Excited to introduce Cognita with @loublanks and @Dr_ASChaudhari.
The field of radiology AI has been developing for years. My observations are (1) Products have emerged over time, but they haven’t unlocked compelling use cases; (2) The hype-driven environment incentivizes people to chase the latest buzzwords rather than solving the problems from first principles.
Therefore, from the very beginning of the journey, we defined and differentiated ourselves through two core principles: (1) a deep understanding of the underlying technologies, and (2) an unwavering drive to create real-world impact. The former enables us to identify the correct problems to solve and innovate effectively; the latter keeps us grounded and prevents us from drifting off course, rather than starting with a shiny new hammer and then looking for nails. These principles guide our need to demonstrate a substantial real-world ROI, rather than staying at the level of academic benchmarks.
Acting upon our principles, we were fortunate to (1) collaborate with an excellent group of radiologists experienced in AI evaluation, education, and deployment from Radiology Partners, (2) collect large-scale and representative datasets, and (3) secure significant compute resources. We connect these dots with our internal technical breakthroughs to design an end-to-end, scalable system that emulates the work of radiologists in interpreting radiology images.
The results have been inspiring: Our models achieve up to four times fewer diagnostic errors than baseline radiologists in clinical practice across different body parts, while also improving the radiologist's efficiency. We are excited to expand this impact to make healthcare accessible. Our speed and level of impact will amplify as we join forces with @Rad_Partners and Mosaic Clinical Technologies, @Rad_Partners' technology division.
To scale our vision, we are actively hiring exceptional people who are the best in their domain to engineer solutions with us. Specifically, if you are excellent at data engineering, modeling, model evaluation, AI infra, or full-stack software development, please consider applying at https://t.co/xGx76P81nN.
Today, @zhjohnchan, @Dr_ASChaudhari and I introduce Cognita.
Over half the world lacks access to sufficient healthcare. We started Cognita a year ago to address this.
Radiology is (1) the first-line diagnostic specialty, (2) facing a worsening workforce shortage, and (3) highly digitized, enabling AI to have an enormous impact. Stage 1 of our company is focused on increasing the world’s access to radiology.
To do this, we are building models that emulate radiologists - describing in detail hundreds of potential imaging findings and comparing to prior studies. This is a departure from existing narrow radiology AI solutions that only provide a yes/no answer for a specific finding.
Our goal is to build copilots that help radiologists perform more accurately, efficiently, and with higher satisfaction, leading to reduced missed diagnoses and shortened patient wait times.
We partnered with @Rad_Partners from day 1 to make this vision a reality. This collaboration has given us access to the required scale of data, and a hand-in-hand partnership with radiologists who have more experience validating and deploying radiology models than any other team on the planet.
What our teams have accomplished together over the past year is extraordinary, seeing unprecedented performance across X-ray and CT.
Stage 2 will focus on adding diagnostic capabilities that extend beyond current radiology practice, such as risk prediction and quantification across time. Stage 3 will incorporate additional data types - clinical notes, medical records, labs, omics, pathology - to deliver improved diagnostics and personalized treatment recommendations.
Because our partnership has been so compelling over the past year, we decided to fully join forces with Mosaic Clinical Technologies, @Rad_Partners' technology division, through an acquisition. This creates further alignment, and is carefully structured to increase Cognita’s velocity. We strongly believe this is the right path forward to increase the world’s access to healthcare.
We are just getting started and the future of healthcare AI is incredibly exciting. If you’re motivated to engineer solutions to one of the most challenging technical problems and impact patient lives every day, there is no better place to be. Please consider applying at https://t.co/kHPplTZBjS.
We are excited to announce the launch of our company - Cognita! We are working towards building the future of radiology through multi-modal AI systems with a great group of founders @loublanks, @Dr_ASChaudhari , and I, and advisors Ajit Singh, Chris Re, and
@curtlanglotz.
We are assembling a lean team of engineers and researchers. If you're interested in making large-scale clinical impact on healthcare with AI, we would love to hear from you!
Let us know here: https://t.co/yW126D8Z1t
Very interesting paper! It would also be interesting to see if LLMs are good evaluators of novel ideas (e.g., by predicting the outstanding ICLR papers this year 😄).
Automating AI research is exciting! But can LLMs actually produce novel, expert-level research ideas?
After a year-long study, we obtained the first statistically significant conclusion: LLM-generated ideas are more novel than ideas written by expert human researchers.
arXiv -> alphaXiv
Students at Stanford have built alphaXiv, an open discussion forum for arXiv papers. @askalphaxiv
You can post questions and comments directly on top of any arXiv paper by changing arXiv to alphaXiv in any URL!
How can we regulate Health #AI to maximize benefits while minimizing risks? @StanfordHAI's Associate Director @curtlanglotz shares his expert insights on creating a safer AI-driven healthcare environment for all.
https://t.co/PSVRhOD6z0
We're very excited to release 🌟DiVA — Distilled Voice Assistant 🔊 @WilliamBarrHeld
✅End-to-end differentiable speech LM; early fusion with Whisper and Llama 3 8B
✅Improves generalization by using distillation rather than supervised loss
✅Trained only using open-access permissively licensed data from the CommonVoice
✅Outperforms existing speech LMs on QA, Emotion Recognition, and Translation Benchmarks
👉Website: https://t.co/U9jw4aVMAk
👉Model Weights: https://t.co/0bV4lgSOaI
💻Try DiVA with our side-by-side comparison to Qwen Audio and SALMONN. Feedback is welcome 🤖
Two weeks of learning, research, mentorship & fun wrapped up with the conclusion of the AIMI Sumer Research Internship & Bootcamp! Thanks to our participants, staff, mentors & featured speakers who made this year's programs a huge success!
Excited to share our R-Tuning got an outstanding paper award@NAACL 2024! Take a look at this paper to see how to align your LLMs to honesty. https://t.co/DWZ1nTQalJ This work is finished during my visit at UIUC. Thanks for Prof. Ji and Prof. Zhang’s supervision!
Introduce HumanPlus - Shadowing part
Humanoids are born for using human data. We build a real-time shadowing system using a single RGB camera and a whole-body policy for cloning human motion. Examples:
- boxing🥊
- playing the piano🎹/ping pong
- tossing
- typing
Open-sourced!
⭐️ Check out Merlin led by @loublanks and mentor @Dr_ASChaudhari in this thread. Trained with diagnostic codes and rad reports and evaluated across various tasks comprehensively. @loublanks really laid a *solid foundation* for CT *foundation* models. :)
🧙 Excited to introduce Merlin, a vision language foundation model for 3D computed tomography 🐈⬛🩻
Trained to understand 3D abdominal CT scans using supervision from:
💾 Structured electronic health records (1.8+ million codes)
🗒️ Natural language radiology reports (6+ million tokens)
Paper: https://t.co/ljneXOpKFC
🧵 1/10
⭐️ Explore CheXpert-Plus in this thread, a CXR dataset including radiology reports, demographics, and structured labels @StanfordAIMI
📄 Paper: https://t.co/T3hWa5lPdu
💾 Dataset: https://t.co/dTr5mDFDjh
Led by @PierreChambon6, @IAMJBDEL, and @curtlanglotz.
Five years ago, thanks to the leadership of @mattlungrenMD, @stanfordAIMI released the CheXpert images: 223K JPG CXRs with labels for 14 conditions. CheXpert has been cited >6000 times, mostly related to development of supervised learning methods. Much has changed since then.🧵
Our clinical #NLP work just published in @NatureMedicine! We present a framework to adapt & evaluate #LLMs for summarization. Physicians 🩺 prefer #LLM summaries to those of #medical experts❗
Big step to reduce documentation 📚 and focus more on personalized care 🙌
A 🧵