Marwan Sati is a distinguished engineer leading AI development in medical imaging and natural language processing for Merge by Merative
The promise of AI to aid radiologists in improving disease detection has blinded us to the many other benefits AI can bring to improve radiology workflow and efficiency. Consider the many workflow steps that comprise radiology: image ordering, scheduling, acquisition, worklist distribution, reading, reporting, patient follow-ups. Each of these steps is an opportunity for targeted AI applications to help improve radiologist efficiency and reduce errors. It’s also an opportunity to create a radiology ecosystem where every workflow step benefits from new insights and efficiencies that AI can bring to the table.
I think of AI adoption in healthcare as following a similar path of AI in the auto industry. The road to fully driverless cars is a long way out, but progress is being made. The adoption of automated features in automobiles has been gradual and specific: blind spot detection, lane detection, front collision detection. These innovations build on each other and pave the way to bigger AI adoptions. Now imagine this same approach in radiology: a series of targeted applications that leverage AI to augment the operator.
This kind of intelligent imaging ecosystem, built on a foundation of focused AI applications, is what will help make radiologists more efficient and drive better outcomes for both clinicians and patients – and not down the road, but right now.
In this piece, I’ll provide a few examples of how AI can help improve radiology workflow efficiency and quality of care.
1. Intelligent worklist distribution
Many PACS rely on individual radiologists’ self-selection or rules-based approaches for assigning imaging studies for review. This leads to variable workflow distribution – in sites where compensation has an RVU component, some radiologists may “cherry pick” studies, leaving other radiologists frustrated. This can also result in uneven image volumes between radiologists, exacerbating burnout and compromising timely, accurate delivery for patient care. There is also room for improvement to optimize which studies go to which radiologists based on their specialty and personal preferences.
An “intelligent worklist” approach that integrates AI can significantly reduce the cherry picking of studies and reduce the friction that comes from uneven workload distribution. Like how Netflix leverages your preferred shows in suggesting new ones, AI can consider radiologist preferences as well as many other parameters, such as fair RVU distribution, radiologist specialty, etc. In a recent internal study, we found that this intelligent worklist approach resulted in a 32% more equitable distribution of studies across radiologists. By dynamically balancing studies based on priority and value, radiologist preference and specialty, and turnaround time, an intelligent worklist can help ensure studies are matched to the radiologists best equipped – in terms of skill set, time available, and preference – to review them, and return speedy interpretations to the patient.
AI can also help to automatically identify relevant prior information to help radiologists read a current study, quickly sifting through dozens of reports on a patient’s complex medical history and summarizing the most relevant information for the clinician.
2. Automated exam follow-up detection
According to the American College of Radiology, up to 10% of radiology reports contain follow-up recommendations – but approximately half of these follow-ups are never acted on. More worryingly, lung nodules represent about half of these missed follow-ups. In fact, for patients with incidentally detected pulmonary nodules, more than 60% do not receive the appropriate follow-up care.
This is an area ripe for improvement – and AI can help fill this gap. Natural language processing tools, for example, can be used to identify findings in prior patient imaging reports that call for follow-ups. That information can then be automatically cross-referenced with the physician’s scheduling system to determine if the patient has been contacted for a follow-up – and if not, can then automatically begin scheduling that follow-up appointment.
This is a relatively narrow and low-touch AI application but one that can carry enormous potential implications for patient care.
3. Reading and reporting
There is a great opportunity for AI to increase efficiency in reading and reporting. Radiology AI solutions have been mostly focused on prioritizing studies in the worklist or flagging urgent studies. There is, however, a great potential to integrate AI results more deeply into radiology viewer and reporting systems workflow. AI findings such as measurements can appear as DICOM image annotations within the viewer. Radiologists can quickly confirm or deny these results and with a single click automatically populate structured reports with the quantitative data. This takes away a lot of the grunt work of tasks like making measurements or tracking lesions over time to assess disease progression. A study we performed with deeply integrated lung nodule detection AI showed a 23% improvement in efficiency over current manual methods for patients with four or more lung nodules.
AI also has the potential to help radiologists follow reporting guidelines and has been shown to do a good job of automating radiology report impressions.
4. Peer review
Something the ACR has been advocating for in recent years is finding a better way for radiologists to conduct, and benefit from, peer reviews of their studies. Nearly 800,000 Americans every year either die or suffer permanent disability due to misdiagnosis. Radiologists review upwards of 800 million images per year. Despite that massive workload the error rate in radiology is only 3-5%. But 3% of 800 million still constitutes millions of studies with potential misdiagnoses or missed findings.
This is an opportunity where AI can help with more efficient peer review. AI-assisted solutions can help identify missed findings retrospectively by comparing what AI found in the image to what was in the report. This can create a more guided peer review process that starts with cases that are more likely to have errors. AI can also be trained on guidelines to help identify reports that are out of compliance with those guidelines. All of this would help improve quality with a more efficient AI-assisted QC process.
Building an intelligent, futureproofed imaging ecosystem
Once upon a time there may have been a fear from some about what AI might mean for radiologists’ job security. Those days are long gone, though. Between aging populations, complexity of diseases, and a healthcare landscape that is still catching up with medical procedures that were deferred during COVID-19, radiologists have never had more work – and never needed more help in doing it.
AI can be a valuable tool in delivering patient care faster and more accurately, but always in a way that is complementary to the radiologist’s judgment and expertise, not in lieu of it. And it doesn’t require radically rebuilding our industry around AI, either.
The four examples above show how targeted AI applications can reshape radiologist workflows into more resilient, intelligent imaging ecosystems. And these are just scratching the surface of what AI can do for radiology and for patients.
This fascinating topic of AI within radiology, along with others will be discussed at the annual Ai-Med Global Summit, scheduled for May 29-31 2024 in Orlando. Book your place now!