During my time as a Lead Computer Vision Algorithm Developer at AiVF (2023-2024), I developed AiVF’s flagship deep learning models, which formed the foundation of the company’s technology. AiVF leverages AI to estimate embryo viability and improve IVF success rates by analyzing time-lapse videos of embryos from insemination (zygote stage) to day 5 (blastocyst stage).
I developed three core deep learning models:
- 3-Day Embryo Viability Prediction Model
- 5-Day Embryo Viability Prediction Model
- Non-Invasive Preimplantation Genetic Testing (PGT) Model
These models predict:
- Embryo viability at days 3 and 5
- The likelihood of an embryo having a negative preliminary genetic test result (PGT)
A visual example of embryo development over time is shown in this time-lapse video:
Data, Model Architecture and Training
To build a robust AI model, we gathered a large dataset of time-lapse embryo videos captured during incubation. Since storing full-length videos is computationally expensive, I implemented a preprocessing pipeline to select key frames that carried the most biologically relevant information.
For training, I used the Video SWIN Transformer—a transformer-based architecture capable of learning both spatial and temporal features. It provided:
- Self-attention mechanisms to capture key developmental markers
- Better generalization across diverse datasets
- Improved efficiency over standard CNN-based models
Since each video contained multiple images, memory usage was a key challenge. To handle this, I implemented the models using PyTorch with Lightning, which allowed for multi-GPU training. Lightning enabled:
- Scalability for handling large video datasets
- Seamless distributed training
- Optimized memory management via automatic gradient accumulation
Evaluation and Real-World Impact
To assess accuracy, we tested the model using categorical cross-entropy on real-world fetal heartbeat data from weeks 5 to 7 of pregnancy. However, beyond standard validation, the true business success of the model was measured by its impact on IVF efficiency.
A key metric for clinics is “transfers per pregnancy”, representing how many embryo transfers are needed before achieving a successful pregnancy. Before using AiVF’s AI models, clinics averaged ~1.6 transfers per pregnancy. After integrating our AI-based selection, this number decreased substantially (exact figures are confidential), directly improving:
- Higher success rates per transfer
- Lower financial and emotional burden for patients
- Increased efficiency in fertility clinics
Conclusion
Developing AiVF’s deep learning models was a challenging but rewarding experience. Our AI didn’t just improve predictive accuracy—it had a measurable real-world impact by reducing the number of transfers needed per pregnancy and improving overall IVF outcomes.
From a technical perspective, leveraging PyTorch Lightning for multi-GPU training was a game-changer, enabling efficient video-based model training at scale.
As AI continues to evolve in reproductive medicine, I believe deep learning-driven embryo assessment will become an industry standard, helping more families achieve successful pregnancies.