ROBUST CLASSIFICATION OF MAMMALIAN EMBRYOS AND OOCYTES BASED ON LABEL-FREE HYPERSPECTRAL IMAGING AND ARTIFICIAL INTELLIGENCE

A major challenge in the IVF clinic remains the robust scoring of oocyte and embryo developmental competence. Brightfield microscopes and time-lapse incubators provide valuable morphological and morphokinetic parameters, however, they provide limited physiological information. Metabolism is a sensitive alternative biomarker for developmental competence. Thus, here we report the development of a direct and non-invasive method to assess oocyte and embryo metabolism and its correlation with developmental potential. We have developed a novel non-invasive Hyper-Spectral (HS) imaging method which uses near infrared light to excite intrinsic metabolic signals of mammalian oocytes and embryos in a safe manner. The data retrieved encodes rich metabolic information derived from the simultaneous acquisition of multiple metabolites (6+). The complex hyperspectral data is transformed via dimensionality reduction, using phasor analysis, and then fed to an Artificial Intelligence (AI) algorithm. We used aSupportVector Machine (SVM) algorithm under the Akaike Information Criterion(AIC) model, and 80%/20% cross validation. For completeness, we used a 5-fold cross-validation repeated 50 times. A total of 96 mouse blastocysts and 178 oocytes were analyzed for the study. Embryo cohorts included control, glucose-starved, pyruvate/lactate-starved, and glucose/pyruvate/lactate-starved blastocysts. The AI algorithm correct classification of control blastocysts achieved an area under the curve (AUC) of 93.7%, significantly outperforming the human graders, which reached an average of 51% AUC. Oocyte cohorts included oocytes obtained from young female mice (< 4-weeks old) analyzed either immediately after collection or after overnight culture (in-vitro aged), and from old female mice (>12-months old). Our algorithm was able to correctly classify control oocytes with an average AUC of 96.2% and predict their probability to reach blastocyst stage after ICSI with an 82.2% AUC. Additional image segmentation was performed by combining two features of the FAD+ metabolite: mitochondrial confinement and a specific spectrum range. Using these criteria, we were able to extract HS enriched signals coming from mitochondria and measure a range of mitochondrial parameters including average size, signal intensity, variance or clustering. The analysis revealed that control embryos are better defined by the total intensity and variability of the signal, while oocytes are characterized by a specific distribution of the mitochondria in the ooplasm. Our HS imaging method offers a 4D (x, y, z, and spectrum) real-time image of the mammalian oocytes and embryo metabolism in a safe fashion. The robust classification of the samples stronglysupportsHS imaging as a promising new tool for the evaluation of embryo and oocyte developmental competence.