Industry Sage Joe Marion: We Have Yet to See Significant ROI on AI in Radiology
In an article published on December 10, just days after the conclusion of the annual RSNA Conference at Chicago’s McCormick Place Convention Center, and sponsored as always by the Oak Brook, Ill.-based Radiological Society of North America, Siddharth Shah and Robin Joffe wrote in Diagnostic Imaging online, “The dust is now beginning to settle on the idea of AI in imaging—an idea that once took the industry by storm. The questions being asked by radiologists have now changed from “will it replace me?” to “how can it help me?”—and rightly so. AI continues to make significant progress in the field of diagnostic imaging, as can be gauged using the recently concluded Radiological Society of North America’s (RSNA) Annual Meeting in Chicago as a barometer.” What’s more, the authors noted, “Last year, there were 49 exhibitors at RSNA tagged as machine learning companies, and 22 of those were first time exhibitors. This year, the same number more than doubled to 104, 25 of which were first-time exhibitors. More importantly, the incumbents of medical imaging equipment have also made notable AI efforts, with each one launching new or enhanced AI capabilities. Clearly, AI was one of the key themes at RSNA.”
The writers also referenced a recent Frost & Sullivan that found that, “of the 114 startups active in the AI for medical imaging space, a significant majority target the image analysis aspect of medical imaging. Identifying and analyzing specific features in an image form the crux of a radiologists’ job, and since they base their findings on this analysis, it forms the most important clinical step in the imaging workflow. The startup disruption rampant in overall healthcare has focused on image analysis in the case of medical imaging AI.” There are artificial intelligence applications being launched “for triaging, worklist assignment, and workflow orchestration,” available from a growing number of vendors.
One of those who has a uniquely authoritative perspective on all this is Joe Marion, a principal in the Waukesha, Wis.-based Healthcare Integration Strategies LLC, has participated in 42 RSNA Conferences—probably among the most of any current attendee. No one has a broader perspective on the imaging informatics vendor market than Marion, who spent years on the vendor side before shifting over to consulting a number of years ago.
Healthcare Informatics Editor-in-Chief Mark Hagland spoke recently with Marion on his perspectives around AI and other key subjects in imaging informatics, at the beginning of 2019. Below are excerpts from that interview.
Here we are in January 2019; at this moment in time in the evolution of the industry, what’s your 40,000-feet-up view of imaging informatics right now?
The interest-grabber, if you will, that gets all the attention, has been artificial intelligence. And the question is where that’s going. Everybody has their placeholder in the context of a platform for AI application development. That probably gets the most attention right now. The reality is that the number of real practical applications that people are using and implementing remains really pretty small. So it’s more hype than reality. The next area that’s important is workflow orchestration. The driver in that context, if you talk with some of the key players, is largely the fact that there’s a lot of consolidation going on, so the intent of a lot of that consolidation is going to be broader reading capabilities.
Just look at Aurora and Advocate as an example. Part of the intent of any merger will be a broader radiologist base, to apply to my facilities, so that if a radiologist is on vacation at one facility, I can pick up the load at another facility. So I think the best way to describe it is intelligent distribution of studies to the available pool of radiologists.
Do you think that a lot of people are too distracted by shiny objects?
To some degree, yes. I think that’s the whole issue with AI—where’s the ROI? Show me that that’s really going to help. The more likely opportunity for AI in that regard is in the context of the shift towards value-based care. In essence, if I could intelligently analyze studies and suggest or route based on the consequences, that could make a difference. Let’s say if it were a complex case, maybe there’s some AI-based analysis to say, it makes more sense to route this to a specialist, as opposed to a generalist radiologist reading it first.
And another twist to that, a number of companies including IBM, have been vocal about it, is the notion of supportive information—the classic example that Change Healthcare always used was, a radiologist was going to review a case and there are spots on the liver, and his first instinct is that it’s a cyst, but he’s presented with the history of the patient and lab information and other information, and all that is made available, and they can use that information to help guide them, and it might change their initial impression that it’s a cyst, to that it might instead be a tumor. It’s in essence improving their read by presenting them with more information than they would normally have. And that’s why I think that workflow orchestration actually offers more short-term benefit than to focus on something described as artificial intelligence.
Is there still work to be done on standards in imaging informatics?
Yes, there’s still work to be done. There’s still a fair amount of confusion. If you look at this XDS, cross-document sharing, if that’s the solution or future, there’s an extension of that called XDSI for imaging. Some people will say, yes, that’s great, but you’ll have to have devices that are XDSI-compliant; so building a repository that’s XDSI-compliant is one thing; but leveraging devices that are XDSI-compliant is another thing. So it’s the early days of the ACR-NEMA imaging format standard, which evolved into DICOM. But in the earliest days of that, it was sort of like the same landscape today for all of this. Yes, there was a standard for communicating, say, CT images. But the CT scanner had to support the DICOM standard, in order to get the images out of the scanner. Conversely, if the scanner was DICOM-compliant, but the viewers weren’t, what was the point of having the DICOM standard?
What do you think will happen this year, in imaging informatics?
I think it’s sort of business as usual. I just saw something about whether PACS deconstruction is real. First of all, a lot of that is terminology and phraseology. It’s hard to sort that out. But I think that there is some question in the context of where everything is headed. And I think the full-service companies are probably still in denial and want to sell you the concept of one throat to choke; whereas the company with the most success in that regard is Visage. There’s just an announcement out that they’ve closed a deal with Partners in Massachusetts. And when you look at the number of entities they’ve picked up—Shands Hospital in Gainesville, Mayo Clinic, Mercy in Iowa, many organizations that they’ve had success with, in saying, we want to be your front-end PACS [picture archiving and communications system] vendor of choice. And they work with a number of different solutions. This will all have to play out over time. But this year will definitely be interesting in this entire area.
