AI facilitated sperm detection in azoospermic samples for use in ICSI

Abstract Research question Can artificial intelligence (AI) improve efficiency and efficacy of sperm searches in azoospermic samples? Design This two-phase proof-of-concept study beginning with a training phase using 8 azoospermic patients (>10000 sperm images) to provide a variety of surgically collected samples for sperm morphology and debris variation to train a convolutional neural network to identify sperm. Secondly, side-by-side testing on 2 cohorts, an embryologist versus the AI identifying all sperm in still images (cohort 1, N=4, 2660 sperm) and then a side-by-side test with deployment of the AI model on an ICSI microscope and the embryologist performing a search with and without the aid of the AI (cohort 2, N=4, >1300 sperm). Time taken, accuracy and precision of sperm identification was measured. Results In cohort 1, the AI model showed improvement in time-taken to identify all sperm per field of view (0.019±0.30 x 10 -5 s versus 36.10±1.18s, P<0.0001) and improved accuracy (91.95±0.81% vs 86.52±1.34%, P<0.001) compared to an embryologist. From a total of 688 sperm in all samples combined, 560 were found by an embryologist and 611 were found by the AI in <1000 th of the time. In cohort 2, the AI-aided embryologist took significantly less time per droplet (98.90±3.19s vs 168.7±7.84s, P<0.0001) and found 1396 sperm, while 1274 were found without AI, although no significant difference was observed. Conclusions AI-powered image analysis has the potential for seamless integration into laboratory workflows, and to reduce time to identify and isolate sperm from surgical sperm samples from hours to minutes, thus increasing success rates from these treatments.