O-136 Artificial intelligence to assist in surgical sperm detection and isolation

Abstract Study question Can an artificial intelligence (AI) improve the speed and accuracy of identifying sperm in complex testicular tissue samples? Summary answer Trained AI can identify real-time instantly with higher accuracy, not only reducing strain on embryologists but increasing sample coverage a shorter time. What is known already Non-obstructive azoospermia (NOA) form severe male-factor infertility, affecting nearly 5% infertile couples seeking treatment. Isolating from macerated for intracytoplasmic injection (ICSI) has changed marginally last two decades, still requires to search arduously through background collateral cells including red blood (RBC’s), white (WBC’s), leydig, sertoli epithelial cells, causing fatigue coverage. Image analysis using trained shapes forms thus presents itself as candidate dramatically reduce processing times surgical cases. design, size, duration This proof-of-concept consists phases over 5 months. A training phase 7 azoospermic patients provide samples train convolutional neural network manual annotation. Secondly, side-by-side live test embryologist versus model comparing time taken identification, precision (false positives false negatives). Results were analysed Mann-Whitney U-test. Differences considered significant when p-value<0.05. Participants/materials, setting, methods Consenting at IVF Australia, Sydney attending collection donated discarded whereby excess was used post-processing conjunction donor sperm, machine learning software. Thereafter, testing conventional ICSI micromanipulator microscope performed taken, recorded between control count, computer vision embryologist. Main results role chance The showed dramatic improvement per field, improved well high level precision. Precision be 100%. Time field view all significantly lower compared (0.019 ± 1.4 x 10-4s vs 22.87 0.98s, P < 0.0001). There also difference (89.88 1.56% 83.22 2.02%, = 0.017) respectively. 91.27 1.27% which correct identification count. From total 688 found, 560 found by 611 less than 1000th it took Limitations, reasons caution AI. Following this study, clinical are needed, trial required prove efficacy usefulness. larger number across multiple indications approaches required. Wider implications findings AI-powered image potential seamless integration into laboratory workflows isolate hours minutes, do success rates these treatments. Trial registration N/A