Artificial intelligence outperforms humans in morphology-based oocyte selection in cattle

Abstract

Evaluating cumulus-oocyte complex (COC) morphology is commonly used to assess oocyte quality. However, clear guidelines on interpreting COC morphology data are lacking as this evaluation method is subjective. In the present study, individual in vitro embryo production was used, allowing follow-up of blastocyst formation for each COC. Images of immature COCs were presented to embryologists and two artificial intelligence (AI) models: deep neural network (DNN) and random forest classifier (RF). The aims were to (1) determine the most relevant morphological characteristics in distinguishing qualitative COCs, (2) review human-made predictions, and (3) build predictive AI models. Our experiments identified cumulus size as pivotal characteristic of COC quality, while embryologists assigned ooplasm morphology as most important. Inspection of COCs by the human eye showed significant limitations, as evidenced by their low predictive ability (balanced accuracy: 42.9%) and fair reliability. Our AI models outperformed the embryologists, yielding a balanced accuracy of 79.3% and 71.2% for DNN and RF, respectively. The first AI models that successfully predict developmental competence of immature bovine oocytes were created, outperforming embryologists and offering an objective perspective for COC morphology assessment. AI has emerged as a novel tool for oocyte appreciation, assisting decision-making in the embryology lab.