Undisturbed morphokinetic imaging & AI-assisted embryo ranking inside a stable-gas benchtop incubator
Time-lapse incubation systems (EmbryoScope, Geri, Miri, Primo Vision and similar platforms) combine a standard benchtop incubator with an integrated microscope and camera positioned beneath each embryo. Instead of removing embryos daily for manual assessment under a conventional microscope, each embryo is loaded once into its own culture well and then never touched again until transfer or freezing.
Conventional IVF workflow requires removing embryos from the incubator daily (sometimes twice daily) to assess development under a standard inverted microscope. Each removal exposes the embryo to room temperature, atmospheric O₂ (~21% vs the intrauterine ~2–8%), pH shifts as CO₂ off-gasses from the culture medium, and light exposure.
Studies tracking incubator microenvironment show that a single door opening on a large box incubator can take 20–60 minutes to fully recover stable temperature, humidity, and gas concentration. A standard 5-day culture with twice-daily checks may subject an embryo to 8–10 such perturbations before transfer.
Benchtop time-lapse incubators solve this by building the optics directly into the culture chamber: a low-intensity LED or laser diode illuminates each embryo from below, and a camera captures the image through the same sealed chamber the embryo never leaves.
A 2018 Cochrane-style review across multiple RCTs found no evidence that time-lapse imaging alone improves live-birth rate over undisturbed culture with once-daily assessment — the primary proven benefit is culture stability, not the imaging itself. Morphokinetic annotation and AI ranking are what extract additional clinical value from the video.
Each embryo is cultured in its own optically flat micro-well (typically 300–400 μm diameter, machined into a polystyrene EmbryoSlide or equivalent dish) rather than a shared droplet. This design serves two purposes:
• Optical: a flat, fixed well floor keeps the embryo at a known, constant focal distance from the objective lens beneath it, which is essential for repeatable multi-focal imaging • Traceability: because each well position is fixed and labeled, the system can automatically associate every captured image and timing annotation with the correct patient and embryo — eliminating identification mix-up risk, a key patient-safety advantage of time-lapse systems
Culture medium (typically a single-step or sequential medium under mineral oil to prevent evaporation and pH drift) fills the well and is not exchanged during time-lapse culture in most single-step protocols, further minimizing disturbance.
The entire multi-well dish sits inside a small, well-insulated chamber (often only a few liters in volume, versus tens of liters for a shared box incubator), which is pre-mixed with tri-gas (N₂, O₂, CO₂) and heated to 37.0°C. Because the chamber is rarely opened and is small, gas recovery after any access is measured in minutes rather than the 20–60 minutes typical of large shared incubators.
Typical stable culture targets:
• Temperature: 37.0°C ± 0.1°C • CO₂: 6.0% ± 0.2% (buffers medium to pH ~7.2–7.4) • O₂: 5% (mimics physiologic oviductal/uterine oxygen tension, lower than ambient 21%)
This stability is the foundation the rest of the workflow depends on: any conclusions drawn from morphokinetic timing are only meaningful if the embryo experienced a consistent, undisturbed environment throughout.
While the embryo sits undisturbed in its stable-gas well, an automated microscope beneath the dish captures a still image every 10–20 minutes at 7–9 different focal planes through the embryo's depth. Over a typical 5-day culture, this produces thousands of images per embryo that are compiled into a continuous, reviewable time-lapse video.
A human embryo is a three-dimensional structure roughly 100–200 μm in diameter, and individual blastomeres (cells) can be arranged at different depths as cleavage proceeds asynchronously in three dimensions. A single 2D image at one focal depth can miss a nucleus, an extra cell hidden behind another, or a fragment obscured by overlap.
By capturing 7–9 images at evenly spaced focal depths (a "z-stack") through the embryo at every time point, the system — and the embryologist or algorithm reviewing the video — can inspect the true 3D arrangement of cells, count nuclei accurately even when blastomeres overlap, and reconstruct cell boundaries with far greater confidence than a single flat image allows.
Multiplying 7–9 focal planes by an image every 10–20 minutes across a 5-day (120-hour) culture yields on the order of 2,500–4,300 image capture events per embryo — an amount of visual data no embryologist could review frame-by-frame manually, which is precisely why compressed time-lapse video review and automated annotation software have become standard.
Repeated imaging over 5 days raises an obvious concern: could the light itself harm the embryo? Time-lapse systems address this with several safeguards:
• Red or near-infrared LED illumination (typically 615–635 nm), wavelengths embryos are minimally sensitive to compared with shorter-wavelength blue/UV light used in some fluorescence applications • Very low irradiance and short exposure times (well under 100 ms per captured frame) • Total cumulative light exposure across an entire 5-day culture (all focal planes, all time points) remains a small fraction of what a single extended manual microscopy session under a heated stage would deliver
Multiple validation studies comparing time-lapse cultured embryos to standard-incubator controls have found no significant difference in blastocyst formation rate, cell number, or clinical pregnancy outcomes attributable to the imaging light itself — the imaging is considered non-invasive to embryo development.
Proprietary software stitches the sequential focal-plane image stacks into a smooth, scrubbable video that can be played forward, reversed, sped up, or paused at any moment across the full culture period. This continuous record is the core innovation over conventional culture: instead of 2–10 discrete daily snapshots, the embryologist has an unbroken visual record of every division, every fragment, every transient event.
The video becomes the substrate for two downstream processes covered in later stages: (1) precise morphokinetic timestamp annotation of cleavage events, and (2) automated computer-vision / deep-learning analysis that extracts these same timestamps algorithmically and feeds them into a viability-ranking model.
With a complete time-lapse video in hand, embryologists (or trained algorithms) mark the exact hour at which each cleavage division occurs. These morphokinetic timestamps — t2, t3, t4, t5, t8, morula (tM), start of blastulation (tSB), and full blastocyst (tB) — form a precise developmental fingerprint that correlates with blastocyst formation and, in aggregate, with ploidy and implantation potential.
A widely used annotation nomenclature (following Meseguer et al., 2011, and subsequent consensus) defines the following key timestamps, each expressed as hours post-insemination or post-ICSI:
• t2 — first cleavage, 1-cell → 2-cell (~25–27 h) • t3 — second cleavage, 2-cell → 3-cell (~35–38 h) • t4 — 3-cell → 4-cell, near-simultaneous with t3 in normal embryos (~37–40 h) • t5 — 4-cell → 5-cell, marks onset of the third cleavage round (~48–53 h) • t8 — reaching the 8-cell stage (~55–60 h) • tM — morula, cells compact into a single mass with smoothed outer border (~85–90 h) • tSB — start of blastulation, first visible fluid cavity (blastocoel) appears (~96–100 h) • tB — full/expanded blastocyst, with distinct inner cell mass and trophectoderm (~106–112 h)
Because the video has a frame every 10–20 minutes, each of these events can be timestamped to within roughly 10–20 minutes — versus a resolution of up to 24 hours with once-daily manual checks.
The interval between t2 and t3 (cc2, "second cell cycle") and the synchrony between t3 and t4 (s2) are two of the most predictive individual morphokinetic parameters identified across multiple cohort studies, because they capture how evenly and how promptly the embryo's genome takes control of cell division after the maternal-to-embryonic transition.
A conventional Day-3 check simply counts how many cells are present at one moment (e.g., "8 cells on Day 3, grade B"). Morphokinetics instead captures the rate and rhythm of division. Two embryos can both reach 8 cells by Day 3 through very different, prognostically distinct paths:
• An embryo with tightly clustered, well-timed divisions (t2 in the normal window, short and synchronous cc2/s2 intervals) tends to correlate with higher blastocyst formation and, in several validated models, higher euploidy likelihood • An embryo with delayed or irregular divisions — reaching the same cell count late, or via uneven intervals — trends toward lower blastocyst formation and lower implantation, even though a single-timepoint snapshot would have scored it identically to the well-timed embryo
This is the central rationale for morphokinetic scoring algorithms: cell count alone under-uses the information already present in a full time-lapse video.
Manual annotation by a trained embryologist remains the gold-standard reference in validation studies, but it is time-consuming — scrubbing through thousands of frames per embryo across an entire IVF cycle's embryo cohort. Modern time-lapse platforms increasingly use convolutional neural network (CNN) models trained on large annotated datasets to automatically detect cell boundaries, count blastomeres per frame, and propose timestamps for each morphokinetic event.
Automated annotation models have reported timestamp agreement with expert embryologists within roughly 1 hour for the majority of cleavage events, and are increasingly used to pre-populate annotations for embryologist review and correction rather than fully replacing human oversight — a "human-in-the-loop" workflow typical of current clinical deployments.
Continuous video reveals a category of developmental events that are invisible to once-daily checks because they occur and resolve within a single day: direct (multipolar) cleavage where one cell divides straight into three or more, reverse cleavage where blastomeres transiently fuse back together, and multinucleation where a blastomere contains more than one nucleus. Each is associated with reduced implantation potential and, for some, elevated aneuploidy risk.
Normally, the first division (t2, 1-cell→2-cell) is followed by the second division (t3, 2-cell→3-cell) after a cell cycle of roughly 8–20 hours. "Direct cleavage" describes an abnormally fast or multipolar division where one cell splits directly into three or more cells with a t2-to-t3 interval of under 5 hours — effectively skipping the expected biphasic division pattern.
This event is invisible on a Day-2 or Day-3 spot check: an embryologist checking once daily would simply see, say, 4 cells on Day 2 and have no way to know whether they arose from two normal sequential divisions or one abnormal multipolar division. Only continuous imaging exposes the mechanism.
Direct cleavage has been associated in multiple published cohorts (e.g., Rubio et al., 2012, and subsequent studies) with a substantially lower likelihood of blastocyst formation and roughly 2–3 fold lower implantation rate compared with embryos that divide through the normal sequential pattern — making it one of the more actionable, clearly negative morphokinetic red flags.
Direct cleavage is believed to reflect chromosome mis-segregation during a compressed or chaotic mitotic event, plausibly explaining its association with mosaic or aneuploid outcomes — underscoring why continuous, undisturbed imaging captures clinically meaningful information that spot-checks structurally cannot.
Reverse cleavage describes an event where two blastomeres that have already separated appear to fuse back together into a single cell, before dividing again later. On daily static imaging this can be entirely missed — the cell count on the following day may simply look "delayed" or "normal," with no evidence the fusion ever occurred.
Reverse cleavage is thought to arise from incomplete cytokinesis (the cell membrane pinches but does not fully separate the daughter cells) or from cortical/cytoskeletal instability during division. Its clinical significance is still an active area of study, but cohort data associate it with altered developmental kinetics and, in some analyses, reduced blastocyst and implantation rates relative to embryos with no fusion events — reinforcing that it is generally treated as a cautionary, if less severe, flag than direct cleavage.
Multinucleation occurs when a single blastomere contains two or more nuclei instead of one, most commonly observed and most clinically significant at the 2-cell stage. It can arise from abnormal cytokinesis (nuclear division without complete cell division) or from chromosome mis-segregation during mitosis.
Multinucleation at the 2-cell stage in particular has been linked in multiple studies to increased chromosomal mosaicism/aneuploidy and reduced developmental competence, even though the affected embryo can sometimes appear morphologically unremarkable by later cleavage stages once the multinucleated blastomere divides again. Because multinucleation is often transient — visible for only one or two frames before resolving — it is another abnormality that is reliably captured only by continuous time-lapse video, not by spot-checks.
Taken together, direct cleavage, reverse cleavage, and multinucleation form the core "dynamic morphology" flags that distinguish time-lapse-based embryo assessment from conventional static grading, and all three are typically incorporated as negative or de-prioritizing features in downstream AI viability-ranking models.
The final step integrates every timing annotation, dynamic-event flag, and morphology grade collected across the full time-lapse video into a single algorithmic viability score for each embryo in the patient's cohort. Deep-learning ranking tools modeled on systems like KIDScore combine this data to help embryologists select the single embryo with the highest probability of implantation for transfer, while lower-ranked embryos are frozen or deprioritized.
Algorithmic ranking models are trained on large retrospective datasets pairing each embryo's time-lapse video (and its extracted morphokinetic timestamps, dynamic-event flags, and morphology grades) with its known clinical outcome — typically whether it produced a clinical pregnancy or live birth after transfer, or its ploidy status from PGT-A testing.
Using this training data, a deep-learning model (often a convolutional neural network operating directly on video frames, sometimes combined with a gradient-boosted or logistic model over extracted morphokinetic features) learns statistical patterns associated with higher versus lower implantation potential — including the timing cutoffs and abnormality flags described in earlier stages, plus subtler pixel-level morphology cues that are difficult for humans to quantify consistently.
The output is typically a numeric score (e.g., on a 1–9.9 scale as popularized by commercial KIDScore-style tools) or a categorical tier, generated automatically in seconds once the embryo's culture period and imaging are complete.
Validation studies of commercial AI/KIDScore-style ranking tools have generally reported modest but statistically real improvements in identifying embryos with higher implantation and live-birth probability compared with morphology grading alone — the models are decision-support aids that augment, rather than replace, embryologist judgment.
In a typical IVF cycle, several embryos reach the blastocyst stage simultaneously. With single-embryo transfer now the recommended standard in many countries (to minimize the health risks of multiple pregnancy), choosing which single embryo to transfer first — and in what order to use frozen embryos in subsequent cycles — is a high-stakes decision.
The AI ranking score, combined with conventional morphology grading (inner cell mass and trophectoderm quality) and any available PGT-A ploidy results, allows the clinical team to rank the entire cohort from most to least likely to implant. The top-ranked embryo is typically selected for fresh or first frozen transfer, with the remainder cryopreserved in ranked order for potential future use.
This workflow converts what was historically a largely subjective, single-timepoint visual assessment into a reproducible, data-driven ranking built on the full developmental history of every embryo in the cohort.
AI viability scores are probabilistic, not deterministic — a highly ranked embryo can still fail to implant, and a lower-ranked embryo can occasionally succeed, because implantation also depends on factors the video cannot capture (endometrial receptivity, immunological factors, embryo genetics beyond what morphokinetics can infer). Ranking models are trained on population-level historical data and can carry biases from the clinics and patient populations used to train them, so most regulatory and professional guidance (e.g., ESHRE time-lapse consensus statements) frames these tools as decision support rather than autonomous decision-makers.
In practice, the embryologist reviews the AI ranking alongside the raw time-lapse video, morphology grading, and any genetic testing results, retaining final judgment — particularly for edge cases such as embryos with a strong AI score but an isolated abnormal cleavage flag, where clinical experience continues to add value beyond what the algorithm alone provides.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| t2 (first cleavage) | ~25–27 h post-insemination/ICSI | Timing of maternal-to-embryonic genome activation onset | Early timing correlates with higher blastocyst rate |
| cc2 (t3 − t2 interval) | Typically 8–20 h | Duration of the second cell cycle | Short, well-timed cc2 predicts higher viability |
| s2 (t4 − t3 synchrony) | Ideally <1 h | Synchrony of the 3-cell → 4-cell transition | Tight synchrony associated with euploidy likelihood |
| Direct cleavage flag | t2 → t3 interval <5 h | Abnormal multipolar division, chromosome mis-segregation risk | Strong negative predictor — deprioritized in ranking |