The Longitudinal Experience of Working With AI: Einstein Healthcare Network’s Radiology Department Experience
What has the experience been for a radiology department at an academic medical center, in committing to being on the leading edge in applying artificial intelligence (AI) strategies and techniques to day-to-day radiological practice? On Monday of RSNA19, being held at Chicago’s McCormick Place Convention Center, and sponsored by the Oak Brook, Ill.-based Radiological Society of North America, the chairman of that department shared his perspectives.
Terence Matalon, M.D., chairman of the Department of Radiology at Einstein Healthcare Network in Philadelphia, shared his perspectives on the experience during a lunch-and-learn sponsored by the Lexington, Mass.-based FUJIFILM Medical Systems. Dr. Matalon was introduced by Bill Lacy, vice president, medical informatics, FUJIFILM Medical Systems USA. Lacy gave the audience an overview of FUJIFILM’s current evolution in the AI space, noting that earlier work in the machine learning space started as far back as the 1980s. The company, he said, has been putting considerable energy into its AI division, called REiLI (pronounced “RAY-lee”).
“Among the many use cases” for AI, Lacy said, “is the example of worklist priority notification. How can we take the worklist and re-prioritize it, based on AI results?” he asked. “How can I communicate directly with the physicians who must most urgently be communicated with? How do we incorporate AI findings into the report? How do we do broader analysis of archived reports? The ecosystem of an AI platform will include third-party vendors, engines from the host like Fuji, etc.,” he said. What remains to be seen is how all the various elements will be integrated. His company is well-positioned to act as an integrating platform, he added. At the end of the program, Steve Worrell, CEO of the Miamisburg, Oh.-based Riverain Technologies, also spoke.
After FUJIFILM’s Lacy presented his opening remarks, Dr. Matalon gave a presentation on the experience to date of the radiology department at Einstein, with its AI tools and platform, with the platform being supplied by FUJIFILM and some of the specific tools provided by Riverain Technologies and by the Kibbutz Shefayim, Israel- and Lakeway, Tex-based Zebra Medical Vision.
“Einstein is a classic urban academic medical center; we serve a very deprived population,” Dr. Matalon noted. “And for me, it has been as important to provide a perception of innovation,” as the innovation itself, in order to engage and retain radiologists. In that, he added, “My colleagues feel more comfortable” since the inception of AI-facilitated work, “that we are innovative in terms of doing the most we can in terms of extracting the most information from the images that we can. I use AI to some degree to deal with burnout,” he said bluntly. “We certainly have seen significant improvements and changes in the way that we do things, as a result of implementing AI.”
One of the most successful use cases so far, Dr. Matalon told the audience, has been around the reworking of the radiologist worklist for chest x-rays. “Normally in the past,” he said, “we had only two designations: STAT and routine.” Now, with the use of an AI-driven algorithm, some cases have had their prioritization changed to “P,” for “priority, when the algorithm has identified a potential case of pneumothorax (collapsed lung). “We had a very bad outcome a number of years ago due to a pneumothorax obtained via x-ray late at night, and which wasn’t identified until the patient had suffered harm.” The implementation of this algorithm, he said, “has definitely changed the order.” Einstein has been using algorithms in that clinical area provided by Zebra Medical Vision. One point that Matalon made is that there are many situations in which “Radiologists may not be working in a 24/7 environment, and there are environments without radiologists.”
With regard to integrating all the tools from Zebra, Riverain, and an additional vendor, the platform element is important, Matalon said. “Each of these tools or algorithms from our different AI vendors involves widgets, and integrations with our PACS [picture archiving and communications system] system, that allow for individual integration and vetting by the radiologist of whether a finding is true or not true. A success finding translates into PowerScribe and provides information that would otherwise have to be manually entered by the radiologist. We now have four widgets on the radiologist desktop. And we’re able to integrate SR findings directly into worklists,” he added, referring to the DICOM structured report format, for long nodule detection, intercranial hemorrhage, and pneumothorax. Further, he added, the department employs an overlay GSPS integration (DICOM Grayscale Softcopy Presentation State) of findings, so that “Findings can be imported directly into the report. And this process or algorithm will automatically look at the prior report and examine findings.”
Benefits of all these tools have been obvious to the practicing radiologists, Matalon noted. For example, he said, “Riverain has a unique technology that suppresses [lung] vessels, and allows lung nodules to be much more easily identified,” leading to improved accuracy. “Our radiologists feel much more comfortable now that no lung modules have been detected.”
Further, the prioritization involved means that “A substantial number of acute findings for intercranial and pneumothorax, are being addressed much more quickly than before prioritization.” And, he added, “Findings are automated into the report, which reduces radiologist burnout” from so much manual inputting of content.
Inevitably, Matalon told his audience, there is some initial loss of efficiency as a system like this is first implemented. “But,” he said, “I think now that there’s no question that there are opportunities to improve efficiency and reduce work.”
A remaining concern, he said, is increased cost. But, he predicted, over time, the value of AI tools will be proven, and hospital administrations will be more willing to invest in them in order to improve their radiologists’ efficiency. “I would encourage people to deploy AI on a trial basis to begin with,” he added, “because your results may not actually match what the vendor has indicated that other sites are seeing. And we did see a number of people fearful of deploying AI,” he reported. “I had no fear that AI would supplant a radiologist. But sticking your head in the sand and thinking that AI won’t come along, is counterproductive. And if you don’t take advantage of all resources, you run the risk of supplanting yourself.”
Indeed, Matalon said, “If I tried to remove AI, I would have a revolt from my radiologists. They feel significantly more confident that no lung nodules have been left unidentified,” he attested. Even so, he acknowledged, there remains a risk of false negatives and positives. AI will never replace clinical judgment. Among the benefits he sees: image optimization and off-axis reconstruction; protocol optimization; triaging in order to better serve underserved population; augmented detection; and the use of natural language processing-based AI tools for follow-up and for the improvement of report generation.
“We are at the transition between multiple separate vendors, and integration,” Matalon concluded. “There are at least seven or eight AI vendors that are presenting at this meeting. Having a platform that makes this comfortable to the physicians is important, rather than having an alternate viewer. I think it’s going to be an exciting time for us in the next five to ten years. What we have now is so different from three months ago and six months ago. We had had Riverain for quite some time—in three years, there are have been so many iterations and improvements in that one product.”