An interpretable and versatile machine learning approach for oocyte phenotyping

Abstract

Meiotic maturation is a crucial step of oocyte development allowing its potential fertilization and embryo development. Elucidating this process is important both for fundamental research and assisted reproductive technology. However, only few computational tools, based on non-invasive measurements, are currently available to characterize oocyte meiotic maturation. Here, we develop a computational framework to phenotype oocytes based on images or movies acquired exclusively in transmitted light. We first trained neural networks to segment the contours of oocytes and their zona pellucida using a diverse cohort of both mouse and human oocytes. We then defined a comprehensive set of morphological features to describe a single oocyte. We have implemented these steps in a versatile and user-friendly open source Fiji plugin available to the mouse and human oocyte community. Then, we present a machine learning pipeline based on selected features to automatically recognize oocyte populations and determine their morphological differences. Its first application is a novel approach to screen oocyte strains and automatically identify their morphological characteristics. We demonstrate its potential by phenotyping a well characterized genetically modified mouse oocyte strain. Its second application is to predict and characterize the maturation potential of oocytes. Importantly, we identify two new features to assess mouse oocyte maturation potential, consisting in the texture of the zona pellucida and the cytoplasmic particles size. Eventually, we tested whether these mouse oocyte quality features were applicable to human oocytes developmental potential.