Announcing the launch of the Medical AI Research Center (MedARC)
Introduction
We are proud to announce the launch of Medical AI Research Center (MedARC), a novel, open, and collaborative approach to research dedicated to advancing the field of AI applied to healthcare.
The progress of AI in other fields has had huge leaps, enabled by the training and utilization of large-scale models: GPT-3/ChatGPT has near-human performance at various text processing tasks, CLIP has enabled new multi-modal applications, and Stable Diffusion has provided efficient photorealistic text-to-image generation to the masses.
These large deep learning models, sometimes called foundation models, have enabled novel, previously inconceivable, applications. However, many of these advances have not had relevant use cases for medical AI for several reasons. This is because these models are domain-agnostic and not trained with medical data specifically, and therefore have limited medical knowledge and understanding.
We therefore believe that the research and development of large deep learning models specific to medical applications shows great promise.
Recently, we’ve seen very successful open-source, decentralized, initiatives result in impactful research in deep learning. For example, EleutherAI has developed the Pile dataset, one of the most-used datasets for training large language models, and released GPT-NeoX-20B, one of the largest publicly-available open-source large language models. OpenBioML has already replicated the results of AlphaFold and completely open-sourced their results, within only a few months of its existence.
We believe similar approaches will be beneficial for medical AI research. We establish an open and collaborative research community with access to computational resources and relevant expertise to pursue research on foundation models tailored to the medical domain.
While we place an important emphasis on the development of foundation models for medicine, we do not limit ourselves to this topic.
Successful medical AI research requires an interdisciplinary team of clinicians with a good understanding of medical problems as well as machine learning (ML) researchers and engineers who can apply the relevant state-of-the-art ML solutions or develop new solutions tailored to a specific clinical need.
For context, our team has worked on another initiative, WAMRI.ai which put together such interdisciplinary teams to tackle specific clinical problems over the period of several months at University of San Francisco. This initiative was quite successful, resulting in a few startups and publications, including a publication in Nature Methods.
MedARC is inspired by the success of WAMRI. Our collaborative and decentralized organization will bring together these disparate groups of experts and build such interdisciplinary teams to address various clinical needs with AI/ML solutions.
We are interested in having machine learning researchers, clinicians, academics, and others get involved in our mission to make a difference in healthcare with AI. We welcome people to join our current projects or also propose new collaborative research projects, which we can help accelerate with our large-scale compute resources!
About MedARC
We believe a novel, open, and collaborative approach to AI research in medicine is needed. Medical AI Research Center (MedARC) was created to develop large AI models, often termed foundation models, for medicine and build interdisciplinary teams to address clinical needs.
An important principle of MedARC is doing science in the open. Models and datasets will be open sourced whenever possible, in accordance with any privacy and legal concerns. We aim to publish all our results as preprints and in reputed peer-reviewed journals and conferences.
We have established a public environment with asynchronous communication, a Discord server, which allows others to discuss, contribute and follow research progress. This open approach has been successful for AI research in NLP (EleutherAI), multimodality (LAION), and biology (OpenBioML). Therefore, we believe a similar environment will be beneficial for medical AI research and enable new opportunities.
We note that OpenBioML has been successfully conducting research at the intersection of AI and biology, and there is some overlap between the mission of OpenBioML and MedARC. However, we differ from OpenBioML by instead focusing on medical and clinical applications, researching how AI can be used more directly to impact patient care.
We have several projects underway already. This includes:
Real-time Reconstructions of Visual Perception from fMRI under the leadership Dr. Paul Scotti. This work is focused on using generative AI to reconstruct images from just fMRI signals.
Fine-tuning Stable Diffusion for Chest X-ray Generation under the guidance of Dr. Akshay Chaudhari from Stanford University. Our preliminary results are shared in our first manuscript in arXiv preprint, we’ve also had news coverage about this and we are currently extending this work.
Projects have various tasks and research directions that collaborators can explore. Computational resources will be allocated accordingly as well. A few other projects are also currently in development.
Our Team
Tanishq Mathew Abraham, founder and CEO is a 5th-year Biomedical Engineering PhD candidate at University of California, Davis who is at the cusp of getting his doctoral degree at 19. Prior to joining the doctoral program, he completed his biomedical engineering degree at 14 from UC Davis.
From a very young age, since 9-years-old, Tanishq has always been interested in the intersection between medicine and technology, and he aims to make a positive difference in this world by developing new technologies to improve clinical outcomes.
He pursued education in biomedical engineering to achieve this goal, and during his PhD he became convinced that AI is the next frontier of medicine, and is pursuing research in applying generative AI to microscopy and digital pathology. He has now founded MedARC to further explore the full potential of AI for clinical applications.
He has presented his work at several prestigious conferences such as SPIE Photonics West, an ICML workshop, BMES, etc. He has also co-written a chapter in the book “Artificial Intelligence and Deep Learning in Pathology”.
Tanishq has contributed to many open-source projects, including being on the founding team for DALL-E mini, being a frequent contributor to the fastai library as well as an instructor in their well-attended fast.ai course, and the lead developer of the UPIT library.
Tanishq is also a part-time employee at Stability AI. In addition to his role leading MedARC, he serves as a researcher for fast.ai and CarperAI and EleutherAI both of which are Stability AI supported research communities.
Tanishq is active on Twitter and known for his well-curated machine learning, generative AI and medical AI research content shared with his >36K followers.
To learn more about Tanishq, visit his website: https://tanishq.ai
Jeremy Howard, President, is a renowned data scientist, researcher, developer, educator, and entrepreneur. Jeremy is a founding researcher at fast.ai, a research institute dedicated to making deep learning more accessible, and is an honorary professor at the University of Queensland. Previously, Jeremy was a Distinguished Research Scientist at the University of San Francisco, where he was the founding chair of the Wicklow Artifical Intelligence in Medical Research Initiative (WAMRI).
Jeremy was the founding CEO of Enlitic, which was the first company to apply deep learning to medicine, and was selected as one of the world’s top 50 smartest companies by MIT Tech Review two years running. He was the founding CEO of two successful Australian startups (FastMail, and Optimal Decisions Group–purchased by Lexis-Nexis).
Earlier to that, he has worked 8 years in management consulting, at McKinsey & Co, and AT Kearney. Jeremy has invested in, mentored, and advised many startups, and contributed to many open-source projects.
To learn more about Jeremy, visit his website: https://jeremy.fast.ai/
Partners
While MedARC is an independent research organization, it is through the support of our partners that we are able to pursue our open approach to medical AI research. Stability AI serves as our main partner.
Stability AI is excited to support open-source AI research that aligns with their mission of AI for the people, by the people. Stability AI provides compute resources, funding, and relevant expertise to help our researchers.
We are also glad to announce our partnership with Stanford Center for Artificial Intelligence in Medicine & Imaging (AIMI) to lead projects in self-supervised learning and foundation models for medicine.
More Information
We would like to invite clinicians and/or machine learning researchers and academics, interested in tackling impactful medical AI challenges, to join our community!
For more information about MedARC and our research community, please visit our website at https://medarc.ai
Contact: Tanishq Mathew Abraham (tanishq@stability.ai)
