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