An artificial intelligence real-time rare sperm detection system for intraoperative microsurgical testicular sperm extraction

Abstract STUDY QUESTION How can rare spermatozoa be identified efficiently during microsurgical testicular sperm extraction (micro-TESE)? SUMMARY ANSWER An artificial intelligence (AI)-assisted system was developed to flag candidate rare spermatozoa in real time during micro-TESE, and it may support embryologists as a decision-support tool. WHAT IS KNOWN ALREADY Patients with non-obstructive azoospermia (NOA) can obtain sperm for procreation through micro-TESE. During this procedure, sperm retrieval primarily relies on embryologists or laboratory technicians visually searching for sperm under a microscope, which is not only laborious and inherently subjective, but also susceptible to errors. Although AI technology has been applied to identify trace amounts of sperm, existing models lack sufficient efficiency and true real-time performance. STUDY DESIGN, SIZE, DURATION This study included model development followed by a single-centre clinical evaluation. An improved YOLO (You Only Look Once)-based rare sperm detection model, termed YOLOv11-RSD, was developed using microscopy data from 1,165 surgical patients, comprising 1,932 image samples containing a total of 5,032 annotated sperm objects with confirmed identification. Clinical evaluation was performed between May 2024 and July 2025. Performance was assessed across confidence thresholds in obstructive azoospermia (OA) patients with normal spermatogenesis, and the system was then applied during micro-TESE in NOA patients and compared with routine embryologist assessment. PARTICIPANTS/MATERIALS, SETTING, METHODS The model was developed using testicular sperm microscopy images collected at a single hospital. Real-time clinical feasibility was evaluated in 10 OA cases and 30 NOA cases. Embryologist assessment was used as the reference standard, and performance was assessed using PPV, sensitivity, F1-score and 95% confidence intervals. Discordant AI-assisted detections were reviewed by embryologists in real time. MAIN RESULTS AND THE ROLE OF CHANCE YOLOv11-RSD achieved real-time detection of candidate spermatozoa in microscopy images with high sensitivity and acceptable PPV under the selected operating threshold. Compared with baseline YOLOv11, YOLOv11-RSD showed improved overall detection performance across representative evaluation settings. In OA cases, the system achieved high sensitivity for sperm detection, reaching up to 96.7% across evaluated thresholds. During micro-TESE in NOA patients, at a confidence threshold of 0.50, positive predictive value (PPV), sensitivity and F1-score were 80.58%, 96.11% and 87.66%, respectively. The system highlighted candidate spermatozoa that were not identified during the initial manual assessment in six NOA cases, including two cases initially classified as sperm-negative; these findings were confirmed upon immediate re-review. Follow-up reproductive outcomes were available for six cases in which AI-assisted detection contributed to the search-and-confirmation workflow: embryo cleavage was achieved in all six cases, and three cases ultimately resulted in live births. Notably, among the two cases initially classified as sperm-negative, one case resulted in a singleton live birth. LARGE SCALE DATA N/A LIMITATIONS, REASONS FOR CAUTION This was a single-centre clinical evaluation with a limited clinical cohort. Although model inference was rapid, procedure-level efficiency was constrained by image acquisition and scanning logistics, and no definitive reduction in total procedure time was demonstrated. External multi-centre validation is required. WIDER IMPLICATIONS OF THE FINDINGS AI-assisted sperm detection may support embryologists during micro-TESE by flagging candidate rare spermatozoa for rapid review. Further prospective multi-centre validation is required to determine whether this approach improves procedure-level efficiency or clinical outcomes. FUNDING This work was supported by grants from Shanghai Science and Technology Innovation Action Plan (24Y12800702), National Natural Science Foundation of China (82301794), Natural Science Foundation of Shanghai (25ZR1401300), National Key Research and Development Program of China (2022YFC270300), China Jiliang University Research Grant (No. H251120), and Shanghai General Hospital Basic and Clinical Collaborative Research Program (JC202612). DISCLOSURES The authors declare no competing interests.