#62 : Time-Lapse Imaging Analysis of Embryo Development: A Dynamic-Image-Based Prediction Model for Blastocyst Formation During Early 80 Hours Development

Abstract

Background and Aims: The objective is to develop a novel algorithm for context learning through time-lapse imaging and compare its performance with static learning for blastocyst formation prediction. Method: Time-lapse imaging (TLI) was used to capture images of over 400 embryos between 19 to 80 hours after insemination with approximately 30,000 frames. The study employed two network input strategies for blastocyst formation prediction - utilizing static (A model) and dynamic image-based (B model) prediction models. The proposed algorithm for context learning through time-lapse imaging involves deep learning methods for identifying subtle changes in the embryo’s developmental stages through the time series of images. The convolutional neural network (CNN) and long short-term memory (LSTM) methods are leveraged to extract features and maintain context over the culture duration. Results: The study found that early-stage prediction of blastocyst formation was challenging, with both network input strategies achieving an overall prediction accuracy of less than 80%. However, B model prediction outperformed A model prediction, with a prediction value of 75% compared to 68% (p 1 days) being correctly predicted compared with 36.4% for the A model. Conclusion: In conclusion, the study demonstrated that a dynamic-image-based prediction model utilizing the proposed algorithm for context learning through TLI outperforms the static-image-based prediction model. The proposed algorithm enables a more comprehensive analysis of the embryos, considering subtle changes over time, which enables improved prediction of blastocyst formation with more robust performance in the future. This study represents a significant step forward in the development of advanced data mining methods that can enable reliable and accurate predictions of embryo development.