Automated Embryo Selection Using Deep Learning: A Comparative Study of 3D-CNN, Hybrid, and Transfer Learning Models

Background The crucial process of choosing viable embryos for in vitro fertilization (IVF) has always depended on the subjective evaluation of embryologists. Automating this process using deep learning may improve pregnancy outcomes and make the evaluation more reliable. Methods A 3D Convolutional Neural Network (3D-CNN), CNN + LSTM hybrid model, and TimeDistributed CNN + GRU were among the deep learning architectures that we investigated. We also used transfer learning using pre-trained models like EfficientNet and ResNet50. To improve model generalization, frame selection tactics and data augmentation techniques were also used. Accuracy, precision, recall, and F1-score metrics on embryo time-lapse films were used to assess performance. Results The 3D-CNN model stood out by not misclassifying any non-viable embryos, which makes it highly reliable in avoiding poor embryo choices. Hybrid models and those based on transfer learning reached up to 82% accuracy and showed a good balance in identifying both viable and non-viable embryos. The TimeDistributed CNN + GRU model, especially when combined with frame selection and augmentation, improved the ability to detect viable embryos even further. Conclusion The proposed models, particularly the 3D-CNN, show strong potential for automating embryo selection with high accuracy, making them suitable for clinical use. Using frame selection and data augmentation helped the models better capture how embryos develop over time. To further improve performance, future research could explore combining multiple models, finding the right balance between recall and specificity, and incorporating more clinical data.