EmbryoSwin++: Enhanced swin transformer with supervised contrastive learning for embryo multi-stage classification in assisted reproductive technology

Accurate embryo stage classification is crucial for enhancing IVF success rates; however, existing automated methods often struggle to handle inter-stage similarity and class imbalance, leading to misclassification across morphologically similar developmental phases. Furthermore, conventional CNN-based approaches tend to lose fine-grained spatial details, which are critical for distinguishing closely related stages. The proposed EmbryoSwin++ model addresses these gaps by integrating a Swin Transformer backbone with a supervised contrastive learning head and a Balanced Batch Sampler, enabling the model to learn discriminative embeddings while ensuring equitable representation of all 15 developmental stages. This dual-loss framework, combining label-smoothed cross-entropy and contrastive loss, enhances robustness, mitigates overfitting, and improves generalization across datasets. Evaluated on the Human Embryo Time-Lapse Video Dataset, the model achieved a validation accuracy of 92.12 %, a macro F1-score of 0.9196, and high AUC values approaching 1.0 for all classes, demonstrating strong discriminative capability. Grad-CAM analysis confirmed the model’s focus on biologically relevant embryo regions, validating its interpretability.