Physician burnout and retirements, combined with fewer qualified technologists amid increasing imaging volume, are straining the radiology workforce.
Radiology AI is expected to buoy those most under pressure, but there is still a vast array of AI tools and clinical applications to consider. In a quest to solve clinical challenges and change the status quo, many radiology departments have been participating in co-developing, piloting, training, or otherwise evaluating AI tools.
For this special edition, AuntMinnie discusses the current state of AI’s impact on radiology workforce challenges and the technologies that will impact the majority of U.S. radiology practices.
By far, radiologist-facing interpretative AI products outweigh technologist-facing and noninterpretive AI use cases, according to the American College of Radiology (ACR), which maintains an online repository of U.S. Food and Drug Administration (FDA)-cleared AI.
Chair of the ACR Commission on Informatics Christoph Wald, MD, PhD, told AuntMinnie that while current FDA-cleared AI for radiology plays only a small role today in addressing workforce challenges in the U.S., other parts of the world are already exercising their options.
Christoph Wald, MD, PhD.
AI exercises abroad
“AI and use of AI in imaging in radiology is a global phenomenon,” said Wald, who trained in internal medicine and radiology in Scotland and Germany, works as senior associate consultant and professor of radiology at Mayo Clinic, and now also serves as vice chair of the ACR board of chancellors.
“U.S. radiology needs to take a global look at the world and figure out what’s working and not working, how it’s working on the efficiency side, how it’s working on the healthcare side, how it’s useful for patients,” Wald continued. “Where we see a positive signal, we should push to adopt that here, and where we see problems, we should not adopt that here.”
Wald recommended paying attention to certain CE-marked products and large-scale experiments in the U.K., where there is a massive shortage of radiologists.
“They are willing to experiment,” he said. “You are going to see the NHS bundling together 10, 15, 20 trusts. The government in the U.K. is basically trying to figure out can we use this technology to close the gap in healthcare delivery.”
Generative AI for some
In the meantime, in the U.S., we’re seeing a potential canary in the coal mine, an exceptional predictor, with AI-assisted report generation, Wald explained.
In this example, radiologists and the information services department at Northwestern Medicine in Chicago have developed a homegrown AI system using clinical data from within its network. Pneumothorax on chest x-ray served as a promising proof-of-concept target.
The study involved deploying a generative AI system over nearly 24,000 radiology reports during a five-month period in 2024. The group compared draft radiograph report creation times and clinical accuracy for 11,980 with and without the AI tool.
Co-author Samir Abboud, MD, PhD, chief of emergency radiology at Northwestern Medicine and clinical assistant professor of radiology at Northwestern University Feinberg School of Medicine, was among those reporting an average 15.5% boost in radiograph report completion efficiency and possibly greater gains.
Wald cautioned the unknown generalizability of the model to general practice — it was developed and is being operated by a single institution for noncommercial internal use under Institutional Review Board oversight — but acknowledged that the technology may prove to be a “serious efficiency enhancer.”
Institutions and imaging suppliers that own large datasets and ground truth will be the most likely to emulate the development of generative AI like Northwestern’s, he said.
Report technology for others
In contrast, most radiology departments will consider other options to drive radiologist efficiency for now, according to Wald.
“My prediction is that you are going to see further evolution of report drafting technology,” he said, explaining that there are three basic AI report-assist approaches:
- Image or pixel-to-text generators (like the Northwestern approach, not widely available)
- Text-to-text impression generators (commercially available)
- Structured output-to-text generators (i.e., inputs piped into a reporting tool such as radiologist dictation, inputs from knowledge databases, image processing software outputs, DICOM files, and translated into natural language in the style of radiologists, commercially available)
“We’re seeing the convergence of this functionality in the new reporting platforms,” Wald explained. “This means that as an end-user radiologist, I don’t have to license multiple products.”
FDA clearance is not currently needed for “text transformation” tools, which are not classified as software as a medical device (SaMD), he added.
“It is important to start sorting whether you’re dealing with a software that is translating images into text in the form of a report,” Wald continued. “If you’re getting into the business of selecting and then using any kind of image-based AI technology in your radiology department, then you must create processes around proper selection, acceptance testing, implementation, user education, and ongoing monitoring of that software.”
That is where the ACR ARCH-AI and Assess-AI programs and technology can guide and support practices in that journey, according to Wald. This first national AI registry (Assess AI) can be easily added by practices to their existing registry arrangements, he noted.
Not all algorithms work equally well in all departments. Importantly, changing scanners or protocols can impact performance, Wald pointed out, highlighting the need for ongoing performance monitoring.
Performance monitoring
To mitigate AI performance risks, machine-learning scientists are developing quality control mechanisms.
“We are going to increasingly see sophisticated automated monitoring methods, like statistical process control, applied to AI applications in radiology and healthcare in general,” predicted David Larson, MD, co-director of the Stanford Radiology AI Development and Evaluation (AIDE) Lab.
This will include real-time monitoring that will automatically identify problematic cases, as well as background monitoring that will identify data shifts, he told AuntMinnie. It will also enable continuous improvement not only in the AI models but also in the local practice’s image quality and workflow.
David Larson, MD.
Ideally, real-time monitoring will alert the relevant individuals in the process, such as the technologists, physicians, IT team, and model developers, according to the type of problem it detects, Larson explained. Each of these stakeholder groups can also use the aggregated data to improve their portions of the process.
Local practices that adopt this approach will see steady improvements in both workflow efficiency and quality, Larson added.
“All of these quality efforts will enable higher confidence in AI models’ output, which will also improve radiologists’ efficiency,” he noted.
The AIDE Lab team envisions ensembled monitoring models and other applications running in parallel with black-box AI models, to not only improve diagnostic accuracy but supply a “confidence” score for physicians using those apps in their practices.
“These automated ‘second opinion’ models can give the radiologist a sort of ‘traffic light’ output, with green implying greater confidence that the primary model is correct, yellow implying no change in confidence level, and red implying a lower confidence level, meaning they should consider the output with greater skepticism,” Larson said.
Ultimate AI orchestrator
Radiologists and imaging teams know that the modern radiology workspace can be curated with any number of AI tools and systems designed to increase efficiency, relieve cognitive burden, and reduce manual tasks that contribute to human and organizational error.
However, Wald predicts that a primary agentic approach will eventually assert dominance as the ultimate AI workflow and radiology production system orchestrator. The jury is still out on whether that will be realized by PACS companies, reporting companies, RIS companies, or workflow companies, he said, adding that there is hope that emerging open-source interoperability standards like the model context protocol will facilitate that approach when they become more widely adopted.
“I predict we will see companies will show initial takes on designing agentic workflows,” Wald said. “The primary system will try to orchestrate the secondary systems to create an efficient workflow in total, rather than right now using the human radiologist as the integration engine. Today, we’re doing all of this manually, but we’re not that far away from being able to do it automatically.”
The key is interoperability, Wald continued. “Companies that are not able to play in that future state are going to lose their customers.”
Big picture
The looming radiology workforce crisis is on the minds of hospital system and imaging center executives, whether it centers on radiologist burnout, fewer qualified radiologic technologists, 24/7 availability of radiologists, or a need for more radiology subspecialists.
Looking broadly, AI will address the radiology workforce shortage in three primary areas: demand management, capacity building, and workflow efficiency, determined a team from the University of Texas MD Anderson Cancer Center in Houston.
“These three approaches are interdependent,” the researchers, including radiologist Naveen Garg, MD, in the department of abdominal imaging, wrote for a paper published in npj Health Systems. In addition, “increasing radiologists’ capacity for high-value tasks amplifies efficiency gains, enabling them to more effectively meet growing demand,” they added.
Wald emphasized that understanding the opportunity of AI-facilitated efficiencies should be evaluated for the entire radiology practice, rather than the radiologist’s work alone. He said that the most reasonable approach considers the life cycle of the imaging study.
Imaging study life cycle
“The first stop where AI can help is ordering or decision support at order entry,” Wald said. “Decision-support software that helps our referring doctors decide what imaging study to order and makes it easier for the ordering doctor to pick the right test and reduce inappropriate imaging at the same time. EHR and decision support companies are trying to use AI to facilitate that transaction.”
AI also presents an opportunity to optimize imaging assets and image acquisition workflow management, compensate for patient no-shows, and leverage AI-facilitated workload balancing to leave no exam slots empty, Wald said.
This is important because radiology services tend to be on the higher end of the cost spectrum and are vulnerable to greater levels of uncaptured revenue per patient visit than primary care when patients don’t show up for appointments.
Even in cases where an appointment is filled with an inpatient or a standby patient, there are sometimes lost costs due to staff preparatory work such as insurance authorization, scheduling time, and protocoling.
Forecasting patient attendance and a system enabled to send targeted reminders can reduce no-show rates, the Garg group noted. In addition, dynamic schedule adjustments combined with predictive analytics tools to forecast demand peaks would better distribute workloads, they added.
Discuss strategies
Workflow efficiencies are appealing to high-performing medical imaging departments, regardless of imaging and patient volume.
The variety of AI use cases, available through the ACR for example, can help imaging department leaders and their stakeholders discuss strategies that would be best suited to their organizational needs.
As radiology leaders continue to warn of the dangers of operational inefficiencies and understaffing, it will be incumbent on hospital, health system, and imaging center executives to begin these discussions sooner rather than later.
Regardless of workforce challenges, greater radiology workflow efficiency could reduce the cost of medical imaging overall. Armed with enough information and proper safeguards, imaging leaders and hospital administrators will be better able to discern the roles that AI will play.