Today marks a career highlight for me. For the last 130 years, modern anatomic pathology has relied on the ability of a human to look at something under a microscope and understand what they’re seeing. Getting AI to help with this has been hard, in part because of the gigantic amount of information in a typical pathology slide, which could contain a billion pixels. I’m honored to have sponsored a research team at Providence, Microsoft and the University of Washington who published a major breakthrough today in the journal Nature. They cracked the computational challenge of modeling whole pathology slides using an adaptation of Microsoft’s LongNet, which allows for long-context modeling. This allowed the team to develop a ‘whole-slide’ foundation model for digital pathology, pretrained on 1.3 billion pathology images. This new model is able to discern between different types of cancer, and even determine which mutations are present in a cancer, just by ‘looking’ at a slide. This foundation model is called Prov-GigaPath, and we’re making it open access for researchers. The medical community finally has an AI tool that’ll help that human with a microscope ‘see’ things they didn’t realize were there. I’m hugely proud of this work and grateful for the opportunity to support it. #AI #medicalresearch #foundationmodel #pathology #science #breakthrough
It’s such a blessing to have you and your brain dedicated to improving health care; your contributions are innumerable and positively impactful. Thank you for all you do. 👏🏻
Gamechanging work 👏. Mindblowing how a single foundation model can now discern between different types of cancer, or even detect which mutations are present 🚀, just by ‘looking’ at a slide. This breakthrough redefines decades of traditional AI approaches to healthcare, and will positively disrupt how pathology🔬 is currently done. Bravo, foundational work that will save millions of lives!!
Ari is a unicorn. Anyone who was privileged to work with him knows that
You never cease to surprise me, but always impress me Still feel lucky I got to work under you even if it was for a short time
Incredible. Amazing to think of the impact this may have. 👏🏼👏🏼👏🏼
I read about this project, Ari, and didn’t know you were part of it. Congratulations. You will assure its success for patients, providers, and systems. So much opportunity for this technology to do even more. Thanks!
Well done Ari Robicsek and team Microsoft, Providence and University of Washington. This is break-thru and shows the power of AI to impact care in a major way! Empowering Pathologists, Researchers and Patients with more precise data! I remember visiting the technicians in Oregon digitizing slides and thinking about the impact this would have! So impressed.
Incredible
Wow, this is really incredible work Ari! I am looking forward to seeing how this is going to be used in clinical practice, which is always the challenge for AI.
Generative AI Engineer and Consultant | Machine Learning Engineer | Ph.D. Biomedical Engineering
1moAmazing work! Ari Robicsek based on the graphic it seems like you've gone for a one sided transformer for the slide-level embedding. Can you give any insight as to what might be expected if using a bidirectional transformer? My suspicion is that the slide pair masking process helps to generalize enough so the bidirectional transformer isn't necessary, but I might be having my own hallucinations! Thanks for your and your team's great work!