From time-lapse to morphokinetics: neural ODE dynamics for reliable embryo stage transition timing

Area of Science:

• In vitro fertilization (IVF)
• Embryology
• Artificial Intelligence in Medicine

Background:

• Time-lapse imaging (TLI) allows non-invasive embryo monitoring during IVF.
• Manual annotation of embryo development is time-consuming and prone to errors.
• Automating the detection of key developmental transitions is difficult due to subtle changes and variable imaging conditions.

Purpose of the Study:

• To develop a robust, clinically relevant framework for detecting morphokinetic phase transitions in embryo TLI videos.
• To create a tool that supports, rather than replaces, human annotation in IVF workflows.

Main Methods:

• Proposed a spatio-temporal Neural Ordinary Differential Equation (Neural ODE) model for continuous-time analysis.
• Implemented a reference-based transition scoring and an online, one-class detection strategy (RB-NODE).
• Validated the model on a public dataset of 704 embryo TLI videos across multiple focal planes.

Main Results:

• RB-NODE achieved high performance with an AUC of 0.988 and F1@frame of 0.975.
• Demonstrated superior performance in AUC, F1@frame, and temporal error compared to a baseline model.
• Showcased strong results for morula/blastocyst transitions and efficient online processing capabilities (171 FPS).

Conclusions:

• Continuous-time modeling with Neural ODEs offers a reproducible method for robust embryo morphokinetic transition detection.
• The RB-NODE framework can enhance standardization in IVF annotation workflows.
• Further external validation is required for clinical deployment.