The Future of Collaborative AI with Keynote Speaker, Paul Thompson, PhD

Can you tell our readers a bit about yourself and your current roles? 

My background is a bit unusual. I did an undergraduate degree in classical languages, which were Greek and Latin, but I was interested in studying medical images of the brain. So, I earned a PhD in neuroscience. Currently, I am Director of the ENIGMA Consortium and  

Associate Director of the University of Southern California (USC) Mark and Mary Stevens Neuroimaging and Informatics Institute. 

What are the main goals of the Mark and Mary Stevens Neuroimaging and Informatics Institute and ENIGMA?  

The main goal of the institute is to really understand human brain diseases – how to better diagnose them and how to better treat them. We can do this with the help of various neuroimaging techniques, which might involve functional mapping of the brain, looking at its activity, the blood flow and other techniques. 

Now, ENIGMA stands for “Enhancing Neuroimaging and Genetics through Meta-Analysis.” We study about 30 different brain diseases, like Alzheimer's, Parkinson's, stroke and epilepsy –  the neurological ones. Then there are others where there have not been a lot of large-scale studies. These include eating disorders in young kids, like anorexia or bulimia, where we are asking, does your brain recover if you get treatment? So what ENIGMA does is bring scientists together from across 45 countries to pool their data expertise and their clinical knowledge of these conditions. It usually starts with someone having a good idea for a question about the brain or a brain disease. Then the scientists can ask one another, “Who has data on this?” or “Does anyone have MRI or diffusion imaging or function maps of the brain for this disease?” They might get 10,000 scans from maybe 20 different countries.

It can help scientists look for patterns. If the pattern that they see in the brain is consistent across the sites, they would say, “Well, this is how the disease looks in general.” Or if it's not consistent, it may be that a disease has a very different expression. In China or Africa, there’s a lot of interest in whether these diagnostic tests will be universal or whether they should be customized. 

How has the ENIGMA Consortium worked to incorporate AI into its research? 

That's been a big change in the last couple of years. Before, we did not use AI; in fact, AI wasn't used in any of the studies that ENIGMA did. Instead, it was more of a standard approach, where a computer program would measure the brain’s function, molecular content and the level of pathology. 

Recently, we won a grant called “AI for AD”. AD stands for Alzheimer's disease. The idea was to see if some of the new AI methods could combine brain scans and diagnose a disease based on those scans. The AI program would learn from tens of thousands of scans to see if it could identify a consistent pattern. Obviously, it scans much faster than humans and can screen through different hypotheses about what the disease may be. This is considered a diagnostic method. 

The second use for AI is called a vision language model. The model reads a brain MRI  scan and tells you what the disease is in spoken language. It's just like if you were using ChatGPT.  Now you could say, “Couldn’t I just use ChatGPT?” or just insert the medical scan in? Well, you could, but it wouldn't have learned what the features are that are useful for medical conditions. 

This method would be very helpful when there is a massive overload of scans that need to be read, or if there is an emergency and it's important that those scans are read quickly. It can get to that scan and flag it as critical and then give a diagnosis or recommend the best treatment. Its confidence is based on how often it's gotten equivalent diagnoses right in the past. With a human in the loop, they can see whether or not its decisions were accurate. The vision for AI to analyze the scan for features that it considers to be diagnostic, then the language part can hedge it a bit; it can say a very helpful tool. 

What about potential incorporation of widespread AI excites you the most? 

A cool new usage of AI involves improving medical imaging, where AI can enhance the quality of brain scans. To give an example, brain scans take a while to collect, but people don't like sitting in the MRI machine for long periods of time and that can affect the quality of the image. So AI is used for “super resolution” to enhance the detail of a brain MRI. It can help us see a condition that was affecting the blood flow to the brain or neural pathways to help conclusively decide what the abnormality is. 

Another project in the lab that's using AI to do something that we cannot do at the moment involves Alzheimer's. What causes Alzheimer’s is amyloid plaque buildup in the brain, but we cannot see that on an MRI since it builds up gradually over many years. A PET scan is the only imaging that can detect the plaque, but the problem is that PET scans are not available all over the world. They're not even available in a lot of medical centers. PET scans are also very expensive, invasive and expose the patient to radiation. 

So, what the AI is doing is taking data from people who have had a brain PET scan and an MRI scan, and it's trying to determine if anything on the cheaper MRI scan could help make a map of these plaques. This is called image enhancement or image synthesis. In this case, it generates a new image, so we are using some of the techniques of generative AI. We are going to provide a huge amount of training data to get these AI methods to lock onto helpful features, and then it can make essentially new images. 

What can attendees expect to hear from you at Emory’s Third Annual AI.Health Symposium? 

We’re in a really exciting time with AI and medicine, but people have concerns about it. Such as, “How does this work?” and “Do we understand how this works?” I plan to talk about when AI methods work and when they don't. I also want people to think creatively about what AI can be used for.  Everyone comes to the table with creative ideas of what might be possible. This is why I think you have one of the best – maybe the best – AI institute in the world. The atmosphere here really fosters the growth of good, creative people who are trying to advance medical applications, which is exciting to be a part of.