We spend most of these updates on our own work, so this month we wanted to step back and share some of what we have learned about the wider field we are part of.
One of the most important studies in this space came out of UCSF in 2021, led by Dr. Rima Arnaout, a cardiologist there. Her team trained a set of neural networks on more than 100,000 ultrasound images taken from about 1,300 pregnancies, then tested the models on scans from thousands more. The goal was to teach the models to recognize the standard views doctors look at during a screening ultrasound, and to tell the difference between a normal heart and a complex defect. When compared to expert readers, the models matched doctors at catching problems and made fewer false alarms. It remains one of the strongest pieces of evidence that this kind of work is possible.
On the commercial side, a Paris-based company called BrightHeart, founded by two pediatric cardiologists, built a similar tool trained on over 90,000 ultrasound exams. In November 2024 it became the first AI software of this kind cleared by the FDA, and it is now being used in a small number of hospitals, including a pilot at Mount Sinai in New York.
Both of these examples focus on reading a scan that has already been captured. What we keep coming back to is the step before that. Getting a clear, usable image in the first place depends heavily on the skill and experience of the person holding the probe. A good detection tool only helps a family if the scan behind it was captured well. And the clinic performing the scan, also has to have access to the tool to begin with. Improving the odds of catching a defect, and making sure that improvement reaches as many families as possible is the whole point of what we are building.
