An Automated Day 3 Embryo Grading Based On Morphological Characteristics Using CNN with Transfer Learning Techniques

In-Vitro fertilization (IVF) is the most successful form of assisted reproductive technology and one of the most used fertility treatments. According to a prior study, the day 3 embryo development stage is a crucial transitional time that significantly impacts the pregnancy and implantation rates of day 5 transfers. Most embryologists currently analyze the morphological characteristics of day 3 embryos visually using a microscope, which is subjective and time-consuming. This study aims to develop a CNN model capable of classifying embryo grades on day 3 based on their morphological characteristics using transfer learning. As a result, it automates the grading procedure. Four image preprocessing techniques, including data vectorization, image resizing, image normalization, and data augmentation, have been shown to improve image features. This project compares ResNet50 and Xception, two pre-trained CNN models. Analysis of the experiment's results reveals that data augmentation can address the imbalanced data issue and increase the data set. Transfer learning using a feature extraction technique outperforms fine-tuning substantially. The Xception model performed better than ResNet50 in classifying the grade of day 3 embryos based on morphological characteristics, with a 98.00% score across all evaluation metrics and the lowest loss value of 0.101.