Deep learning-based segmentation of human oocytes with cross-dataset evaluation

BACKGROUND AND OBJECTIVE: Accurate segmentation of human oocyte subregions in microscopy images is an important step towards developing objective and automated systems to support oocyte quality assessment in assisted reproductive technologies. This study aims to segment four clinically relevant oocyte regions-ooplasm, perivitelline space (PVS), zona pellucida (ZP), and first polar body (PBI)-using deep learning models and to evaluate their performance across multiple clinical datasets. METHODS: Four segmentation models (UNet++, DeepLabV3+, SegFormer, and a transformer-based architecture) were trained on a clinically annotated dataset. The models were evaluated using the Intersection Over Union (IoU), Dice Similarity Coefficient (DSC), precision, and recall. To assess performance, external datasets collected under different imaging conditions from other clinical sites were used for testing. Domain adaptation techniques were applied by fine-tuning the pretrained models on external clinical datasets to improve cross-site robustness. RESULTS: SegFormer achieved the highest segmentation performance, consistently outperforming convolutional neural network (CNN) models in oocyte segmentation. On the primary training dataset, the SegFormer model achieved high DSC across all regions, with mean values exceeding 99% for ooplasm, 87% for PVS, 94% for ZP, and 86% for PBI. Fine-tuned models pretrained on a single oocyte dataset with external clinical images yielded better performance than ImageNet-based initialisation, highlighting the advantage of task-specific pretraining. CONCLUSIONS: These findings confirm the feasibility of multi-component oocyte segmentation and establish a basis for future quality assessment studies.