Interviewed on October 27, 2020
Dr. Lungren is the Co-Director of the Stanford Center for Artificial Intelligence in Medicine and Imaging and an Associate Professor Clinician Scientist at Stanford University Medical Center. Dr. Lungren’s NIH and NSF funded research is in the field of AI and deep learning in medical imaging, precision medicine, and predictive health outcomes. His work has been featured in national news outlets such as NPR, Vice News, Scientific American, Forbes and he regularly speaks at national and international scientific meetings on the topic of AI in healthcare. He teaches the Stanford AI in Healthcare class for learners with a non-technical background that is also now available on Coursera.
Interviewed by Satvik Tripathi, Founder & Head, Techvik
Satvik Tripathi (ST): How did the AIMI Center start, what work are you doing, and what change would you like to see in the future?
Matthew P. Lungren (MPL): Yeah, so my work dovetails into a lot of what you’re seeing in the popular media about artificial intelligence and medical imaging in particular. I’m currently the Co-director of the Artificial Intelligence in Medicine Imaging Center, here at Stanford. The AIMI center was started by Prof. Curtis Langlotz with merely 4 people and just a few years later we have more than a hundred faculty and students. The primary function of the center has been to catalyze multi-disciplinary teams consisting of medical domain experts and machine learning students and faculty. We have found that there are many physicians with strong computer science backgrounds as well as computer science faculty with deep knowledge of many aspects of medicine. And now we have tried to make sure that the center can serve as a resource for groups all over the world with open access to our machine learning in healthcare journal clubs, panel discussions with experts around the globe, and open access medical imaging datasets and code; this is very similar to your work that is focused on online resources and education for machine learning.
ST: In one of your talks on Deep Blue, you spoke about the importance of datasets. The algorithm used to make Deep blue was discovered back in 1983, but the model was trained after the chess dataset was released in 1991 and in just 6 years it defeated Gary Kasparov.
ST: So, even now, one of the major issues research faces today, is the lack of datasets or adequate datasets. How do you think this could be resolved, and is synthetic data an alternative?
MPL: Yeah, I mean, that is one of the major issues preventing the rapid advancement of this field is the lack of open data - and this is not unique to healthcare. There have been several innovative solutions offered to get around the challenges of releasing medical data such as federated learning, but ultimately the single best way to advance the science of machine learning in healthcare is to have well-annotated data that is properly de-identified out in the public domain, preferably as part of a research publication and code, so that others can build on the work. This is the lesson learned with prior big machine learning breakthroughs such as deep blue or IBM Watson - that data is the key catalyst to advancements.
ST: Now as we are transitioning into AI, how clinicians and support staff at medical institutions and facilities could learn about these upcoming AI-based medical solutions?
MPL: Yeah, there are plenty of resources out there for people, like you, who are interested in technology and are ready to take up challenges as coders. We are trying to bridge the existing gap between the ends of clinicians and computer scientists. At the AIMI Center, we have learned a lot in a very short period of time, and it's an ongoing process, and this is also happening while the field is rapidly changing. So, we have tried to share information and resources, and do our best to coordinate with others doing the same such as the NIH and MIT. Eventually, we will get deeper into the application science via clinical trials and so more medical institutions will require competencies in AI, and educating enough leaders to help facilitate AI in healthcare is a big upcoming challenge that we hope our efforts can contribute to.
ST: Do you think there should be a new field of study or discipline that combines AI and clinical areas to provide better predictions and diagnosis, just like a symbolic system which combines AI with humanities aspects?
MPL: Absolutely - symbolic systems serve as an example of how two fields could be forged into one to bring out better outcomes, and this is something Stanford is known for! As more and more researchers from both the medical and the computer science field come together in this interdisciplinary science of clinical machine learning, it is possible one day we would be able to develop a new study or discipline out of the small bits of information and resources we are collecting today.
ST: And how do you see a career with a major like that, because I and many others would pick it in a heartbeat.
MPL: I think it holds great potential. As we see the outcomes of current projects and researches, it really amazes us just by thinking about the impact this work could have worldwide. We have a long way to go before realizing the full potential and so getting into this field provides both challenges and opportunities and we need as many of the best and brightest as to help us achieve what we know will be possible.
ST: How do you plan to carry forward the research approaches developed at AIMI to broader communities and how far is that vision?
MPL: So one of my personal passions is in global health, and I served as a vice president for a large global health organization that was focused on imaging, which is now the largest in the world, called RAD-AID. And in that process, we wrote a textbook on global health imaging and established core education around this topic - and we've learned a lot about pitfalls and opportunities around the world in different situations. And now with new applications in AI, we find that there are many potential opportunities to achieve the democratization of knowledge and, unlike all other major medical advancements before, this could be an inexpensive way to deliver higher quality care. So we think about a future vision where we could achieve delivering expert-level clinical AI models directly to every clinician around the world who would potentially use it or need it.
ST: There's kind of a concern for people talking about jobs as a medical student in the future, what does that look like in 10-20 years?
MLP: There have been a lot of opinion pieces and editorials written about the future of radiology, all of which ask some version of “does AI lead to the end of radiology?” And from a very superficial viewpoint it is possible to draw those conclusions because you're seeing AI model results in a test setting perform some of the tasks that we do as radiologists, and in some cases even better than radiologists. But for those of us who are doing the research and the work to study and implement these models, that narrative has not really caught on. We are still far from capturing the immense complexity that constitutes the routine cognitive work of a radiologist. Now do I think in ten years things will change? Yes of course things are going to be different - but as it stands now we have tried to focus on creating tools and exploring clinical AI applications that can best help the patients that we're serving. Because if our work leads to a future where someone can come to the emergency room in a rural part of the country, or the world, and receive the proper treatment at the same expert level of care that they would receive in large centers, then this becomes our primary mission and guides our work.
ST: I recently came across very intriguing research on the application of GANs for novel Drug Discovery methods
ST: So how at AIMI do you hope to involve Drug Discovery and Drug Delivery in future or current projects?
MPL: Yeah, absolutely, drug discovery has also been a part of clinical AI research applications, especially using Recurrent Neural Networks, as it allows you to add decision boundaries. GANs have been coming out as one of the upcoming techniques, there have been new courses starting here at Stanford, and Andrew’s Deep Learning has launched new courses on GANs on Coursera and it would be really fun to see how people use this knowledge to build better models. Also, one of the techniques which have a lot of promise in medical applications is multi-modal data fusion, which could potentially provide many features from the healthcare data ecosystem to achieve even better performance.
ST: What are your views on our initiative, Techvik?
MPL: I recently came across it and I am very impressed by the work you are doing. I always support initiatives that could provide a platform for others and help them share knowledge with others. So, congratulations on your achievements, and I really hope to see more of you in the future!
ST: This brings me to the last question of this interview
ST: Sir, any final words for our reader?
MPL: I think it is very important to keep learning about new things, utilizing the vast resources available online, through various institutions and individuals, and using that knowledge for the betterment of the world around you.
ST: Thank you so much once again, sir, it was an honor speaking to you!
MLP: No problem, I really enjoyed connecting with you and all the very best for your future!
ST: I will end this interview with a picture of ours.
MLP: Sure, I would love that.
Thank you so much, Prof. Matthew P. Lungren!