P-112 Predictive model of good quality blastocyst development based on static image of fresh mature oocytes

Abstract Study question Can good quality blastocyst development be predicted from a static images of fresh mature oocytes? Summary answer At deep learning model can be used to predict the blastocyst outcome with the 0.694 confidence based on a static image of a mature oocyte. What is known already The application of deep learning in in vitro fertilization(IVF) laboratories has been a rapidly evolving area of research, aimed at improving the efficiency, accuracy, and outcomes of IVF treatments. One of the most significant applications of deep learning in IVF laboratories is the grading and selection of embryos for implantation. But recently, oocyte is gaining more interest in the context of fertility preservation and oocyte donation. The development of reliable models to predict blastocyst development from a single image of metaphase II oocyte may improve the counselling and management of fertility preservation cycles as well as oocyte donors and recipients. Study design, size, duration This study aimed to develop a model for good-quality blastocyst development prediction using non-invasive imaging of fresh metaphase II oocytes. A dataset of 747 oocyte images with an imbalanced class distribution (70%blastocyst, 30%non-blastocyst) was used. Images collected from one clinic required patient-wise data separation to ensure meaningful evaluation. Preprocessing included grayscale conversion, normalization to [0,1], resizing to 224 × 224pixels, and cross-validation with a stratified group k-fold split to maintain class balance and patient separation across folds Participants/materials, setting, methods The designed machine learning model is based on a modified pre-trained VGG16-architecture to benefit from transfer learning. The model was trained to distinguish oocytes likely to develop into blastocysts. The imbalance class distribution was addressed using Focal Loss, enabling the model to prioritize harder-to-classify images while balancing minority class gradients. The training involved two phases: classifier-only training (three epochs) and end-to-end fine-tuning (ten epochs). Augmentation techniques like image rotations, zoom, and intensity adjustments enhanced robustness. Main results and the role of chance The preprocessing pipeline included grayscale conversion, normalization, and cross-validation with a stratified group k-fold split to ensure robust evaluation and prevent data leakage. Given the dataset’s small size, a data-centric approach was adopted, focusing on collecting high-quality oocyte images and cleaning to remove noise and artifacts, maximizing data utility and clinical relevance. The use of Focal Loss further addressed class imbalance, balancing sensitivity and specificity while prioritizing harder-to-classify cases. Building on these mentioned methods, the model demonstrated strong predictive performance. Indeed, the model achieved an AUC-ROC of 0.694, demonstrating good performance in predicting good quality blastocyst development (Grade A and B based on Gardner grading system). Sensitivity and specificity were balanced at 0.65 and 0.672, respectively, reflecting the model’s ability to handle the class imbalance. The negative predictive value (NPV) (non-blastocyst development) was 0.382, while the positive predictive value (PPV) (blastocyst development) reached 0.86, indicating superior performance in identifying good quality blastocyst development. These results were validated on an independent test set including 224 images with patient-wise separation, ensuring clinical relevance and mitigating potential data leakage. The balanced performance across sensitivity and specificity metrics supports the model’s potential as a non-invasive predicting support tool for embryologists and clinicians. Limitations, reasons for caution The dataset included oocytes that developed into blastocysts and oocytes that did not reach the blastocyst stage, excluding other developmental outcomes. Data from a single clinic limits generalizability. Additionally, the relatively low PPV underscores the need for larger, multi-clinic datasets to validate the model’s robustness and clinical applicability. Wider implications of the findings This study highlights the potential of AI in non-invasive oocyte quality evaluation, supporting professionals in counselling fertility preservation and IVF patients. By predicting good quality blastocyst development, the model is expected to reduce subjective assessment and improve the prediction of success rates. Expanding dataset will enhance clinical impact and generalizability. Trial registration number No

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.