P–064 Application of an artificial intelligence model for morphologic prediction of fertilization-competent human spermatozoa

Abstract

Abstract Study question What is the role of artificial intelligence in selecting fertilization-competent human spermatozoa according to their morphological characteristics? Summary answer: The established AI model in this study can be potentially used to select semen samples with superior fertilization potential in clinical settings. What is known already Defective spermatozoa-zona pellucida (ZP) interaction causes subfertility and is a major cause of low IVF fertilization rates. While ICSI benefits patients with defective spermatozoa-ZP binding, a standard method to identify such patients prior to conventional IVF is lacking. The application of artificial intelligence to sperm morphology analysis has become a topic of growing interest owing to the fact that the conventional assessment is highly subjective and time-consuming. Deep-learning, a core element of artificial intelligence (AI), incorporates the convolutional neural networks (CNN) to process all the data composing a digital image through successive layers to identify the underlying pattern. Study design, size, duration The fertilization-competent spermatozoa were isolated according to their binding ability to the ZP. The ZP-bound and -unbound spermatozoa were collected for functional assays and to establish an AI model for morphologic prediction of sperm fertilization potential. Human spermatozoa (n = 289) were isolated from normozoospermic samples. Human oocytes (n = 562) were collected from an assisted reproduction program in Hong Kong. Sample collection has been ongoing and will continue until the end of this study in November 2021. Participants/materials, setting, methods Sperm-ZP binding assay was employed to collect ZP-bound and -unbound spermatozoa. The fertilization potential and genetic quality of the collected spermatozoa were evaluated by our established protocols. Diff-Quik- stained images of ZP-bound and -unbound spermatozoa were collected respectively for the establishment of an AI model. A novel algorithm for sperm image transformation and segmentation was developed to pre-process the images. CNN architecture was then applied on these pre-processed images for feature extraction and model training. Main results and the role of chance Our result showed that the sperm-ZP binding assay had no detrimental effect on sperm viability when compared with the raw samples and unbound-sperm subpopulations. ZP-bound spermatozoa were found with statistically higher acrosome reaction rates, improved DNA integrity, better morphology, lower protamine deficiency and higher methylation level when compared with the unbound spermatozoa. A deep-learning model was trained and validated by analyzing a total of 1,334 and 885 of ZP-bound/unbound spermatozoa to evaluate the predictive power of sperm morphology for ZP binding ability. Our newly trained AI-based model showed initial success in classifying the ZP-bound/ unbound spermatozoa according to their morphological characteristics with high accuracy of 85% and low computational complexity. Limitations, reasons for caution This sperm selection method requires micromanipulation and relatively long processing time to recover ZP-bound spermatozoa. In addition to limited availability, the use of human materials may result in interassay variations affecting the reproducibility of this method among laboratories. Wider implications of the findings: In light of current findings, AI-based sperm selection method may provide high predictive values of sperm fertilization potential for clinical purposes. This method is particularly applicable to patients who had poor fertilization outcomes after conventional IVF treatments or those with high degree of defective sperm-ZP binding ability. Trial registration number Not applicable