Artificial Intelligence (AI) for Embryo Ranking and its Use in Human Assisted Reproduction

Defined as computer algorithms/software that can perform tasks that typically require human intelligence (e.g. learning, problem-solving, visual perception), Artificial Intelligence (AI) is rapidly growing, with vast applications across various fields. Regarding medical imaging, X-rays, MRIs, predictive modelling from patient data, ophthalmology, dermatology, radiology, pathology and assisted reproduction technologies (ART) all could benefit. In ART, the practice of IVF has the potential to benefit in several areas, including patient management, gamete/embryo identification, ranking and selection. This thesis aims to facilitate selecting and ranking IVF embryos for transfer using static images and AI analysis. In particular:

• To test prototypes for predicting pregnancy test results after embryo transfer by image feature extraction and analysis using machine learning. This was achieved using an approach named AIR-e. The hypothesis that this novel approach is effective was accepted, indicating the feasibility of using AIR-e® to predict implantation potential from a single digital image.

• To test an Embryo Ranking Intelligent Classification Algorithm (ERICA), an AI clinical assistant with embryo ploidy and implantation predicting capabilities. This was achieved: the hypothesis that ERICA can be used to predict ploidy status in IVF embryos was accepted. Following training and validation, ERICA was more successful than random selection and experienced embryologists in ranking embryos with the highest implantation potential based on a static picture as the only source of information.

• To use ERICA to test the hypothesis that it can predict first-trimester pregnancy loss. In a retrospective pilot study, the hypothesis was accepted. Results support a correlation between the risk of spontaneous abortion and embryo rank as determined by ERICA.

• To develop the first automatic method for segmenting all morphological structures during the developmental stages of a human blastocyst. his was achieved. The approach can automatically segment blastocysts from different laboratory settings and developmental phases within a single pipeline. A sensitivity analysis established that this method is robust.

• To perform a systematic evaluation of predictive AI models in reproductive medicine. Here, a critical appraisal of AI models in reproductive medicine is discussed, conveying the importance of transparency and standardisation in reporting AI models so that the risk of bias and the potential clinical utility of AI can be assessed. Four reasons to interpret reproduction AI studies with caution are given, as is a guide to appraise AI's efficacy in reproductive medicine critically. Finally, a quick reference reader and referee consideration guide are provided.

AI’s ability to make decisions based on facts and data makes the decision process reproducible and repeatable. AI can learn and analyse complex patterns at an increased resolution and with more variables far beyond most humans’ capabilities. This thesis thus demonstrates that AI has the potential to be utilised as a promising tool to resolve many longstanding challenges in ART and assist clinicians in decision-making to achieve the ultimate goal of a healthy live birth. Similar principles may apply to sperm and egg analysis and other challenges in ARTs. Barriers include health record privacy terms, paper records and variations in electronic medical record systems. AI will play a significant role in the future of IVF that, realistically, will only be achieved if the artisanal approach of manual handling, basic microscopy and subjective analysis is replaced. The most likely combination of modalities to achieve scalability and improved access to IVF will be automation and AI.