Human Embryo Quality Assessment with Deep Learning Models.

Abstract

Background: Embryo quality assessment plays a pivotal role in assisted reproductive technology (ART) for selecting viable embryos for implantation. Accurate evaluation is essential for improving success rates in fertility treatments. Traditional assessment methods rely on subjective visual grading by embryologists, which can lead to inconsistencies. The application of deep learning in this domain offers the potential for objective and reproducible assessments.

Materials and methods: This study investigates the use of deep learning models to classify embryo images as good or not good at the day-3 and day-5 stages. A dataset obtained from Hung Vuong Hospital in Ho Chi Minh City was used to train and evaluate four convolutional neural network (CNN) architectures: VGG-19, ResNet-50, InceptionV3, and EfficientNetV2. Performance metrics, including accuracy, precision, and recall, were used to assess model effectiveness.

Results: Among the tested models, EfficientNetV2 demonstrated superior performance, achieving an accuracy of 95.26%, a precision of 96.30%, and a recall of 97.25%. These results indicate that deep learning models, particularly EfficientNetV2, can provide highly accurate and consistent assessments of embryo quality.

Conclusion: The high classification accuracy of EfficientNetV2 underscores its potential as a valuable tool for fertility specialists. By offering objective and consistent evaluations, this approach can enhance fertility treatment efficiency and support prospective parents in their reproductive journey.