EmbryoCST: A Stage-Aware ResNet with Layered CNN-3D Swin Transformer for Embryo Stage Classification

The classification of embryonic developmental stages serves as the foundation for embryo quality assessment in In-Vitro Fertilization (IVF) and related clinical research. Traditional methods rely on subjective morphological criteria, leading to significant limitations in consistency and objectivity. This paper proposes a stage-aware ResNet framework (EmbryoCST) that replaces traditional residual units with Spatiotemporal Fusion Residual Units (STFRUs), designed to efficiently capture complex spatiotemporal patterns and subtle morphological changes during embryonic development in an end-to-end manner. The STFRU integrates a Layered Conv-3D Swin Transformer Block (LCST), which performs subsampling and local feature fusion through convolutions in shallow layers, thereby reducing computational complexity while enhancing the recognition of subtle morphological changes. In deeper layers, a 3D Swin Transformer module enables parallel spatiotemporal modeling with global context awareness, effectively capturing complex dynamic features throughout embryonic development. Due to the continuity of embryonic development and the morphological similarity between adjacent stage image frames, the model faces significant difficulties in classifying adjacent stages. To alleviate this issue, a Stage-Aware Loss is designed to reduce misclassification between adjacent stages and mitigate stage class imbalance. Comparative experiments on two public embryo image datasets demonstrate that EmbryoCST outperforms state-of-the-art methods in embryonic developmental stage classification.