Abstract
This paper presents a comprehensive analysis of the application of deep learning models for embryo quality assessment in the field of in vitro fertilization (IVF). As embryo selection plays a crucial role in the success rates of IVF treatments, it is important to adopt an automated and accurate system to evaluate embryo viability. Our study focuses on comparing the effectiveness of four state-of-the-art deep learning models: VGG-19, EfficientNet, MobileNet, and ResNet, in classifying embryos based on their Inner cell mass (ICM) and trophectoderm (TE) cell characteristics from microscopic images. Utilizing a dataset of 1,020 embryonic images from two significant developmental stages distributed through the World Championship in Data Science and Artificial Intelligence 2023 - ISODS, we systematically trained and evaluated each model to address the challenges posed by small and imbalanced datasets commonly encountered in medical imaging. Through a series of experiments, including the application of data augmentation techniques and advanced model training strategies, we aimed to optimize model performance and minimize overfitting. The results revealed that EfficientNet-B0 has a high accuracy and robustness in distinguishing between viable and non-viable embryos. Additionally, we explored the potential of explainable AI techniques, such as Grad-CAM visualizations, to provide insights into the decision-making processes of the models. This study not only contributes to the advancement of automated embryo assessment systems but also lays the groundwork for future research to enhance IVF success rates through improved embryo selection methodologies.