In vitro experimental study demonstrates high classification accuracy of transfer learning models in bovine embryos, highlighting potential for automated time-lapse monitoring.
Key Points
To assess the developmental kinetics of bovine embryos cultured in two different media using time-lapse imaging and evaluate pretrained convolutional neural networks for automated developmental stage classification.
Cultured 311 individual bovine zygotes under standard in vitro fertilization conditions, comparing synthetic oviductal fluid (SOF, n=152) and Gx-TL™ medium (n=159) using a MIRI time-lapse system.
Manually annotated time-lapse frames from two-cell through blastocyst stages and augmented the image dataset to 5,000 images.
Trained and evaluated three pretrained transfer learning architectures: ResNet18, DenseNet121, and EfficientNet-B0 for two-class, nine-class, and ten-class embryo stage classification.
In the SOF cohort, 81 embryos reached the two-cell stage, 10 reached morula, and 6 reached blastocyst; in the Gx-TL™ cohort, 77 reached the two-cell stage, 18 reached morula, and 10 reached blastocyst.
All three convolutional neural network architectures achieved 100% accuracy in the two-class classification task across both culture media.
Classification accuracy across architectures ranged from 95% to 100% for the nine-class models and from 98% to 100% for the ten-class models.