The use of machine learning in assessing mammalian pre-implantation embryo quality

In Vitro Fertilisation (IVF) is quickly becoming an extremely important medical intervention as the prevalence of infertility increases. Therefore, it is vital to ensure that IVF procedures are as safe and successful as possible. There are still many challenges to be addressed, including the embryo assessment process used to select an embryo for transfer. This is a difficult task, as knowledge of pre-implantation development is incomplete and there is still a high degree of subjectivity involved in assessing embryos for viability. In this thesis, we work towards more accurate, objective and versatile embryo selection by employing machine learning (ML) techniques and investigating a range of potential morphological quality markers, including currently neglected sub-cellular features. We first present CNN models trained to predict live birth from a variety of developmental stages. These include the first DL models predicting live birth using solely pre-blastocyst stages, which could allow for earlier embryo transfer, mitigating the harmful effects of prolonged culture. We also showed that information from earlier stages can assist selection at blastocyst stage, allowing for the previously unachievable ranking of high-quality blastocysts. In developing these models, we explored the time period of pre-implantation development to identify the best developmental moments for predicting live birth, therefore providing crucial information for embryo assessment procedures. In order to achieve the best possible assessment of embryo quality at pre-blastocyst stage, we next identified morphological features correlated with transfer outcome and combined these with the CNN model outputs to get an improved overall prediction of live birth. We experimented with different supervised learning techniques and found that linear regression gave the best performance. We also investigated the structure of our dataset via dimensionality reduction techniques and unsupervised clustering, gaining a deeper insight into the challenge of embryo selection. Finally, we investigated changes in nuclear size and appearance during preimplantation development, a sub-cellular feature not currently used to assess embryo viability past the first two embryonic cycles. For this we used the mouse embryo, which mirrors key developmental stages in human embryos. We discovered trends in size and shape both over development, and across different lineages at the same developmental stage. This research could pave the way for better understanding of standard nuclei appearance during pre-implantation development, allowing for the existing embryo selection criteria to be extended. In this thesis we have demonstrated the potential of ML techniques to increase knowledge of the pre-implantation development period, and ultimately lead to improved embryo selection procedures in IVF.