Counting the oocytes in an ovary has usually meant serial sectioning and extrapolation. By imaging the whole organ and segmenting every oocyte with deep learning, a team at the Centre for Genomic Regulation in Barcelona has drawn up an inventory of the ovarian reserve across the entire reproductive lifespan of the mouse — and found an unexpected invariance: the share of the stock that sits, at any given moment, in the activation window does not move, while the stock itself collapses.

Source: biorxiv.org

In plain terms

A woman is born with a stock of eggs she will never renew, and that stock dwindles with age. It is tempting to picture this reservoir as a draining bathtub: the more water remains, the stronger the flow. This study, carried out in mice, observes something else. At any given moment, the same proportion of the remaining stock — about 14% — is in the transition phase where an oocyte leaves dormancy. That share does not change even as the stock is divided by more than ten. In other words, what appears to be regulated is a proportion, not a number. Three caveats up front: all of this is measured in mice; the study observes a regularity without experimentally testing the mechanism that would produce it; and there is neither a treatment nor a clinical promise here.

Technical explanation

1. Why the ovary resisted counting — The mouse ovary is several millimetres across and irregularly shaped. The historical method consists of cutting it into series of histological sections, counting follicles on a fraction of the sections, then multiplying by a correction factor. The authors note that these section-based methods introduce inaccuracies in estimating primordial oocyte numbers and sacrifice all spatial information. One may add, as a methodological reminder, that two sources of error classically arise: an oocyte crossed by two sections may be counted twice or not at all, and extrapolation assumes a homogeneous distribution — precisely the assumption this study undermines. Light-sheet microscopy sidesteps the obstacle by illuminating the specimen with a thin plane perpendicular to the observation axis, which allows optical sections to be acquired without touching the tissue. The light still has to get through: clearing replaces the tissue medium with a mixture whose refractive index is close to that of proteins, here benzyl alcohol/benzyl benzoate, which suppresses scattering at interfaces and renders the organ translucent.

2. The labels — The anti-DDX4 antibody serves as an oocyte marker; the TO-PRO-3 dye counterstains DNA and provides tissue context, which makes it possible to situate oocytes among the other nuclei. The authors indicate that their optimised protocol also allows the stages of folliculogenesis to be distinguished — on the criterion of size, as point 4 sets out. They claim identification of all oocytes in situ following the joint optimisation of the immunostaining and clearing protocols — and it is on that optimisation, not on a cross-check between the two channels, that they base this claim.

3. Segmentation, and what its agreement rate actually bounds — A whole ovary contains thousands of adjoining oocytes; separating them one by one is a matter of instance segmentation, harder than mere detection. The authors manually annotated volumes, trained a network within the BiaPy framework, then imaged five independent ovaries to compare the automatic output against expert annotation: 95.3% agreement before correction, 96.7% after light correction. A qualification is needed on the scope of that figure: this validation was carried out on 5-week-old ovaries, that is, in the densest regime. Nothing establishes that it holds at 60 weeks, where local density has fallen by a factor of four. An age-dependent drift in performance therefore remains possible, and it would affect the measurement of ageing itself.

4. How an oocyte is classified — a diameter threshold, calibrated on morphology — This is the most important point to grasp in order to read the main result, and it is easy to state wrongly. It is in confocal microscopy, and not on the SPIM volumes, that the authors go looking for granulosa cell morphology; the main pipeline itself measures oocyte diameters. They therefore calibrated a threshold on this confocal-imaged subset, where the shift from squamous (flattened) to cuboidal cells — which marks entry into growth — is directly observable. Taking the mean ± the standard deviation of the corresponding diameters, they obtain the bounds 16.30 µm and 20.11 µm, and then classify the entire dataset by size alone: primordial from 10 to 16.30 µm, transitioning from 16.30 to 20.11 µm, growing beyond 20.11 µm. The "transitioning" category, which carries the central result, is therefore a narrow diameter window, grounded in a morphological observation but applied metrically.

5. The central result: a share, not a number — Between 5 and 60 weeks, the total reserve falls by more than an order of magnitude. Over the same period, the composition of the pool shifts markedly: primordial oocytes go from about 70% to 55% of the total, growing oocytes make the reverse journey, from about 15% to 30%. It is this contrast that makes the third figure remarkable — the share of transitioning oocytes stays close to 14% at every age. Writing $R(t)$R(t) for the reserve and $T(t)$T(t) for the transitioning count, the measurement gives $T(t)/R(t) \approx 0{,}14$T(t)/R(t)≈0,14, whereas recruitment at a constant absolute rate would impose $T(t) \approx \text{const}$T(t)≈const, and hence a ratio that would explode as $R$R decreases. The data settle the two regimes in favour of proportional recruitment. Take care not to confuse the two quantities: 14% is a share of the population at a given instant, not a speed. The recruitment rate itself is estimated at about 2.5% of the dormant reserve per week (point 6).

6. The kinetic model — Progression is described as a one-way cascade: an oocyte that has left dormancy cannot return to it; at each step it advances to the next compartment at rate $k_i$ki​ or dies at rate $d_i$di​. The preprint does not write the equations explicitly; the system can be put in the form
   $\frac{dP_1}{dt} = -(k_1 + d_1)\,P_1 \qquad \frac{dP_i}{dt} = k_{i-1}P_{i-1} - (k_i + d_i)\,P_i \ \ (i = 2,3) \qquad \frac{dP_4}{dt} = k_3 P_3 - d_4 P_4$dtdP1​​=−(k1​+d1​)P1​dtdPi​​=ki−1​Pi−1​−(ki​+di​)Pi​  (i=2,3)dtdP4​​=k3​P3​−d4​P4​
   For this fit — and for it alone — the size distribution is re-partitioned into four non-overlapping classes: primordial (10–20 µm), primary (> 20–40 µm), secondary (> 40–60 µm), antral (> 60 µm). This four-compartment partition is distinct from the three-stage scheme of point 4: the "transitioning" category does not appear in it. Mind the vocabulary, which changes meaning from one scheme to the other — the model's "primordial" compartment (10–20 µm) encompasses the transition window of point 4 (16.30–20.11 µm). The 2.5% per week cited above is therefore an exit rate from this enlarged compartment, and not from the strictly dormant pool alone. The system is integrated from the counts measured at 5 weeks and fitted by constrained least squares (SciPy's SLSQP algorithm). The rates obtained are $k_1 = 0{,}0258 \pm 0{,}0053$k1​=0,0258±0,0053, $k_2 = 0{,}2173 \pm 0{,}0400$k2​=0,2173±0,0400, $k_3 = 0{,}3375 \pm 0{,}0567$k3​=0,3375±0,0567 week⁻¹; on the loss side, $d_1 = 0{,}0179 \pm 0{,}0074$d1​=0,0179±0,0074 and $d_4 = 0{,}7224 \pm 0{,}1470$d4​=0,7224±0,1470 week⁻¹ — the latter being an exit rate from the antral stage that aggregates degeneration and ovulation.

   One point of reading deserves to be set down, because it changes the status of two parameters. The authors state in their methods that the constrained minimisation led to zero death rates at the primary and secondary stages, and that they then fixed $d_2$d2​ and $d_3$d3​ to zero for the final fit, on the grounds that the literature points that way. The "$\pm\, 0{,}0000$±0,0000" displayed is therefore not a measurement precision: it is the signature of frozen parameters, and the other rates were fitted in the knowledge of that choice. The authors see in it a confirmation of the robustness of their assumptions; one may just as well see an assumption injected into the model, which the model can no longer be asked to test.

7. The 60 µm bottleneck — Once activated, an oocyte increases its volume roughly 64-fold — a figure the authors take from the literature, not measured here. If activation and growth were both regular, the size distribution would form a smooth continuum. It is in fact bimodal: a dominant, sharp peak near 15 µm corresponding to the dormant reservoir, then a second peak around 60 µm. A peak in a size distribution signals a slowdown — cells accumulate where they progress most slowly, just as a queue forms at the slowest point of a chain. And 60 µm corresponds to the late secondary/early antral transition, that is, the moment when the follicle becomes responsive to gonadotropins. The interpretation put forward is therefore that a change in growth regime occurs upon acquisition of this hormonal sensitivity. This peak persists at every age, which indicates that the phenomenon does not depend on the size of the remaining stock.

8. The geometry: where the oocytes sit — A spline curve is manually positioned along the organ and serves as an internal ruler: each oocyte is located by its position and its angle around that axis, which makes ovaries comparable between animals and between ages. The frame is therefore angular, not longitudinal: 0° designates the superior face, 180° the hilum, and the sectors 45–135° and 225–315° the medio-lateral regions. The result is clear-cut: primordial oocytes are depleted at the hilum (concave face) as well as at the opposite convex pole, and enriched in the medio-lateral regions. The same bias holds for transitioning and growing oocytes, and it is maintained with age. The manual placement of the spline introduces an element of subjectivity that should be kept in mind, since the entire angular frame depends on it.

1. Local density and activation — A separate analysis from the previous one: for each primordial oocyte, the authors count its primordial neighbours within a 100 µm radius. This local density falls with age — more than 10 neighbours on average at 5 weeks, about 2.5 at 60 weeks. The interesting point is the direction of the correlation. A long-standing hypothesis holds that a dense neighbourhood inhibits activation, dormant follicles mutually repressing one another. The data go the other way: regions of high primordial density harbour more transitioning oocytes, and the reciprocal relation is positive as well. This result must be read for what it is — a spatial correlation, measured without perturbation. It undermines the neighbourhood-inhibition model, proposed, it is true, for the prepubertal ovary, whereas these data begin at 5 weeks; but it does not demonstrate that high density causes activation. A third variable, local vascularisation for instance, could carry both effects.

2. The pregnancy test — Females were mated from 7 weeks onwards, over 18 weeks, carrying several pregnancies to term; their reserve was compared at 31 weeks with that of virgins of the same age. No significant difference appears, either in the total stock or in any of the subcategories. The authors infer that suppressing ovulation does not appear to slow oocyte loss, and that follicular recruitment would be regulated at tissue level, independently of ovulatory activity. Two qualifications for the reader: this is a null result measured at a single point, which does not amount to a demonstration; and the intervention does not isolate ovulation, since repeated pregnancies also alter the hormonal axis, metabolism and vascularisation.

Why It Worked

The contribution lies less in a novel technique than in the combination of three elements, each of which corrects a weakness in the other two. Volume imaging eliminates the extrapolation bias of sections; automatic segmentation makes an exhaustive count practicable across a hundred organs, a scale at which manual counting would be unfeasible; and validation against expert annotation bounds the algorithm's error, at least in the regime where it was carried out. The authors themselves place their work in a lineage: two recent studies had already combined clearing, microscopy and machine learning to automate oocyte counting. What they claim as new is the generalisation — a pipeline that holds up on the adult ovary, across the whole reproductive lifespan, and transposable to human tissue.

The most telling comparison is not before/after method, but between two competing models of ovarian depletion. Recruitment at a constant absolute rate predicts a fixed number of oocytes activated per week; proportional recruitment predicts a fixed share. The data settle it: the share stays close to 14% while the stock is divided by more than ten, and while the two other fractions of the pool shift by some fifteen points each.

There remains the gap between what is measured and what is asserted. The phrase "the ovary counts its own eggs", widely relayed, does not come from the institutional press release alone: it is carried by the authors themselves, whose discussion advances that "the ovary possesses a mechanism to monitor the total number of oocytes", and whose abstract concludes to an organ-scale regulatory mechanism. The preprint frames these sentences in the conditional, and that is where the calibration is at stake: what is measured is an invariance of proportion and a correlation between local density and activation. The existence of a mechanism that would "read" the size of the stock in order to wake a percentage of it is a hypothesis compatible with these observations, but no experiment here perturbs the size of the reserve in order to test it: no partial ablation followed by a measurement of the activation rate, no mutant, no graft. The study is descriptive and correlative, and it is on that basis that it is solid.

Three further limitations deserve to be set down. Zero mortality at the primary and secondary stages is not a measurement but a parameter fixed by the authors. The model further assumes that growth and death rates remain broadly stable over the course of ageing — an assumption the authors state explicitly and give as liable to revision, even though it conditions all the published rates. Finally, human validation amounts to a demonstration that the pipeline works on ovarian cortex from two donors aged 25: a proof of technical feasibility, in no way a study of human dynamics.

Causal Chain

Oocyte reserve fixed during fetal life → historical counting by histological sections and extrapolation, at the cost of inaccuracies on primordial oocytes and the loss of spatial information → development of light-sheet microscopy and refractive-index-matched clearing protocols → DDX4 labelling and nuclear counterstaining allowing each oocyte to be situated in intact tissue → manual annotation of 3D volumes and training of an instance segmentation model → validation at 95.3% against expert on five 5-week-old ovaries → calibration of a diameter threshold (16.30–20.11 µm) on granulosa morphology observed by confocal → individualised counting of more than 85,000 oocytes across 56 mice spread over seven ages → finding of a transitioning share invariant at ≈ 14% despite a collapse of the stock by an order of magnitude and a shift in the two other fractions → fitting of a system of differential equations giving ≈ 2.5% of the dormant reserve recruited per week → reframing of the ovarian reserve as a regulated system rather than a passive reservoir → hypothesis of proportional control at organ scale, as yet untested by perturbation → were it confirmed, a potential target for modulating the rate of depletion, on an entirely preclinical horizon.

Anecdote

The most unsettling detail of the study is not the main result but the observation that preceded it. It is the known variability of anti-Müllerian hormone — the reserve marker used in human clinical practice, whose values scatter widely from one woman to another — that led the authors to ask whether the number of oocytes itself varied as much between individuals, including within genetically identical populations. They therefore imaged eight five-week-old C57BL/6J females, inbred animals, hence genetically identical. The answer is yes, and the gap is considerable: up to three times more oocytes in one than in another. The gap is already measurable at five weeks, which suggests that it sets in before sexual maturity. One may infer, without the authors writing it, that part of the initial endowment escapes both genetics and reproductive history.

Legacy and Current Data

The human orders of magnitude recalled by the authors frame what is at stake: about 1 to 2 million oocytes at birth, roughly 400,000 at puberty, fewer than 1,000 at menopause — for only some 400 ovulations over a lifetime. The overwhelming majority of the stock is therefore lost without ever being ovulated, and it is the pace of that loss, not the number of ovulations, that sets the length of reproductive life.

The quantitative framework proposed here is presented by the authors as the first numerical estimates of follicular activation and growth rates in a mammal. What those rates become in women remains entirely to be established: no data in the study allow them to be transposed.

Two notes of transparency. The preprint carries an internal inconsistency on data volume: its abstract announces "more than 85,000 oocytes", its discussion "more than 100,000". Both figures appear in the same document without being reconciled; the present article retains 85,000, the abstract's value, also taken up by the institutional press release. Furthermore, the exact reference of the version published in Nature Aging could not be verified directly: all the figures cited here come from the bioRxiv preprint, and not from the journal version. Finally, the study is cross-sectional — seven distinct ages, on different animals — and not longitudinal: current technologies do not allow the same ovary to be followed over time. The starting point at 5 weeks was chosen as the maximum reserve after the first wave of activation and oocyte death, and not at the very beginning of reproductive life.

The researcher's view — open questions

The experiments below do not appear in the study; they indicate what would have to be done to turn the observed correlation into a demonstrated mechanism.

• Partial ablation, the decisive test of proportional control. Remove a known fraction of the reserve, then measure the transitioning share in the remaining tissue. If regulation is indeed proportional, that share must return to 14% after readjustment; if it is absolute, the activated count must remain unchanged. This is the experiment that settles the two models, and unilateral ovariectomy provides a simple version of it.
• Decoupling density from position. Nothing in the study separates the density–activation correlation from the medio-lateral enrichment, and the two analyses are never crossed. A position-controlled analysis would say which of the two variables carries the effect.
• Testing the 60 µm bottleneck through hormonal sensitivity. If the size peak does indeed reflect the acquisition of gonadotropin responsiveness, manipulating FSH signalling should shift that peak in size, or abolish it. An unresponsive peak would invalidate the interpretation.
• Lifting the model's stationarity assumption. The rates were fitted assuming they do not drift with age. Fitting the model by age brackets, and comparing the $k_i$ki​ obtained, would say whether that assumption holds — it is the most accessible internal control.
• Taking the pipeline to a human series. Feasibility is established on two donors. A cohort spanning several decades would say whether the invariance of proportion is specific to the mouse or general among mammals: that is the question which decides the clinical reach of the whole edifice.

Sources

• D'Angelo A., Franco-Barranco D., Musy M., Sharpe J., Arganda-Carreras I., Böke E., "3D Mapping of Intact Ovaries Reveals the Aging Dynamics of the Ovarian Reserve", bioRxiv, posted 8 November 2025 — doi:10.1101/2025.11.07.686728. Status: preprint, not peer-reviewed at the date of posting; peer-reviewed version announced in Nature Aging.
• Centre for Genomic Regulation (CRG), "First 3D map of ovary throughout reproductive lifespan reveals mouse reproductive organ counts its own eggs", press release of 12 August 2026 — crg.eu, also distributed by EurekAlert!. Status: institutional press release, announcing publication in Nature Aging.

Background references

These references do not report the results of the study discussed; all three appear in its bibliography and support the general mechanistic reminders made above.

• Huisken J., Swoger J., Del Bene F., Wittbrodt J., Stelzer E. H. K., "Optical sectioning deep inside live embryos by selective plane illumination microscopy", Science, 2004, vol. 305, pp. 1007–1009 — doi:10.1126/science.1100035. Principle of light-sheet microscopy.
• Edson M. A., Nagaraja A. K., Matzuk M. M., "The mammalian ovary from genesis to revelation", Endocrine Reviews, 2009, vol. 30, no. 6, pp. 624–712 — doi:10.1210/er.2009-0012. Folliculogenesis and the acquisition of gonadotropin sensitivity at the late secondary/early antral stage.
• Winship A. L., Sarma U. C., Alesi L. R., Hutt K. J., "Accurate Follicle Enumeration in Adult Mouse Ovaries", Journal of Visualized Experiments, 2020 — doi:10.3791/61782. Estimation biases in primordial follicle numbers from serial-section methods.