Towards Automated Selection Of Embryos For Ivf By The Investigation Of Changes In Relative Entropy Over Time.

We propose a novel image recognition framework, based on machine learning, to assess embryo quality and viability for in-vitro fertilization (IVF) treatment using time-lapse images obtained from a morphokinetic system (EmbryoScope). We quantified morphometric changes in time-lapse images obtained from the EmbryoScope using relative entropy, which measures changes in the distribution of pixel intensities over time. We processed the time-series data using dynamic time warping (DTW) to automatically assess the pace of development of embryos relative to each other. Lastly, we employed machine learning to predict embryo quality and interpreted results by computing the marginal effect of input features on the predicted outcome. Our data consisted of time-lapse images for 395 embryos obtained from frozen donor eggs at a university affiliated IVF clinic. Images were taken at an interval of 20 mins at 7 different focal planes over 1-6 days by the EmbryoScope. In addition, each embryo was associated with an outcome (cryo-preserved, transferred or discarded) and ground-truth assessment of embryo quality by embryologists at different time points. We preprocessed the images (segmentation and registration) prior to relative entropy calculation. For each patient, the RE data was normalized to a reference embryo, using the DTW algorithm to identify lag/lead during development. Embryo quality prediction was performed using a multi-layer perceptron and results were interpreted using partial dependence plots and the Local Interpretable Model-Agnostic Explanations (LIME) algorithm. Our investigation revealed that early milestones in embryo development (disappearance of pronuclei, stages of cleavage, compaction, etc.) correspond to spikes in relative entropy that can be easily detected. Embryos that stall in development exhibit persistently low entropy which indicates lack of cell division and rearrangement. The pace of development of an embryo, relative to a reference embryo was quantified using DTW and interesting cases where an embryo that lags in development early but is able to catch up later could be readily identified. Our preliminary findings from supervised classification suggest that the time taken to reach developmental milestones can be used to identify promising candidates for implantation as soon as day 3. We achieved classification accuracy of approx. 70% (excl. embryos that failed to develop) with limited observation up to advanced cleavage stage (9+ cells). Accuracy increased by 10-15% upon including time taken to form morula and blastocyst. We have successfully developed a novel methodology to assess embryo development in a clinical setting. Compared to traditional assessment by embryologists, our approach provides unbiased, quantifiable determination of changes in embryo morphology. Combined with our novel framework of DTW and machine learning, this approach provides an excellent tool for automated selection of quality embryos, which is transparent and easily interpretable.