Original AI-assisted conceptual illustration: symbolic cell detection and trajectories, not an experimental micrograph or patient result.
An F1 score of 0.98 is not a 98% chance of having a baby. A new study in Reproduction describes an open-source system that detects sperm heads in microscope images and reconstructs their movement. It is an interesting development in measurement technology. It is not yet evidence that a new fertility test improves treatment decisions, pregnancy rates or live births.
For couples undergoing fertility assessment, that distinction matters. A precise-looking number can sound reassuring, particularly when artificial intelligence is involved. But before asking how impressive a number is, we need to ask what was measured, in whose samples, and against which reference.
What did the researchers build?
The workflow combines two different tasks. First, a deep-learning tool based on Cellpose identifies sperm heads in individual microscopy frames. Next, tracking software, Trackpy or TrackMate, links detections over time to reconstruct individual trajectories. Movement parameters can then be calculated from those paths.
The researchers used species-specific models for human, mouse, rat and bovine sperm. Human sperm were also examined under baseline conditions and conditions promoting hyperactivation, a change in movement relevant to sperm physiology. This is a laboratory-methods study, not a randomized trial of fertility treatment.
Its appeal is transparency and flexibility. Instead of viewing sperm movement only through population averages, researchers can examine the behaviour of individual cells. Averages can conceal substantial differences: two samples may have similar mean speeds while containing cells with very different patterns of movement.
Detection, tracking and fertility prediction are different tasks
Imagine filming people moving across a railway platform. Identifying each person’s position in one frame is detection. Following the same person through successive frames is tracking. Predicting whether that person reaches a particular destination is a different task again. Success at the first does not automatically establish success at the second or third.
The same logic applies here. Detecting a sperm head correctly is essential, but trajectories can still be interrupted, confused when cells cross, or affected by cells entering and leaving the field. Even a perfectly reconstructed trajectory would remain a measurement of movement—not a complete assessment of fertilizing ability or a prediction of parenthood.
The reported F1 scores of 0.96–0.98 refer to sperm-head detection on held-out image datasets. They should not be described as the percentage of pregnancies predicted correctly, the proportion of cells tracked without error for an entire video, or the probability of a live birth.
What does F1 measure?
F1 combines precision and recall. Precision asks how many objects labelled as sperm heads really were sperm heads. Recall asks how many of the annotated sperm heads the system found. The score balances these two aspects of detection performance.
This is useful information about a defined technical task. It is not a substitute for a clinical endpoint. A high F1 score also does not tell us whether performance remains stable on another microscope, in samples with substantial debris, or in patients whose samples differ from those used during development.
The human detection-validation datasets included images from eight negative phase-contrast videos, containing 307 annotated sperm cells, and six positive phase-contrast videos, containing 182 cells. These are counts of videos and cells—not hundreds of independent patients. Confusing these denominators would exaggerate the clinical evidence.
The human samples were selected
In the study, motile human sperm were selected using a swim-up procedure. In simple terms, this enriches the sample for cells able to move into the surrounding medium. It is a legitimate laboratory preparation, but it means the analysed material should not be treated as interchangeable with every unprocessed semen sample seen in a fertility clinic.
Routine practice includes low sperm concentrations, poor motility, debris and other challenging conditions. Performance in selected motile cells cannot establish reliable performance across that whole spectrum. Before clinical adoption, validation would need to include representative samples and explicitly report where the method fails.
Results across four species demonstrate methodological breadth. They do not replace human clinical validation. A system working in mouse or bovine sperm has not thereby been shown to improve care for a couple undergoing IVF.
Were the test data genuinely independent?
The paper reports held-out datasets that were not used for model training. That is an important safeguard. However, independence has several levels: different images, different videos, different donors and different clinical centres are not equivalent.
The available report does not establish donor-level separation clearly enough to assume it. This is not proof that data leakage occurred. It is a reason not to claim a stronger form of independence than has been demonstrated. Closely related material from the same donor or imaging session may share characteristics that make testing easier than deployment elsewhere.
A clinically persuasive next step would use a locked model tested on new donors at an independent centre, with predefined criteria and prospective data collection. Performance should be reported for important sample subgroups and different imaging conditions, including difficult cases rather than only technically clean examples.
Does it agree with existing CASA systems?
Computer-assisted sperm analysis, or CASA, is an established approach to analysing movement. The researchers compared several descriptors from their workflow with a commercial CASA system, including straight-line velocity, curvilinear velocity and lateral head displacement.
There is an important qualification. The comparison was restricted to motile trajectories with straight-line velocity above 10 micrometres per second. The paper reports values in similar ranges, but explicitly states that a formal analysis of measurement agreement was not performed.
Similar averages are not enough to prove interchangeability. Two methods could give the same average across a group while disagreeing substantially for individual samples. If a laboratory changes equipment or software, the question is not simply whether the new output looks plausible. It is whether differences are small enough for the intended clinical use and whether decisions would remain dependable.
What benefit was shown for patients?
The study did not test whether using this workflow changed treatment decisions or improved pregnancy or live-birth outcomes. It did not establish that selecting treatment on the basis of its measurements is better than the current pathway.
That does not make the work unimportant. Better research tools can reveal aspects of cell behaviour that existing summaries miss. They may help generate hypotheses and eventually support better diagnostic methods. The appropriate conclusion is that this workflow merits further evaluation—not that it is already a validated predictor of fertility.
For a patient, “we can measure more detail” and “this additional detail helps you” are separate claims. The second requires its own evidence. A new test should have a clearly defined role, and the team offering it should be able to explain what would change depending on the result.
What should a laboratory check before implementation?
The following checklist is a proposed implementation standard, not a list of benefits demonstrated by this study:
- Intended use: is the system for exploratory research, routine measurement, or a treatment decision? Each purpose needs an appropriate validation plan.
- External testing: are new donors, another centre and realistic sample difficulties represented?
- Error handling: can staff recognize unreliable detections or trajectories, and is there a defined fallback?
- Method comparison: are individual-sample differences assessed, rather than only group averages?
- Clinical impact: does the additional information improve decisions or outcomes, rather than merely increase the number of reported parameters?
- Governance: who controls access, software versions, quality assurance, maintenance and incident review?
Open-source software does not mean implementation is free. Staff training, computing, storage, maintenance and quality control all consume resources. The authors themselves note the storage and processing demands of video and trajectory datasets. Local privacy and cybersecurity arrangements also require assessment; a technical performance score does not certify them.
Questions worth asking during a consultation
If a clinic offers an AI-enhanced semen assessment, ask which part of the process uses AI. Does it detect cells, measure movement, classify samples, or predict an outcome? Ask whether the validation population resembles your situation and whether the result will change management.
You can also ask what happens when the algorithm is uncertain, whether a qualified professional reviews the output, and what evidence supports any additional fee. These are reasonable questions about any new diagnostic technology, not an argument against innovation.
No decision to start, stop or change fertility treatment should be based on the headline F1 score from this paper. Individual assessment and interpretation remain essential.
The useful promise—and the boundary
This study offers a flexible way to investigate sperm movement at single-cell resolution. Its strongest contribution is a transparent measurement workflow, with encouraging detection performance in the tested datasets.
The boundary is equally clear: detection accuracy is not fertility prediction. External validation, formal method comparison and evidence of decision or patient benefit are still needed before stronger clinical claims can be made. Responsible communication preserves the value of the innovation by being precise about what it has—and has not—shown.
Source and transparency
Ameijeiras P, Kaplan E and colleagues. Automated Sperm Tracking Using Deep Learning for High-Resolution Analysis of Motility. Reproduction, accepted manuscript, 2026. Read the source article, DOI: 10.1093/reprod/xaag114.
The authors report support from the Walter Scott Family Foundation, the US National Science Foundation, the Chan Zuckerberg Initiative and The Company of Biologists, and declare no conflicts of interest. These declarations provide context; they do not replace appraisal of the methods.
This article is educational and does not replace an individual medical consultation. Evidence reviewed on 9 September 2026.