JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):1
Oral Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263631

P-57. Development of Artificial Intelligence software for assessing ploidy in human embryos from blastocyst images

Murilo Costa Maffeis¹, Mariana Nicolielo², Dóris Spinosa Chéles², Bruna Lourenço², José Roberto Alegretti², Aline Rodrigues Lorenzon², Marcelo Fábio Gouveia Nogueira¹, José Celso Rocha¹

¹ Universidade Estadual Paulista "Júlio de Mesquita Filho" - UNESP - Assis - São Paulo - Brasil
² Huntington Medicina Reprodutiva - Eugin Group - São Paulo - SP - Brasil

Objective: To develop artificial intelligence software capable of predicting the ploidy of human embryos in a non-invasive manner, using segmented images of the three main morphological regions of the blastocyst (inner cell mass, trophectoderm, and blastocoel), combined with quantitative morphological variables extracted automatically by digital processing techniques.
Methods: A total of 1,428 blastocyst images (D5/D6) obtained by EmbryoScope^® time-lapse incubators were used, from in vitro fertilization (IVF) cycles submitted to preimplantation genetic testing for aneuploidy (PGT-A). The images were resized and segmented to isolate the three regions of interest. Three convolutional neural networks (CNNs), based on the ResNet-50, VGG16, and InceptionV3 architectures, were trained separately for each segmented structure. In parallel, 33 quantitative morphological variables were automatically extracted by an in-house software using digital image processing techniques. The feature vectors generated by the CNNs and the analysis of the variables were integrated by a multilayer perceptron network, which performed the final prediction of ploidy (euploid or aneuploid). The system's performance was evaluated by validation and blind testing, using statistical metrics such as confusion matrix and AUC-ROC.
Results: The artificial intelligence used in the research, with the best analyzed performance, showed 69.20% accuracy in training, 56.62% accuracy in validation, and 55.33% accuracy in the blind test set. The combination of segmented neural networks and morphological variable analysis showed good results for training, but with room for improvement in validation and blind testing. The results obtained contribute to the advancement of the field and support the continued development of the proposed approach. This strategy allowed for greater consistency and reproducibility between tests, bringing the system closer to real clinical application conditions.
Conclusion: The proposal offers an innovative and non-invasive approach to assist in embryo selection in assisted reproduction clinics. By integrating segmented images and morphological variables into an interpretable model, the system can increase IVF success rates, reduce costs and risks associated with PGT-A, and provide embryologists with a reliable decision support tool. To improve accuracy, it is necessary to expand the dataset and optimize the neural network architecture.