Deng, L.; Lin, P.; Xie, L.

Quantitative changes in zebrafish embryonic developmental tempo are key phenotypes in drug screening, genetic perturbation, and environmental stress studies. However, manual staging is labor-intensive and subjective, and nominal hours post fertilization (hpf) often fail to reflect true developmental progression under condition shifts such as temperature changes. Existing automated approaches predominantly rely on single-frame image analysis and do not explicitly enforce temporal coherence when applied to dense time-lapse sequences; moreover, treating correlated frames or overlapping clips as independent observations can introduce pseudo-replication and overconfident statistical inference. To address these challenges, we propose EmbryoTempoFormer (ETF), a clip-based CNN-Transformer model that predicts developmental time from short brightfield time-lapse clips and is trained with a within-embryo temporal-difference consistency regularizer to improve trajectory self-consistency across overlapping predictions. In a downstream inference and statistical workflow, correlated clip predictions are summarized into interpretable embryo-level tempo readouts and compared across conditions using embryo-bootstrap confidence intervals, treating embryos as independent statistical units. Overall, ETF and the accompanying embryo-resolved workflow provide a robust and reproducible solution for high-throughput developmental phenotyping from time-lapse imaging.