Automatic identification of human spermatozoa with zona pellucida-binding capability using deep learning

Can a deep-learning algorithm, independent of World Health Organization (WHO) sperm morphology grading, be used to identify human spermatozoa with zona pellucida (ZP)-binding capability in assisted reproductive technology (ART)?

A novel deep-learning model, irrespective of the conventional semen analysis, was established to identify human spermatozoa capable of binding to ZP for predicting their fertilization potential.

Sperm morphology evaluation is crucial in semen analysis to investigate male infertility and to determine the appropriate insemination methods in ART. The current manual assessment, which relies on microscopically examining individual spermatozoa based on WHO criteria, has shown limited predictive power for fertilization outcomes due to its highly subjective, labour-intensive nature, and high inter-/intra-assay variations. Deep learning is a rapidly evolving method for automated image analysis. Recent studies have explored its potential for automating sperm morphology analysis. However, algorithms trained on manually annotated datasets using existing WHO criteria have had little success in predicting ART outcomes. To date, no study has established an independent set of morphology evaluation standards based on sperm fertilizing ability for clinical prediction.

Spare semen samples were collected from men undergoing premarital check-ups at a family planning clinic. Immature oocytes at germinal vesicle/metaphase I stage, or mature metaphase II oocytes were donated from women attending the infertility clinic for assisted reproduction treatments. Acrosome-intact, ZP-bound spermatozoa were collected by our previously modified spermatozoa-ZP coincubation assay. ZP-unbound spermatozoa were collected from normozoospermic samples with defective ZP-binding ability, as evidenced by complete fertilization failure following conventional in vitro fertilization (IVF) and the absence of ZP-bound spermatozoa on the inseminated oocytes. A total of 1083 Diff-Quik stained images of ZP-bound and unbound spermatozoa were collected to create a training database, with an additional 220 images serving as an independent test set. Clinical data were obtained from 117 men undergoing IVF due to male factor or unexplained infertility to validate the model's ability to generalize to new data. These participants were categorized into three groups based on their fertilization rates following IVF: low (0-40%), intermediate (41-70%), and high (71-100%).

A pre-trained VGG13 model was fine-tuned using our database to classify individual spermatozoa as either ZP-bound or unbound based on their automatically extracted morphological features. Confusion matrix was used to assess the model's classification performance, expressed in terms of accuracy, specificity, sensitivity, and precision rates. The area under the receiver-operating characteristic (ROC) curve (AUC) was utilized to measure the model's discriminative power. A 5-fold cross-validation was conducted on the training dataset to assess the model's performance on randomized subgroups. Saliency mapping was used to analyse pixel importance localized to the morphological features of sperm images. Clinical data of spermatozoa from three fertilization groups were used for clinical validation. Logistic ROC regression analysis was performed to evaluate the differences in predicted values between high and low fertilization groups, as indicated by AUC and P-values. Additionally, Youden's index was applied to determine a clinical threshold for predicting IVF fertilization outcome using the model.

A VGG13 model was fine-tuned to distinguish images of spermatozoa capable of binding to the ZP based on their morphological features with high sensitivity (97.6%), specificity (96.0%), accuracy (96.7%), and precision (95.2%). The model exhibited low learning variance (average accuracy: 97.4%; sensitivity: 96.0%; and specificity: 98.5%) across subgroups, with primary emphasis on the sperm head and mid-pieces in all images as indicated by the pixel importance. Its discriminative performance was clinically validated on over 33 000 sperm images collected from three fertilization groups. Overall, the model exhibited excellent generalization ability as reflected by the strong correlation between the predicted percentages of spermatozoa with ZP-binding per sample and their fertilization rates. A clinical threshold of 4.9% (specificity: 89.3%; sensitivity: 90.0%) was established to differentiate sperm samples with normal and defective ZP-binding ability. By conducting pairwise comparisons among 30 patients, the predicted values generated by the model outperformed conventional semen analysis assessed by our in-house embryologists in identifying patients who were likely to experience failure with conventional IVF.

N/A.

The model is currently designed for high-resolution, air-dried, Diff-Quik stained sperm samples, and further research is required to validate its classification performance across different image qualities with a larger sample size.

This newly established method can identify couples at high risk of unexpected IVF fertilization failure, enabling clinicians to offer alternative insemination methods to reduce the likelihood of suboptimal fertilization outcomes.

This study was supported in part by two Health and Medical Research Funds, the Food and Health Bureau, The Government of the HKSAR (07182446 and 11222236), and the Sanming Project of Medicine in Shenzhen (SZSM 202211014). Two provisional patent applications related to the data presented here have been filed on behalf of The University of Hong Kong (i. application no. 63/511,375; filing date: 30 June 2023; current status: active; applicant: The University of Hong Kong; ii. application no. US 63/567,147; filing date: 19 March 2024; current status: active; applicant: The University of Hong Kong). The authors declare that they have no other competing interests.