Automated identification of aneuploid cells within the inner cell mass of an embryo using a numerical extraction of morphological signatures

Abstract

STUDY QUESTION: Can artificial intelligence distinguish between euploid and aneuploid cells within the inner cell mass of mouse embryos using brightfield images? SUMMARY ANSWER: A deep morphological signature (DMS) generated by deep learning followed by swarm intelligence and discriminative analysis can identify the ploidy state of inner cell mass (ICM) in the mouse blastocyst-stage embryo. WHAT IS KNOWN ALREADY: The presence of aneuploidy, a deviation from the expected number of chromosomes, is predicted to cause early pregnancy loss or congenital disorders. To date, available techniques to detect embryo aneuploidy in IVF clinics involve an invasive biopsy of trophectoderm cells or a non-invasive analysis of cell-free DNA from spent media. These approaches, however, are not specific to the ICM and will consequently not always give an accurate indication of the presence of aneuploid cells with known ploidy therein. STUDY DESIGN, SIZE, DURATION: The effect of aneuploidy on the morphology of ICMs from mouse embryos was studied using images taken using a standard brightfield microscope. Aneuploidy was induced using the spindle assembly checkpoint inhibitor, reversine (n = 13 euploid and n = 9 aneuploid). The morphology of primary human fibroblast cells with known ploidy was also assessed. PARTICIPANTS/MATERIALS, SETTING, METHODS: Two models were applied to investigate whether the morphological details captured by brightfield microscopy could be used to identify aneuploidy. First, primary human fibroblasts with known karyotypes (two euploid and trisomy: 21, 18, 13, 15, 22, XXX and XXY) were imaged. An advanced methodology of deep learning followed by swarm intelligence and discriminative analysis was used to train a deep morphological signature (DMS). Testing of the DMS demonstrated that there are common cellular features across different forms of aneuploidy detectable by this approach. Second, the same approach was applied to ICM images from control and reversine treated embryos. Karyotype of ICMs was confirmed by mechanical dissection and whole genome sequencing. MAIN RESULTS AND THE ROLE OF CHANCE: The DMS for discriminating euploid and aneuploid fibroblasts had an area under the receiver operator characteristic curve (AUC-ROC) of 0.89. The presence of aneuploidy also had a strong impact on ICM morphology (AUC-ROC = 0.98). Aneuploid fibroblasts treated with reversine and projected onto the DMS space mapped with untreated aneuploid fibroblasts, supported that the DMS is sensitive to aneuploidy in the ICMs, and not a non-specific effect of the reversine treatment. Consistent findings in different contexts suggests that the role of chance low. LARGE SCALE DATA: N/A LIMITATIONS, REASON FOR CAUTION: Confirmation of this approach in humans is necessary for translation. WIDER IMPLICATIONS OF THE FINDINGS: The application of deep learning followed by swarm intelligence and discriminative analysis for the development of a DMS to detect euploidy and aneuploidy in the ICM has high potential for clinical implementation as the only equipment it requires is a brightfield microscope, which are already present in any embryology laboratory. This makes it a low cost, a non-invasive approach compared to other types of pre-implantation genetic testing for aneuploidy. This study gives proof of concept for a novel strategy with the potential to enhance the treatment efficacy and prognosis capability for infertility patients.