Development of Artificial Intelligence Software for Evaluating Ploidy in Human Embryos Based on Blastocyst Images

• Grant number: 25/20180-0
• Support Opportunities: Scholarships in Brazil - Master
• Start date: April 01, 2026
• End date: March 31, 2027
• Field of knowledge: Health Sciences - Medicine
• Principal Investigator: José Celso Rocha
• Grantee: Murilo Costa Maffeis
• Host Institution: Faculdade de Ciências e Letras (FCL-ASSIS). Universidade Estadual Paulista (UNESP). Campus de Assis. Assis , SP, Brazil
• Associated research grant: 23/16156-1 - Prediction of gestational success, using the variables of morphology and morphokinetics of the human embryo and the patient, using the Artificial Intelligence technique, through deep learning and multilayer perceptron., AP.R

Abstract

This project proposes the development of artificial intelligence software for the non-invasive prediction of human embryo ploidy, using static images of blastocysts obtained by time-lapse incubators. The proposal aims to improve embryo selection in in vitro fertilization cycles, reducing dependence on invasive methods such as preimplantation genetic test for aneuploidy and contributing to increased pregnancy success rates. The model will be structured with three convolutional neural networks, each specializing in the analysis of one of the main morphological regions of the blastocyst: inner cell mass, trophectoderm, and blastocoel. The characteristics extracted from these regions will be integrated by a multilayer perceptron neural network, responsible for the final prediction of ploidy status (euploid or aneuploid). In addition to the segmented images, the system will also incorporate 33 automatically extracted quantitative morphological variables, allowing for hybrid and multimodal analysis. The model's performance will be evaluated by cross-validation (K-Fold) and blind testing, using metrics such as confusion matrix, accuracy, sensitivity, F1-score, and AUC-ROC. Pre-trained architectures (such as ResNet, VGG, and Inception) will be explored, in addition to fine-tuning strategies and hyperparameter adjustments. Interpretability techniques, such as Grad-CAM, will be applied to ensure that the model's decisions are based on biologically plausible regions, promoting greater clinical confidence. At the end of the project, we will consider the possibility of making the software available through a user-friendly interface, which could facilitate its practical application in assisted reproduction clinics. (AU)