SpermSearch.AI clinical study at IVF Australia published
Our clinical study is out in Reproductive BioMedicine Online
The clinical study of SpermSearch.AI spearheaded by our CSO, Dale Goss, has been accepted in Reproductive BioMedicine Online (RBMO), and the journal pre-proof is now online. The full paper is free to read: Real-time artificial intelligence-assisted rare-sperm searching: a prospective, multi-centre, paired sister-oocyte pilot study
The study set out to answer a fairly narrow question. In severe male factor cases, does running real-time AI detection alongside an embryologist during a rare sperm search change how long the search takes and how much sperm is recovered? And in the patients where both methods produced sperm, does the choice of search method indicate improvements in outcomes?
It also reports what we believe to be the first live births following real-time AI-assisted sperm searching from January 2025 and January 2026.
Why rare sperm searching is a bottleneck
Between 10 and 20% of men investigated for male factor infertility are azoospermic, meaning no sperm at all in the ejaculate, and around 60% of those have non-obstructive azoospermia (testicular or pretesticular azoospermia). For most of these patients, sperm retrieval means microdissection testicular sperm extraction, and whatever the surgeon recovers then has to be found by an embryologist working through the tissue at a microscope. In many cases an extended search of a semen sample comes first, to rule out rare ejaculated sperm before surgery is considered.
The search is where a large proportion of the laboratory time gets spent and sperm are extremely sparse and sit among a large volume of background cells and debris. This means the work is slow and depends heavily on both the individual searching and how many staff the laboratory can dedicate to this task. In many cases, this search is performed in parallel with egg collection or egg warming, which turns it into a rate-limiting step in an already tight schedule.
The cost is not only scheduling and logistics, the longer search and handling times have been associated with reduced fertilisation rates in testicular tissue cases. Extended handling outside the body also gives sperm more time to accumulate DNA damage, which mature sperm are unable to repair. Prolonged microscope work carries a documented occupational health burden for embryologists on top of that.
What was missing from the evidence
AI sperm detection is not new. Deep learning models have been shown to identify rare sperm accurately in testicular tissue and in concentrated semen, including our own detection work published in RBMO in 2024. That body of work established that the models can do the job under controlled conditions when trained on representitive datasets.
What it had not established was what happens inside a live clinical workflow. Performance on stored images tells you the model works, but not what a search looks like when an embryologist is using it during a laboratory workflow in clinical conditions. No prior study had compared assisted and unassisted searching of the same sample, under identical conditions, in routine practice.
How the study was run
The study ran at the IVF Australia Alexandria and Greenwich laboratories in Sydney, under ethics approval from the IVF Australia Human Research Ethics Committee under informed consent.
Forty patients were included, all with poor prognosis non-obstructive azoospermia or severe oligozoospermia. Thirty-five were surgical cases and five were extended semen searches. Across those patients, 147 dishes were searched by 15 different scientists.
The design is what makes the comparison work. For each patient, matched dishes were prepared in an identical format, and the sample was loaded into every dish before any dish was allocated to a search method. One set was then searched the usual way through the eyepiece, and the other with SpermSearch.AI running. Because both halves came from the same patient, the same preparation and the same dish format, each patient effectively acts as their own control.
Searches were run at the same time as egg collection or egg warming, exactly as they would be on any other day.
What the tool does, and what it does not do
SpermSearch.AI is a real-time detection model trained on clinical material, so that it sees the debris, cell density and variability that a severe male factor search actually involves.
In the lab it runs alongside the microscope, drawing boxes around candidate sperm on a monitor at eye level while the embryologist pans across the sample as they normally would. The system localises candidates for the embryologist to confirm the candidate, assess sperm, and select for injection into eggs. This leaves the embryologist’s expert judgement in place while accelerating the speed at which sperm are identified.
SpermSearch.AI live detection during a cryptozoospermic semen search
Search time and sperm recovery
Average time per dish fell from just over 40 minutes without assistance to just under 24 minutes with it, a reduction of 41%. In the patients where sperm were found by both methods, average time to find each sperm fell from around 27 minutes to under 13 minutes, a reduction of 53%.
Sperm recovery also improved with the use of AI, where the average number of sperm recovered per dish rose from 2.02 to 3.71, an increase of roughly 1.8 times. Every sperm found in these cases is another chance at injecting an egg, and thus creating an embryo.
Cycles that went ahead
It is important to note that 20/40 cases went ahead with oocyte injection in the study. But the number we keep coming back to is 8 out of 40.
In four patients, sperm were found with AI assistance and not found at all without it. Two of those four yielded only immotile, poor quality sperm and the cycle could not proceed. In a further six patients, sperm were found both ways, but the only motile sperm suitable for injection came from the AI-assisted dishes. Following each clinic's standard protocol of injecting the best available sperm, those were the sperm that were used.
Altogether, eight ICSI cycles went ahead on sperm that had been found only with AI assistance.
What happened at fertilisation
Twelve patients had eggs injected using sperm from both arms, which allowed a direct like-for-like comparison within the same couple. Across that group, 31 of the 73 eggs injected with AI-found sperm fertilised, compared with 9 of 38 in the unassisted arm. Seventeen usable embryos came from the AI-assisted sperm and four from the unassisted sperm.
These numbers should be treated as early and exploratory. Twelve patients is a small group, several produced no usable embryos in either arm, and results from a sample that size can shift a long way on one or two cases. Everything we looked at pointed in the same direction, but the study was not built to prove fertilisation or embryo development as these would require larger numbers.
Pregnancies and live births
Across the full group of 40 patients, 13 of them (32.5%) produced usable embryos only from eggs injected with sperm found during AI-assisted searching. There were no cases where the reverse happened.
Four clinical pregnancies and two live births came from embryos created with AI-assisted sperm.
To our knowledge these are the first live births reported following real-time AI-assisted sperm searching, where the embryologist performs the search and the system assists as they go. They were not endpoints of the study and are reported descriptively rather than as a result the study was designed to produce.
What about operator variability?
The most obvious challenge to a design like this is that the two searches ran in parallel, performed by different people. If the AI-assisted searches happened to be done by faster or more experienced embryologists, that alone could explain the result.
We describe in the paper the influence of operator variability and found that the embryologists searching without assistance worked at broadly similar speeds to one another, so baseline skill across the team was reasonably even. When we re-ran the analysis with the individual embryologist accounted for, the time saving did not significantly move. Far more of the variation in how long a dish took came down to the patient's case than to who was at the microscope.
The sperm recovery finding held up less firmly. Once the individual embryologist was accounted for, the advantage shrank and became less certain, although it still pointed in the same direction, and the same embryologist recovered more sperm when using the tool than when searching without it. We have reported that as it stands rather than leaning on the raw comparison.
Read the paper
If you’re based in Australia and want to integrate SpermSearch.AI into your clinical workflow for severe male factor cases, contact us to arrange a demo.
Acknowledgements
This study was only possible because of the embryology teams at the IVF Australia Alexandria and Greenwich laboratories. Our thanks in particular to Karina Evans, Elizabeth Gardiner, Rebecca McMahon, Cassandra Nader and Charlotte Knight, to the managing doctors involved in patient care throughout, and to Virtus Health for supporting and enabling the study.
