Abstract
The use of time lapse systems (TLS) in In Vitro Fertilization (IVF) labs to record developing embryos has paved the way for deep-learning based computer vision algorithms to assist embryologists in their morphokinetic evaluation. Today, most of the literature has characterized algorithms that predict pregnancy, ploidy or blastocyst quality, leaving to the side the task of identifying key morphokinetic events. Using a dataset of N = 1909 embryos collected from multiple clinics equipped with EMBRYOSCOPE/EMBRYOSCOPE+ (Vitrolife), GERI (Genea Biomedx) or MIRI (Esco Medical), this study proposes a novel deep-learning architecture to automatically detect 11 kinetic events (from 1-cell to blastocyst). First, a Transformer based video backbone was trained with a custom metric inspired by reverse cross-entropy which enables the model to learn the ordinal structure of the events. Second, embeddings were extracted from the backbone and passed into a Gated Recurrent Unit (GRU) sequence model to account for kinetic dependencies. A weighted average of 66.0%, 67.6% and 66.3% in timing precision, recall and F1-score respectively was reached on a test set of 278 embryos, with a model applicable to multiple TLS.