Automating Embryo Quality Classification in IVF by Comparing the Performance of ResNet50, VGG16, and InceptionV3

Infertility affects millions of couples worldwide, and the success of In Vitro Fertilization (IVF) heavily depends on accurate assessment of embryo quality, which is currently performed manually and is prone to subjectivity. This study investigates the use of deep learning for automating embryo classification based on microscopic images from Day 3 and Day 5 of development. Four models ResNet50, VGG16, InceptionV3, and a custom CNN were trained on 1,432 images (1,076 from Day 3 and 356 from Day 5) using the Adam optimizer at a learning rate of 0.0001 for 100 epochs. Results show InceptionV3 achieving the highest accuracy on Day 3 (99.09%) and Day 5 (89.5%), with precision, recall, and F1-scores all at 0.99 and 0.89, respectively. ResNet50 followed closely with Day 3 accuracy of 97.27% and Day 5 accuracy of 86.8%. The custom CNN demonstrated moderate performance, achieving 91.82% accuracy on Day 3 and 81.6% on Day 5, while VGG16 showed the lowest performance. These findings highlight InceptionV3 as the most reliable model, while the custom CNN offers practical solutions for smaller datasets or computationally constrained environments. This approach enhances IVF precision and efficiency, minimizing subjectivity and accelerating analysis.