 31,642 and 99,324 images, respectively.

Main Results:

• The ploidy prediction model achieved accuracies of 0.59 (3 hr) and 0.63 (10 hr) without U-NET segmentation.
• Implementing U-NET improved model accuracy to 0.61 (3 hr) and 0.66 (10 hr).
• A 10-hour image extraction period resulted in higher accuracy than a 3-hour period.

Conclusions:

• Combining U-NET segmentation with 10-hour time-lapse image sequences enhances CNN-based models for ploidy status prediction.
• This integrated approach offers improved accuracy in identifying embryonic ploidy status.