Purpose: The advent of time-lapse incubators has expanded the scope and volume of information accessible to embryologists, facilitating more informed decisions regarding embryo viability. One predictor that is gaining promise is accurate estimations of mophokinetic changes. Unfortunately, manually labeling these events is both susceptible to embryologist variation and is resource-intensive. Addressing these challenges, we harnessed AI to automate annotations and foster transparency by open-sourcing our model.
Methods: We utilized imaging data from a Vitrolife EmbryoScope, capturing images every 20 minutes from 413 patients across five fertility clinics in New Zealand. This dataset encompassed 104,418 images from 1612 embryos. Manual annotation by a senior embryologist defined 13 morphokinetic stages, including tPNa, tPNf, t2, t3, t4, t5, t6-7, t8, t9+, tM, tB, tEB.
For automated morphokinetic estimation, we trained a Convolutional Neural Network (CNN) on the images. Data was partitioned into training (87.2%), testing (3.2%), and validation (9.6%) sets. The CNN employed a Resnet-50 backbone for feature extraction, which was combined with elapsed time since fertilization converted to a time vector. A transformer layer with a multi-headed attention mechanism estimated interrelationships between imaging features and time, culminating in predictions from a fully connected layer. The final time for predictions are made via a sliding window approach.
Results: The model exhibited an impressive overall accuracy of 81.2% and when taking embryologist subjectivity into account the model accuracy rose to 96%. This high performance is due to the fusion of images with timing information. This amalgamation enhanced the model's comprehension of both the image and temporal context, enabling more precise predictions than those derived solely from single images.
Conclusion: In conclusion, our study introduces a powerful deep-learning model that automates human embryo morphokinetic annotation, attaining high accuracy and showcasing promising implications for integration into the clinic.