Prediction of Progression from Oocyte to Useable Blastocyst Using Deep Radiomic Signatures: Preliminary Results

Abstract

Background: Oocyte cryopreservation is increasingly used for fertility preservation by young women with cancer prior to chemo/radiotherapy and by single women wishing to defer pregnancy beyond their mid-30’s, thereby avoiding the increased risks of infertility, miscarriage and chromosomally abnormal pregnancy seen as consequences of female ageing. Oocyte cryopreservation is inefficient, with a large number of oocytes being required to give a realistic possibility of even a single pregnancy. We report preliminary data that demonstrate the ability of deep radiomic imaging to predict progression from oocyte to useable blastocyst as a first step towards improving the efficiency of oocyte cryopreservation. Aim: To develop and apply a deep radiomic signature (DRS) based on an artificial intelligence model to identify oocytes that progressed to useable blastocysts. Method: 40× brightfield images (n = 45) of human oocytes from 6 patients, taken prior to ICSI using an Olympus DP27 camera and IX73 inverted microscope. An unsupervised deep learning approach was applied to extract informative DRS of oocyte morphology. Unsupervised analysis by a fully blinded operator was conducted to identify data clusters. The data were colour-coded to identify useable (suitable for cryopreservation or transfer) blastocyst formation and a classifier was developed to separate useable blastocysts. This study was approved by the South Eastern Sydney Human Research Ethics Committee (2020/ETH01767). Results: Blinded analysis identified an effective DRS signature with two separate clusters of oocytes. Oocytes that led to useable blastocysts formed a single cluster which was denser than that from oocytes with no good blastocyst outcome, implying less morphological heterogeneity of the higher quality oocytes. The classifier could successfully distinguish oocytes which did/did not lead to useable blastocysts with strong performance (AUC=0.86). Conclusion: Our preliminary data demonstrate that non-invasive DRS can be applied to human oocytes, identifying morphological features that predict useable blastocyst formation with high accuracy.