For children born with congenital heart defects (CHDs), early diagnosis before birth is essential to ensure that the baby will receive optimized care after delivery for the best possible outcome. However, only 34% of congenital heart defects are detected prenatally. BrightHeart, a medical technology company based in Paris, has made it their mission to improve fetal heart screening. The company has created AI-powered medical software with the support of Meta’s DINOv2 model to help clinicians identify or rule out signs suggestive of congenital heart defects faster and more accurately, with the aim of improving the prognosis for affected children.
BrightHeart recently achieved FDA 510(k) clearance for its first artificial intelligence software at an unprecedented pace—only two years after the founding of the company—in part thanks to the efficient deployment of advanced tools like Meta’s open source DINOv2 model to accelerate R&D. Building on the foundations of the 2021 DINO model, DINOv2 uses self-supervised learning to gain a deeper understanding of images and video. This allows for extremely accurate video classification performance, which can aid in the early detection of congenital heart defects. By leveraging the power of AI and machine learning, BrightHeart aims to revolutionize fetal heart screening and give children with congenital heart defects the best possible chance at a healthy and happy life.
BrightHeart was founded by two pediatric cardiologists who recognized a significant gap in the detection of CHDs, which are often overlooked due to the complexity of the ultrasound exam, the high level of expertise and experience required for proper detection, and the variety of morphological presentations of CHD in a fetal heart that typically measures less than 1 cm. This oversight can have serious consequences, significantly impacting morbidity and mortality.
Given the significant negative consequences surrounding undetected CHD’s, the BrightHeart team is urgently working to bring their innovative software to market as quickly as possible. Based on this, it was important during the R&D stage to select tools that not only met their strict standards for privacy and security, but also delivered unparalleled outcomes at record speed. From the start, DINOv2 stood out not only for its quality and innovation, but also the fact that its open source nature allows the model to be downloaded and used privately, with all the data staying strictly within the BrightHeart ecosystem.
“Our main goal when selecting our neural network tool was to have a high quality, innovative, and efficient neural network that we could also run within a matter of minutes,” says Eric Askinazi, a lead data scientist at Brightheart. “Our evaluation process started with nearly a dozen candidates for classification, and DINOv2 was the clear choice.”
The team leveraged DINOv2's open source nature to accelerate their product development process. With time being their biggest constraint, they found that a pre-trained model was incredibly convenient and allowed them to focus on integrating the technology into their solution rather than building it from scratch. BrightHeart trained their models using DINOv2 as a foundation in order to analyze video clips from ultrasound examinations and identify whether the examination is normal, or flag if it exhibits signs that may indicate the presence of a CHD.
“It’s allowed us to really get up to speed and to get better results faster,” Askinazi notes.
Once they are ready to go into production, the team hopes that the prenatal diagnosis rate of CHDs will increase significantly, thereby helping to avoid possible complications after birth. Askinazi says that achieving the recent FDA clearance in record time would not have been possible without DINOv2. He’s confident that BrightHeart can make a significant impact on the future of diagnoses and standard of care for newborns.
“We haven’t seen anything which comes close in terms of the ratio between performance and inference speed, and so DINOv2 has been a great tool for us,” Askinazi says. “We are pleased with our rapid product development traction so far, and now the goal is to get it in the hands of clinicians to accelerate improved patient care.”
Our latest updates delivered to your inbox
Subscribe to our newsletter to keep up with Meta AI news, events, research breakthroughs, and more.
Join us in the pursuit of what’s possible with AI.
Foundational models
Our approach
Latest news
Foundational models