Disclaimer: EMA by AIVF has CE mark. In the US, an AI component of EMA is considered an investigational device. Limited by Federal (or United States) law to investigational use.
©All rights reserved AIVF 2023

Understanding Artificial Intelligence in Embryology

Yishay Tauber PhD, Head of AI at AIVF | Nov 16, 2023

Introduction to Artificial Intelligence

Artificial intelligence (AI) is the application of computer science and technology to problem solve and perform tasks that are typically associated with human intelligence. Machine learning (ML) is a subfield of AI – it refers to the development of systems that automatically learn tasks and inferences from data.

In recent years, the application of AI/ML in healthcare has revolutionized the way we prognose, treat, and counsel patients in almost every sector of medicine. Healthcare AI/ML is often applied to analyze the relationships between clinical data and its associated clinical and/or patient outcomes. Health care applied AI/ML typically has the following pattern: the system is fed vast amounts of data, mathematical algorithms are applied to automatically learn insights and recognize patterns in the data, this information is then used to make better clinical decisions.

These systems have the ability to draw insights from clinical data that exceed human capacity, thereby allowing more accurate clinical diagnoses, optimized patient care, and more personalized treatment.

Deep Learning in the Field of Artificial Intelligence

Although there are multiple AI/ML approaches, when talking about the application of AI in healthcare, we usually refer to the use of supervised deep learning -based techniques. Supervised deep learning is a type of AI/ML that relies on teaching AI systems to process data in a way that is inspired by the human brain. Deep learning uses layers of neural networks in a manner that mimics the brain’s learning process; each layer progressively extracts more features and processes data until a final output is computed.

Algorithms learn the relationship between the input data and the desired output using trained labels found in the dataset.  Through iterative learning, these algorithms grasp underlying patterns in the labeled data, thereby enabling them to make accurate predictions or classifications when presented with new, unseen data.

Current Need

Just as other medical professions are experiencing shortages around the world, there is a severe lack of trained and experienced embryologists to perform necessary IVF laboratory workflows. In the US, an estimated 5% of the existing  embryology workforce is beyond retirement age (65-67 years old) and an additional 20-40% will reach retirement age within the next seven years [1].

Furthermore, conventional manual tasks, often referred to as “dry” procedures, account for the majority (~60%) of occupational IVF laboratory workload. These tasks include: embryo monitoring, manual embryo quality annotation,  treatment data collection, and data transfer to the electronic medical record (EMR). These tasks are typically performed by skilled embryologists and require sufficient laboratory staff and occupational hours, which comes at a significant cost to the clinic.

Relying on manual (human operator) performance for these IVF laboratory tasks can pose various challenges. These include the risk of manual data transfer errors, occupational stress/fatigue, communication failure, unnecessary repetitive  workflows, decreased laboratory efficiency, and operator variation. To enhance overall clinic productivity and efficacy, the IVF clinic should consider implementing streamlined processes and leveraging technology whenever possible.

Furthermore, conventional embryo evaluation methods performed by embryologists typically rely on visual assessments of embryo quality and manual annotations. Though standard in practice, this method is time-consuming for embryologists and fraught with intra – and inter – observer variability. There is a need for objective, data-driven tools to improve embryo evaluation and pregnancy success rates.

The Application of AI in Embryology

The application of AI in embryology relies on a combination of deep learning algorithms and computer vision. It is a rapidly growing area of research due to its potential to drastically improve clinical outcomes in the IVF clinic and overcome the current birth rate plateau. For example, AI can optimize embryo evaluation/selection by detecting bio-features in the embryo that are important to its likelihood of pregnancy success. AI systems can also take into account patient data that may influence IVF success to provide more personalized patient counseling and prognosis. These systems have the potential to reduce the number of cycles and time it takes to get pregnant. The result: more efficient and data-driven IVF.

Using a clinical dataset that consists of time-lapse embryo images and videos, associated patient metadata, known clinical  outcomes, and embryologist annotations, deep learning algorithms can be built to predict IVF treatment success.

Clinical Outcome-Based Approach to Deep Learning

The most straightforward way to use AI in embryology is to train a neural network using a clinical outcome-based approach. For example, algorithms rely on embryo time-lapse video sequences along with their known clinical outcomes to identify visual features encapsulated in the embryo that are important to its likelihood of pregnancy success. Here, the clinical outcome is the desired target of the network training.

During the training process, the network can automatically identify the critical factors that contribute to a successful clinical pregnancy. The network’s predictions are strongly correlated with embryo quality assessments conducted by embryologists. Deep learning algorithms may significantly improve conventional embryo evaluation workflow by sensitively detecting features in the embryo that are invisible to the naked eye. For example, algorithms can distinguish between similar-looking embryos by their likelihoods of pregnancy success in a way that is difficult for an embryologist to do using conventional visual methods only. Furthermore, these algorithms are also not subject to bias and observer variability. In this way, the algorithms introduce streamlined objectivity and data-driven decision-making to the embryo evaluation workflow.

Knowledge-based Approach to Deep Learning

In spite of the fact that outcome-based networks are powerful tools for predictive modeling in IVF, they come with several limitations. As they heavily rely on existing data to make predictions, encountering anomalous events that aren’t well represented in the dataset can present a problem. In addition, since the network automatically learns the features that are important to the clinical outcome, it is hard to ensure that well established factors are indeed a part of the network decision. Laboratory interventions and/or biological scenarios that naturally occur during in vitro embryo development may also affect network results if not taken into account.

2PN detected


In order to address these issues, rather than only rely on outcome-based networks, a suite of specialized networks may be trained to detect key events in embryo development and incorporate them into the final algorithm. For example, algorithms can be trained to automatically detect known morphokinetic and developmental parameters that are critical to embryo development success.

One example of this type of approach is a knowledge-based automated pronuclei quantification algorithm that measures pronuclei appearance and fading times as well as kinetics to ensure adequate fertilization. Another example includes an image segmentation algorithm, which relies on AI and computer vision to accurately detect and quantify embryo shape, size, and dynamics.

A third example includes a knowledge-based morphokinetics network, applied to track key morphokinetic parameters during embryo development, count and detect abnormal cell cleavage, detect proper morula and blastocyst formation, and track kinetic over time.


The application of AI in the IVF clinic holds significant promise. By addressing the current challenges in the field, AI has the potential to introduce streamlined workflows and increase embryo selection accuracy without compromising on clinic confidence, transparency, or outcomes. Future research should be directed toward AI-based technologies that can accomplish this.