P–123 How to develop accurate Computer Assisted Sperm Analysis (CASA) AI in the absence of protocol standardization and abundance of human error when performing semen analyses?

Study question How can an automation & artificial intelligent tools be developed to perform according to WHO recommendations? Summary answer Developing CASA performs at < 20% error margin requires AI trained with high quality datasets and a robotic system adheres to WHO guidelines. What is known already A survey of 40 andrology laboratories, in 22 countries, revealed that > 90% had nonconformities in correct use of equipment, standardisation of protocols and quality control, leading to a lack of compliance to WHO protocols. Conventional CASA systems can standardize analysis, but controversy has occurred due to differences between manual and automated analyses stemming from: 1) all cells in a semen sample are detected including debris; 2) protocol variation when compared to top-notch manual analysis. The first point can be addressed by AI. The second point can be addressed by robotics designed to adhere to WHO guidelines. Study design, size, duration A mojo AISA (AI-powered semen analysis) system was placed in four clinical laboratories mentioned above capturing images of over 300 samples, one million images were generated over a course of 2 years. Mojo AISA’s AI was trained on data collected from the four clinics using robotic system is developed according to WHO guidelines. Participants/materials, setting, methods For an AI to detect sperm accurately, sperm samples were captured using mojo AISA smart microscopy and then the extracted sperm images expertly annotated. To evaluate the system-ability for semen analysis, fresh sample were analysed for concentration and motility by a manual operator and compared to a mojo AISA test. Main results and the role of chance To train the sperm detection AI, representative sperm images were carefully captured using mojo AISA and processed according to the following criteria: the number of images and videos to train and to test the model: 50,000 spermatozoon head and tails with various variations the variety of images: data used to train the AI has to be representative of the population that will undergo the analysis: 1) wide concentration ranges from 0 to 300 M/ml, 2) high and low density of debris and cells, 3) Presence of slight aggregations careful and precise annotation: expert andrology scientists annotated sperm images and identify objects to exclude, such as debris in seminal plasma, Mojo AISA is an attempt strictly build CASA AI system to WHO-guidelines. The marriage of AI and robotics automation has shown a promising results to mimic humans when measuring a semen sample and attempt to obtain results comparable to the manual analysis. mojo AISA’s performance improved three-fold (from 0,85 to 0,95 Pearson sperm count correlation and from >100% means relative error to 25% mean relative error). Limitations, reasons for caution Lack of standardization for semen analysis laboratory process globally is a bottleneck towards building a robust multi-center study, on-site CASA testing and generating an actionable data pool for studying the causes behind male fertility declineWider implications of the findings: Key learnings for parties advancing developing AI based on images and videos for application in the fertility space. Trial registration number Not applicable