In the field of assisted reproduction, time-lapse technology can collect embryo images across multiple focal planes, which helps embryologists stage embryos and dynamically evaluate their quality, thereby improving the success rate of transplantation. Clinical practitioners rely on integrated information from various focal planes, as each plane encompasses information from all cells, considering the influence of depth of field. However, existing methods predominantly focus on single-plane image acquisition, either neglecting comprehensive information or failing to exploit internal correlations. To address this issue, we propose a method named Focal Consistency Network (FC-Net) for processing time-lapse embryo video datasets and classifying embryo developmental stages. The FC-Net comprises a classification head and a multi-focal consistency head. While the classification head learns the categories of images from different focal planes at the same time, the multi-focal consistency head ensures consistency between the predictions of other focal planes and the main focal plane, facilitating the model’s learning of more stable feature information. The method demonstrates significantly superior performance on publicly available time-lapse embryo video datasets compared to other models, achieving a success rate increase of 3% points. Furthermore, visual analysis of the results confirms the alignment of the predicted embryo developmental stage results with the actual scenario, further validating the effectiveness and superiority of the proposed method.