.png){kind=icon}
.png){kind=icon}
Return
From time-lapse to morphokinetics: neural ODE dynamics for reliable embryo stage transition timing
.png){kind=icon}
PRE
AI
M
J
M
F
A
N
DOI：10.1007/s10815-026-03960-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Time-lapse imaging (TLI) enables longitudinal and non-invasive embryo monitoring in IVF, yet morphokinetic annotation remains labor-intensive and subject to inter-operator variability. Automated detection of developmental phase transitions is challenging due to subtle morphological changes, heterogeneous acquisition conditions, discrete and potentially irregular image sampling, focal-plane variability, and the need for temporally consistent predictions. We aim to detect morphokinetic phase transitions in embryo TLI videos using a robust, clinically oriented framework designed to support human-in-the-loop annotation rather than replace clinical decision-making or commercial embryo selection systems. We propose a spatio-temporal Neural Ordinary Differential Equation (Neural ODE) model for continuous-time latent representation learning between discretely acquired observations, coupled with a reference-based transition scoring mechanism and an online, one-class detection strategy. The method, named Reference-Based Neural ODE Change Detector (RB-NODE), is transition-agnostic and detects when the evolving latent state deviates from a prototype representation of the current developmental phase. Experiments are conducted on the public dataset of Gomez et al., a single-center TLI benchmark comprising 704 embryo videos acquired across seven focal planes. All reported results are computed over the union of test samples generated from the seven focal-plane views, with strict embryo-level data splitting to prevent leakage across focal planes. RB-NODE achieves an AUC of 0.988, an F1@frame of 0.975, a Det@5f of 0.873, a Det@10f of 0.956, and a mean absolute localization error of 2.64 frames. Although a ResNet18+LSTM baseline obtains a marginally higher Det@5f, RB-NODE provides the best AUC, F1@frame, Det@10f, and mean temporal error among the evaluated methods. Event-group analysis shows particularly strong performance on morula/blastocyst-related transitions, with Det@5f = 0.910, Det@10f = 0.978, and a mean temporal error of 1.90 frames. RB-NODE also remains compatible with online use, processing frames at approximately 171 FPS on an NVIDIA RTX 6000 Ada Generation GPU. 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. The proposed framework is intended as an annotation-support tool, not as an autonomous embryo selection system. External validation across multiple IVF centers, TLI platforms, acquisition intervals, annotation protocols, and focal-plane configurations remains necessary before clinical deployment.
Keywords:
Embryo time-lapse imaging
IVF
Morphokinetics
Morphokinetic transition detection
Neural ordinary differential equations
Human-in-the-loop AI
Journal
IF:
2.7
Papers:
5.9K
Citations:
9.4K
Organization
Cited Papers