Computer vision for in-vitro fertilization (IVF)
Artificial intelligence has a lot of use cases in the healthcare industry. These applications are often based on computer vision and medical imaging. AI assists doctors in diagnosing cancers, for bone age calculations or for DNA sequence analysis.
Recently, I discovered an interesting application of computer vision in the healthcare industry. It makes in-vitro fertilization (IVF) easier and increases the success rate. This technique allows better decision-making for embryologists by giving objective metrics calculated by a computer vision model.
Let me explain how it works!
How does in-vitro fertilization works?
Infertility is a common problem that affects 186 million people around the world. And every year percentage of Europeans suffering from infertility increases by 8 to 9%. In France, it concerns about one couple out of 6.
Many couples can rely only on IVF to have a child. During IVF, the doctor fertilizes with sperm an egg removed from the woman’s ovaries. The fertilized egg is what we call the embryo. It’s cultivated in a laboratory under strict conditions for a few days. After this process, the embryologists inject this embryo back to the woman’s womb.
Embryologists can follow the process of development of the embryo through a device that allows reproduces the perfect development conditions. This device helps doctors to monitor the process continuously (non-invasively). This tool allows having this type of sequence that show the state of the embryo at regular time intervals. This is called time-lapse imaging (TLI):
Doctors often grow several embryos, monitor their development and decide at the end of the process which one will be injected. This can be expensive, complex and emotionally difficult for the patient.
The success rate of IVF is about 35%, and has been increasing very slowly since the 2000s. Moreover, many couples have to repeat the process several times without any guarantee.
Computer vision allows the selection of the “right” embryo
The images extracted during the culture allow professionals to determine which embryo is most likely to lead to a pregnancy. This selection is based on the experience and knowledge of the doctor, and the criteria are not always objective. That’s why we can see a big variability in embryo selection between doctors.
Like many medical techniques, in-vitro fertilization could benefit from recent advances in computer vision.
Modern computer vision techniques can make it possible to objectify this choice. We can compute a score that gives the probability for a given embryo to lead to a pregnancy.
To make this prediction, we combine convolution neural networks techniques ( ResNet or U-Net architectures for example) with a temporal approach (LSTM type). Convolution models allow to extract features for each frame of the sequence, while the temporal models allow to quantify the relationship between frames.
These models are trained either with data labeled by doctors or with data for which the outcome of the process is known (pregnant or not).
One of the start-ups working on this subject is ImVitro. It is a French start-up that offers embryologists and hospitals a computer vision model, which allows them to classify embryos according to their quality.
Conclusion
Although promising, this technique has its limits. First, even if technically the models used are quite “simple” and do not present any innovation, the amount of available data is limited, and their use is regulated because of their sensitivity.
Moreover, we do not only want to have a confidence score for each embryo, but we want this score to be explainable and based on measurable metrics. Here we find the famous performance-explainability dilemma.
