Facing a backlog of work during non-core working hours is nothing new to most radiologists. Particularly when working evening or overnight shifts, cases can pile up with no way of knowing which are the most critical. But now, radiologists can get assistance in the form of AI algorithms that, when combined with workflow orchestration software, can help triage cases, moving studies with potentially critical findings to the top of the worklist and improving patient care. At Lahey Hospital & Medical Center in Burlington, Massachusetts, Christoph Wald, MD, PhD, MBA, FACR, chair of Lahey’s radiology department and chair of the ACR Commission on Informatics, and his colleagues from the neuroimaging and emergency radiology departments led the way on deploying this type of findings-detection software to improve their workflow and, in general, alleviate the pressure on radiologists. During a recent interview with the Bulletin, Wald discussed the importance of integrating AI algorithms alongside workflow-orchestration software to help radiologists prioritize imaging findings and how, despite popular media narratives, AI isn’t poised to take radiologists’ jobs anytime soon.
How has a suite of AI algorithms coupled with new workflow orchestration software helped your group move images with critical findings to the top of the worklist?
The first set of FDA-cleared algorithms that we implemented focus on computer-aided triage (CADt). This iteration of technological support determined whether a finding was present or absent for the purpose of prioritizing work, so that radiologists could then triage the cases to be interpreted by likely urgency. It is important to emphasize that the CADt AI wasn’t FDA-cleared for making the primary and final diagnosis, and radiologists must take care to not use AI to automate making a diagnosis.
That original approach has actually held up quite nicely during times of the week when there is a mismatch between studies to be read and radiologist capacity to interpret them. Specifically at nighttime when the ED is busy and we’ve got very few radiologists in the department, it becomes important to leverage this technology to help the radiologists direct their attention to the most urgent studies where a finding might be present.
These algorithms were cleared for triage, but they accomplish this through carrying out a narrow diagnostic task. The radiologist essentially gains a “look over the shoulder.” There are certainly some instances when we as radiologists have not detected very subtle findings, for example a small peripheral segmental pulmonary embolus, which the software may detect. However, that doesn’t mean the algorithm is anywhere near ready to replace a radiologist.
And of course, sometimes it goes the other way, where we humans detect something that the software didn’t see. The software might flag something that it was not trained on, mistaking it for a real finding, and we determine that it’s an artifact. So, there’s a little bit of a back and forth where the radiologist always determines whether or not to include an AI’s finding in the report. We prefer not so much to compare radiologist performance to that of the AI, but rather compare the performance of the radiologist in collaboration with AI against the radiologist alone.
Media portrayals often show AI being poised to take radiologists’ jobs. Can you address this perception?
In our practice, we don’t see AI as a threat. On the contrary, it assists our radiologists in prioritizing their work more appropriately. We anticipate that the intensity of imaging utilization will only rise in the future as the so-called baby boomers age into the high-consumption years of medical care.
At the same time, there are extreme cost pressures. For example, CMS has continually decreased payments to radiologists in the recent past, and there seems to be no end in sight for this trend. Added to this is the fact that, in recent years, we haven’t really changed the rate at which we graduate radiology residents. The bottom line is that, most likely, radiologists will be asked to continue doing more work for less pay; consequently, using AI to lessen certain aspects of the work burden and make radiologists more efficient will be welcomed by radiology.
We must also not forget that AI is increasingly used to perform routine quantitative analyses on patient studies — brain thickness mapping and emphysema quantification being two examples — which we cannot easily perform ourselves during conventional image interpretation. Inclusion of the derived quantitative maps and outputs into the radiology report makes our work product more valuable to our clinical colleagues and patients.
Can you tell us about the process of teaming up with a vendor?
For the vast majority of radiology groups, including my own department, making our own AI is beyond the scope of what we do. We just don’t have the core competency in house. Clinical departments in this circumstance need to familiarize themselves with an ever-growing number of commercial AI offerings. The ACR has created a resource called AI Central to help members navigate this growing space (see sidebar). There were about 160 commercially available algorithms relevant to radiology listed on AI Central as of March 2022, and this number is rapidly growing.
As a first step for someone looking to familiarize themselves with this process, I would encourage them to speak with colleagues who have already put these systems in place. In addition, they should reach out to the vendors themselves and inquire about the architecture because vendors can share the standalone performance testing of their software. This will give interested parties an idea of the performance of the product under the conditions that led to its FDA clearance.
In our practice, we don't see AI as a threat. On the contrary, it assists our radiologists in prioritizing their work more appropriately.
Furthermore, anyone looking to work with a vendor in this space should find out what your IT security department is comfortable with implementing. Some can live entirely behind your institution’s firewall. Others employ a local server that sends de-identified information to the cloud for processing. And some vendors will give the customer a choice.
One way to make reviewing, trialing, and ultimately implementing more than one AI algorithm easier is by contracting with so-called platform companies or AI marketplaces. Contracting with a marketplace vendor provides a single platform where you can more or less turn on or off algorithms without having to write individual, complex contracts and stand up separate servers and solutions.
In 2018, your team asked for referring provider input when it came to deciding which algorithms would be incorporated into the workflow. Can you discuss why you brought referrers into that decision-making process?
Imaging departments should gain an understanding of what their clinical customers need and expect from them as radiologists. Right around the time we began exploring the use of AI, Lahey was transitioning to becoming a comprehensive stroke center. In gearing up for this change, our department looked at best practices and noticed that the majority of stroke revascularization trials had incorporated a particular AI which claimed to be able to distinguish dead brain tissue from recoverable brain tissue.
Since the technology also included a perfusion analysis and had been successfully included in multicenter trials, this solution seemed like a good candidate. Proof of its efficacy from the trials proved especially attractive to our referring clinician colleagues, so we certainly wanted to follow that precedent.
Fortunately, our practice had established itself at our institution as the go-to division for undertaking such technological projects. We’d already laid the groundwork for being recognized as honest partners who would work well with the clinical team and address their needs. It’s important for the decision-makers in our institutions — whether that be the C-suite or others — to recognize radiologists in this role. The last thing you want is a clinical service other than radiology independently contracting for an AI installation without your knowledge.
What role do you see AI playing in radiology in the next 10 years, and how should radiologists be preparing for the future?
Ignoring AI integration is probably not a good long-term strategy for radiologists. Rather, I think radiologists should seek out every opportunity where AI either enhances their work product or makes them more efficient — or both. I see it impacting the entire value chain of radiology, including how we manage our practices.
Radiology as the first truly digital specialty lends itself to AI applications. For instance, AI will most likely influence how we schedule patients, how we predict missed-care opportunities, and how we schedule our workforce to match imaging volume on a given day. In addition, clever software is being introduced that can make similar or better-quality images based on lower doses when it comes to CT or limited case-based sampling and MR. There are also exciting new developments when it comes to speeding up image acquisition in ways that were never possible before. We’re also seeing the emergence of opportunistic screening, where AI extracts features and information from imaging studies done for another reason. The purpose behind this is to make risk predictions for certain diseases in patients.
At the end of the day, the future of AI in radiology will involve unlocking more value from what we do. I think that the combination of radiologists plus AI is most likely to win the day in the long run.
To view this podcast episode and others, please visit acr.org/bulletin-podcast.