AI-Powered Embryo Selection for IVF

Figure 1: Demo of the embryo-assessment pipeline analyzing time-lapse microscopy video.

In IVF, embryo selection is where clinical outcomes are won or lost—and it still runs on manual morphological assessment. Embryologists review static images or time-lapse videos to spot developmental milestones and grade quality, a process with well-documented inter-observer variability, a dependence on scarce expertise, and hard time constraints that make it impossible to thoroughly review days of development for every embryo.

Time-lapse incubators made the problem sharper, not easier. They capture continuous video from fertilization through blastocyst formation—thousands of frames per embryo—creating an opportunity for quantitative analysis that no human team can exhaust. Subtle morphokinetic patterns predictive of developmental potential and chromosomal abnormalities sit in that footage, beyond what manual review can reliably extract.

The client needed a production system, not a research prototype: one platform that localizes the embryo in every frame, filters unusable footage, detects developmental stages, grades morphology, reconstructs the morphokinetic timeline, and estimates clinical outcomes. It had to work across the multiple commercial time-lapse imaging platforms found in real clinics—each with different optics, spatial resolutions, and temporal sampling rates—handle the messy realities of embryo imaging like variable positioning, inconsistent illumination, and partial occlusion, and serve results through standardized APIs with inference fast enough to slot into daily lab workflow. Most prior automation efforts tackle isolated prediction tasks; this engagement demanded the full pipeline, from pixels to clinical report.

We built the system as a seven-stage pipeline of specialized models, each optimizable independently while feeding the next:

Because embryo videos contain thousands of frames—far beyond what dense sampling can afford—each temporal task gets its own sampling strategy, with non-linear schedules that sample early development densely and later stages sparsely, matching the biological weight of early cleavage timing.

The most striking capability is non-invasive euploidy prediction: estimating chromosomal normality before genetic testing, from video alone plus routine clinical data. A two-stage ensemble combines Video Swin Transformer predictions on pre-biopsy video (60–128 hours post-insemination) with maternal age and Day-5 morphological scores through calibrated logistic regression. It does not replace preimplantation genetic testing—but it gives clinics a principled way to prioritize embryos when testing is cost-prohibitive.

Deployed on model-serving infrastructure with hardware-accelerated inference, single-frame analysis runs with sub-second latency, and video sequences process through efficient batching.

The system demonstrably:

Two honest boundaries define the system's role. Morphological grading performance remains below the inter-observer agreement levels of expert embryologists on reference datasets. And clinical outcome prediction, while significantly exceeding random baselines, does not reach thresholds sufficient for autonomous decision-making without clinician oversight—outcomes depend on maternal factors, endometrial receptivity, and transfer procedures the model never sees. The system is decision support that scales embryologist attention, not a replacement for it.

Event detection built for biology, not just vision. Morphokinetic events are transitions, so we detect them as change: a Siamese network learns frame embeddings where distance tracks morphological difference, and the timeline algorithm computes distance signals between temporally adjacent frames, detects peaks with adaptive thresholds, maps classifier predictions to an initial event sequence (t2–t9, tB, tM), then refines timing with isotonic regression—enforcing the biological constraint that developmental events occur in order. Contrastive training also stretched limited labels, since pairwise comparisons exploit unlabeled frames. The result is a timeline that cannot report four cells before two.

A modular pipeline that earns its complexity. Decomposing assessment into seven specialized stages means each component can be retrained, improved, or debugged independently without touching the rest—segmentation upgrades do not risk grading regressions, and new prediction heads bolt onto proven upstream stages. It also stretches expensive expert annotation: multi-task heads share backbone representations across correlated labels, so every embryologist-labeled frame trains several capabilities at once.

Cross-device generalization as an architectural commitment. Rather than hoping augmentation would paper over hardware differences, the pipeline maintains per-device preprocessing branches: some platforms use centroid-based cropping with scale correction derived from known microns-per-pixel ratios, others use direct spatial resizing—normalizing what each optical system actually delivers before any model sees a pixel. Combined with mixed-platform training and device-specific fine-tuning, this made one system deployable across clinics instead of one model per incubator brand.

Evaluation discipline that survives clinical scrutiny. Data splits were stratified at the patient level to prevent leakage, and temporal models were tested on embryos from dates chronologically after the training period—the honest simulation of deployment. That rigor is also why we can state the system's limits precisely: we know where it falls short of expert agreement because we measured it the hard way.