From Time-Lapse to Morphokinetics: Neural ODE Dynamics for Reliable Embryo Stage Transition Timing

Problem : Time-lapse imaging (TLI) enables continuous embryo monitoring in IVF, yet morphokinetic annotation remains labor-intensive and subject to interoperator variability. Automated detection of developmental phase transitions is challenging due to subtle morphological changes, heterogeneous acquisition conditions, and the need for temporally consistent predictions. Aim : We aim to detect morphokinetic phase transitions in embryo TLI videos using a robust, clinically oriented framework designed to support human-in-theloop annotation rather than replace clinical decision-making. Methods : We propose a spatio-temporal Neural Ordinary Differential Equation (Neural ODE) model for continuous-time representation learning, coupled with a transition scoring mechanism and an online, one-class training strategy. The method named Reference-Based Neural ODE Change Detector (RBNODE) is transition-agnostic and relies on a reference trajectory learned from normal developmental dynamics to detect change points. Experiments are conducted on the public dataset of Gomez et al. (704 embryo videos, seven focal planes, 16 developmental phases). Results : The proposed method achieves an AUC of 0.988 for transition detection and reaches 0.873 detection accuracy within a 5-frame tolerance (Det@5f). The estimated transition time is close to expert annotations, with a mean absolute error of 2.64 frames and a median error of 1 frame. Performance remains stable across key transitions, including blastocyst-related stages, highlighting the benefit of continuous-time temporal modeling for reducing temporal instability. Conclusion : Continuous-time spatio-temporal modeling with Neural ODEs provides a promising and reproducible approach for robust morphokinetic transition detection in embryo TLI, offering practical support for standardized annotation workflows and future prospective validation in clinical settings.