Automated embryo selection using deep learning: a comparative study of 3D-CNN, hybrid, and transfer learning models

Embryo selection is a critical step in in vitro fertilization (IVF) and is traditionally based on subjective embryologist assessment, which can introduce variability. Deep learning–based automation has the potential to improve consistency and support clinical decision-making. We evaluated multiple temporal deep learning architectures for embryo viability assessment from time-lapse videos, including a 3D Convolutional Neural Network (3D-CNN), CNN plus LSTM, and TimeDistributed CNN plus GRU models. Transfer learning approaches using ResNet50 and EfficientNet backbones were also investigated. Frame selection and data augmentation strategies were applied to enhance generalization. Models were evaluated using accuracy, precision, recall, and F1-score. The 3D-CNN model achieved perfect specificity by avoiding false-positive predictions of non-viable embryos, a desirable property in clinical practice. Hybrid temporal models and transfer-learning-based approaches achieved accuracies up to 82 percent, showing a balanced ability to classify both viable and non-viable embryos. The TimeDistributed CNN plus GRU model benefited further from frame selection and augmentation, improving sensitivity to viable embryos. Temporal deep learning models demonstrate strong potential for automated embryo viability assessment using time-lapse imaging. In particular, the 3D-CNN model offers high reliability for avoiding suboptimal embryo selection. Incorporating temporal modeling, frame selection, and augmentation improves performance by capturing developmental dynamics. Future work will focus on model ensembling, external validation, and integration of additional clinical information.