ASH Preview: LymphoVision

I'm finally getting around to looking at ASH abstracts.

For those of you who are new around here, ASH stands for the American Society for Hematology, the organization for doctors who specialize in blood disorders, including blood cancers. Every year in early December, they hold their annual meeting, and it's where some of the biggest news about Follicular Lymphoma gets shared. About a month or so before they meeting, they publish the abstracts -- the summaries of the research that will be presented at the meeting. If you're interested in reading on your own, you can find the abstracts here. You can enter a search term like "Follicular Lymphoma" or the name of a particular treatment, and you'll get the list of every abstract that deals with that thing.

Every year, I like to do a little ASH preview, and talk about some of the abstracts that I'm interested in. To be clear -- these aren't necessarily the most important abstracts (though I do try to guess what those will be, which is usually easy because of all the talk online about them). Usually it's just something that I find interesting. Like, this year, there's some research being presented about the long term success of Zevalin, an ImmunoRadioTherapy that I first wrote about in 2009. If you search for "Zevalin" in the blog, you'll see that I wrote about a whole bunch. But it never really gained acceptance as a treatment, for a specific reason. I'll write more about all of that soon when I do a preview of that ASH presentation. 

My point is, I usually write about things that interest me, not necessarily the things that experts find important. Some medical conferences end up featuring some groundbreaking research that really changes things for us as patients. But most of the time, it's the kind of small, incremental moves forward that are typical of cancer research. Science tends to move in inches (or centimeters) over many years, rather than in miles (or kilometers) all at once. 

At a meeting like ASH, that means lots of new research about treatments that are already in use, and maybe a few reports from phase 1 or phase 2 clinical trials that are very hopeful but still a long way from showing up in the treatment room. I read an article a couple of days ago where someone gave an ASH preview, looking at research for a whole bunch of blood cancers, not just FL. He says there is lots of research about "rival BTK degraders," as well as bispecifics and CAR-T treatments. Not necessarily big news about something new. But lots of Presentations showing that one company's version of a treatment might be a little better than someone else's. That's not a bad thing. Keep moving forward, inch by inch, centimeter by centimeter.

So all of that brings us to the first abstract that I want to talk about: "#117: Lymphovision: A lymphoma-specialized foundation model for histology-based lymphoma classification and subtyping."

I'll be honest -- I went through all of the abstracts and marked a bunch to look at more closely later on, and the only reason I marked this one was because the name "LymphoVision" made me laugh. It's definitely something that a hero from the Marvel Universe would have as a super power. Lympho Bob loves Lympho Anything. 

But it's actually a very cool presentation.

The research looks at Image Foundation Models, a type of huge AI-powered database. For cancer research, Image Foundation Models can be trained to look at tissue samples and determine if one is cancerous. They aren't used much with Lymphoma cells yet, but in research on other cancers, Foundation Models have actually been able to look at a sample that seems normal but that they accurately predict will mutate and become cancerous in the future. The researchers for this abstract have developed a Foundation Model that looks specifically as Lymphoma samples -- and they call it LymphoVision.

(They actually call it Lymphovision, without the upper case V, but I think it looks cooler that way.)  

For their research, they used 37 million tissue samples to train the program to recognize Lymphoma. They asked the program to perform several specific tasks, including finding FL in the sample and then classifying it as either grade 1-2 or grade 3. For that particular task, the program found 660 cases of FL -- 500 cases of grade 1-2 and 166 cases of grade 3, for an accuracy of 89.2%. When asked to identify a range of different types of Lymphoma, it had "high diagnostic accuracy" (you can see the results in the abstract). If you're wondering how all of that compares to humans, the abstract says "Based on historical
benchmarks from the 1997 International Lymphoma Study Group classification project, expert
hematopathologists achieved diagnostic agreement rates of 55% to 84%" using the same methods. 

The researchers know that more work on this is necessary, but they see a lot of promise in using tools like this for diagnosis in the future.

I don't know how much you all deal with Artificial Intelligence, but I think there is a general ambivalence about it -- one that I share. Research like this is wonderful, and shows how an AI program can save time and cut down on human error. But the technology is still in its very early stages and make lots of mistakes. That 89.2% accuracy still leaves 1 out of 10 patients getting an incorrect diagnosis. So it's not perfect. 

But I do agree that it holds some promise, and it will likely get better over time. (I'm in that weird space where I want some AI to improve quickly to make my health better, but other AI to improve slowly so I can hang on to my job for a few more years until I'm ready to retire.)

More ASH previews to come soon.