May 19, 2020
So many embryos look so similar to each other, and until now, to determine which is best for implantation, they have been subject to human analysis, which unfortunately is flawed. While some embryologists may annotate time lapse images for hours in the lab, this approach is still reliant on human beings – each of which may have a different interpretation – and even if it were a perfect approach, it is far from scalable to cover every embryo in every IVF process.
As a self-proclaimed AI geek, together with my co-founders, we set out to find a better way, using computers. Annotating photos is the computer equivalent of what is called “supervised machine learning.” A simpler example is if you want a computer to recognize a square, you can give the computer a rule that squares have four corners, then when you show it a picture of a circle it will identify that it is not a square. A much more powerful tool is “unsupervised machine learning,” which is to feed the computer with examples of circles and squares, and let the computer figure out its own rules for differentiating between the two. Believe it or not, computers can do the same thing, on a much more complex level, of course, with images of embryos.
So, when selecting the best blastocyst to implant in an IVF procedure, unsupervised machine learning can be a major help. By feeding the computer with a massive number of examples of embryo images, the computer can, by itself, determine the relevant features to consider, which could make the implantation a success or a failure. For example, the timing of the appearance and disappearance of two pronuclei can be a very strong predictor.
I discuss this much more in detail in this webinar hosted by the International IVF Initiative. I was part of a panel discussion with several of my peers in a session called “Decision-making, AI and ART.” My part starts at 45:47, but I encourage you to listen to the whole presentation.
Join us on our quest to make IVF more consistent, predictable and cost-effective.