P-049 Identification of human spermatozoa with zona pellucida-binding ability using deep learning

Abstract

Abstract Study question Can a deep learning model be used for identification of fertilization-competent human spermatozoa based on their distinctive morphological features? Summary answer Our results validated the clinical significance of image analysis of sperm morphology using deep learning to identify high-quality spermatozoa with zona pellucida (ZP)-binding ability. What is known already Sperm morphology is manually evaluated as part of the conventional semen analysis in clinical settings. However, the clinical predictive power of sperm morphology remains debatable due to its lack of association with the fertilization potential of spermatozoa. Human zona pellucida (ZP)-binding ability is a measurable, independent sperm parameter associated with in vitro fertilization (IVF) and pregnancy outcome. Supervised deep learning is an advanced branch to systematically identify the underlying relationship between the input images and designated output. A well-trained model with high generalization ability can generate plausible predictions for unseen images collected under the same condition as the training datasets. Study design, size, duration Human spermatozoa and oocytes were obtained with written consents from the donors attending our assisted reproduction program. Acrosome intact, ZP-bound spermatozoa were recovered after 30 min of gamete coincubation. Unbound spermatozoa were collected from clinical samples of spermatozoa with defective ZP-binding ability leading to fertilization failure following conventional insemination. Sperm images of infertile men (n = 165) with fertilization rates including 0-40% (low), 41-70% (intermediate) and 71-100% (high) following conventional insemination were collected for validation. Participants/materials, setting, methods Diff-Quik stained images of spermatozoa were processed by the K-means clustering algorithm to generate grey-scaled images of individual sperm heads. All sperm images of clinical samples were subjected an additional image conversion step using Cycle Generative Adversarial Network (CycleGAN) to ensure that their microenvironments are similar to those of the laboratory samples. A VGG 13 model was fine-tuned to classify a total of 1,083 images of ZP-bound and unbound spermatozoa over 50 epochs. Main results and the role of chance The newly fine-tuned VGG-13 model demonstrated excellent performance in classifying the ZP-bound and unbound spermatozoa, as reflected by the high sensitivity of 97.6%, specificity of 96.0%, and accuracy of 96.4%. It had a high discriminative power to distinguish between ZP-bound and unbound spermatozoa with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.992. Saliency maps indicated that the model focused on the sperm head and mid-pieces in all images, with a primary emphasis on the anterior regions (43.0%±5.0%). The model was further validated on a vast collection of over 40,000 sperm images collected from the aforementioned fertilization groups. Our results revealed that the high fertilization rate group had a significantly higher percentage of predicted ZP-binding ability than the low fertilization rate group, as determined by the model (16.8%±7.2 vs. 3.2%±1.4; p < 0.05). By analyzing sperm images from infertile men (n = 113) in the high and low fertilization rates groups, a clinical threshold of 4.8% was identified using ROS curve (AUC:0.970) and Youden’s index, which can be used to distinguish between ZP-bound and unbound spermatozoa of clinical samples with a specificity of 93.8% and a sensitivity of 90.9%. Limitations, reasons for caution Our deep learning model is only applicable to diff-quik stained images of spermatozoa captured at high magnification. Further investigation is needed to examine its classification performance on images captured by different platforms and settings, ensuring consistent and reliable prediction accuracy regardless of varying degrees of image quality. Wider implications of the findings Our newly fine-tuned model can be used to evaluate sperm fertilizing potential, irrespective of the parameters evaluated in conventional semen analysis and to identify normospermic patients with defective ZP-binding ability who may benefit from intracytoplasmic sperm injection (ICSI) to improve their fertilization outcome. Trial registration number not applicable