Back to list
Six consecutive time-lapse images over 2 hours on day 3 can predict blastulation better than a single image
Oct 25, 2022
Hye Jun Lee, M.D. , Sung Han Woo, B.S. , Jong Hyuk Park, Ph.D. , Ji-Ye Jung, M.D. , Sungwook Choi, M.D. Hyung Min Kim, M.S. , Taehoon Ko, Ph.D.


To evaluate the convolutional neural network (CNN) model to predict blastulation of day 3 embryos using time-lapse images.


We retrospectively collected 10,223 time-lapse videos from 1,015 patients at two private fertility clinics and developed a blastulation prediction model using two datasets. The first dataset consisted of only a single frame image at 66 hours post-insemination on day 3, and the second one extracted six frame images from 66 to 68 hours on the same day. We used a pre-trained VGG-16 model and applied fine-tuning methods to train our datasets. Four performance metrics were compared: area under the receiver operating characteristic curves (AUROC), F1-score, and accuracy. We also used gradient-weighted class activation mapping (Grad-CAM) to verify that the model concentrated on the relevant features of the embryos.


In predicting blastulation on day 5 from a single frame image of day 3, the VGG16 pretrained model showed performance of AUROC 0.71 and accuracy 0.61. When six consecutive images of day 3 embryos were used, the model showed higher performance of AUROC 0.85 and accuracy 0.76. The cut-off value that maximized the F1-score was selected for the performance metrics.


Accurate prediction of blastulation potentially increases pregnancy rates as embryos unlikely to survive till day 5 can be successfully transferred or frozen at a cleavage stage. In this study, we confirmed that a single Time-lapse image of day 3 embryos is very useful in predicting blastulation at fair robustness of AUROC 0.71. Furthermore, we demonstrated that using 6 consecutive images improved prediction accuracy. The 6 images were closely examined by the experienced embryologists but the developmental stage did not advance notably over the 2 hours. It is well known that the CNN model extracts features from each image, but does not reflect associations between successive frames. Therefore, we expect that our model can further improve its performance by combining CNN and recurrent neural network (RNN) that reflects successive features.


The blastulation prediction model using a single image on day 3 can empower embryologists to make better clinical decisions, potentially leading to less embryo waste and higher pregnancy rates. It is noteworthy that multiple images consecutively taken over a couple of hours can further improve the predictive power for blastulation.