Where Rubber Meets the Road
By Kathy Hardy
Vol. 21 No. 1 P. 10
AI is making inroads in radiology, but opening new avenues for use will require much more work.
As radiologists face increased demand for precision medicine and complex analysis, they’re looking for tools to assist them in meeting these needs. This is where AI is coming into its own—as a sidekick, not a replacement, for radiologists.
AI-assisted workflow is emerging. When looking to the next phase of AI adoption, decision support is an area where machine learning can bring clinical benefits to radiology. AI is finding a place in decision support as a prioritization tool for detecting urgent cases faster, fine-tuning where physicians should focus immediate medical attention.
Another clinical use of AI that is showing early promise is in improving efficiency and quality in image acquisition.
“Application of deep learning in image acquisition is one of the early practical examples of AI that we’re seeing,” says Paul Chang, MD, FSIIM, a professor of radiology and vice chair of radiology informatics at University of Chicago Medicine. “Improving image acquisition throughput by using AI to improve the efficiency of image reconstruction is a palatable use case to justify to decision makers.”
AI in Clinical Settings
In addition to image acquisition, Chang sees AI making inroads in triage settings. In the case of emergency medicine, current algorithms, while not perfect, support the use of AI to make quick decisions regarding which findings need immediate medical attention, he says. The decisions involve identifying where the most serious problems exist, not about coming up with a diagnosis.
“We’re currently seeing AI used in applications that don’t put a lot of stress on the algorithms,” Chang says. “That’s why we’re seeing it used in triage situations and not for diagnosis. We can use AI for prioritization of emergent findings of concern; these use cases conveniently avoid the issue that we still don’t have enough rigorous clinical evidence to confidently use AI for diagnosis.”
In addition, increased knowledge of how AI algorithms reach conclusions will be vital to any FDA approvals for diagnostic solutions, adds Bradley Erickson, MD, PhD, a radiologist at the Mayo Clinic in Rochester, Minnesota.
“The challenge in using AI for diagnostics is we’re not clear how the system is making the decision,” Erickson says. “This makes the FDA more wary to approve new tools. If you can’t explain how it reaches diagnostic conclusions, you need more data.”
For example, a diagnostic AI system could be keying off unimportant data, affecting the merit of the findings.
“If you don’t understand what the system is using to get results, it’s difficult to build confidence in the applicability of the solution,” Erickson says.
A third area where Chang sees AI currently being used is essentially recapitulating the traditional computer-aided detection (CAD) model that many radiologists have used for years. Similar to existing CAD systems, radiologists receive cues from AI algorithms regarding where to view potential findings; the radiologist can ignore or embrace this suggestion.
“This approach requires minimal integration with existing PACS,” he says.
Sharpening the Tool
Industry also sees AI as a tool, rather than a replacement for radiologists. Ariella Shoham, vice president of marketing for Aidoc, sees AI’s role in the clinical setting as “augmenting humans, making them more efficient.”
Aidoc develops health care AI-based decision support software that analyzes medical imaging to aid in detecting acute abnormalities throughout the body. Specifically, Aidoc’s AI technology identifies critical findings in CT scans, prioritizing them in a radiologist’s worklist, and ensuring that patients with the most critical conditions are diagnosed and treated first. Aidoc is PACS agnostic, does not need dedicated hardware, and is both HIPAA and GDPR (General Data Protection Regulation, which requires businesses to protect the personal data and privacy of EU citizens for transactions that occur within EU member states) compliant.
“For an AI solution to bring value, it should be comprehensive and provide value within an existing workflow,” Shoham says. “Beginning three years ago, we saw an acute need in radiology for prioritization of abnormalities found throughout the body and started working to develop a solution for prioritizing and routing cases for radiologists. We focused on usability within an actual work environment.”
Aidoc’s first FDA-cleared algorithm for flagging and prioritizing intracranial hemorrhage on head CTs was approved in August 2018. It was followed by clearance for algorithms identifying cervical spine fractures and pulmonary embolisms. Today, Aidoc is being utilized in more than 250 facilities in the United States.
The focus on CT images comes from the predominant use of CT in emergency medicine, Shoham says. “We wanted to be able to provide value to the broadest set of pathologies,” she explains. “Today, we are able to address close to 80% of the most common acute pathologies in CTs.”
Several car accident victims can arrive in the emergency department (ED) at one time, and the typical protocol is to order CT scans to help diagnose injuries.
“All those images are sent to the PACS at the same time,” Shoham says. “The radiologist on call has no way to know which victim has the most acute injuries without opening the case.”
The company’s solutions analyze medical images directly after they are performed and notify the radiologists of cases with suspected findings directly in their work environment.
“The most urgent case is immediately prioritized,” she says. “Aidoc recommends the exact slice where the medical problem is recognized and flags the abnormality to the radiologist. This process has already shown a reduction in turnaround time of 12% to 14%.”
As the FDA grants more adaptable clearances for new products and solutions, Shoham believes AI will enhance the overall standard of care. In general, more AI solutions are receiving FDA clearance, and more processes are being adapted to accommodate the need for these solutions in the market.
“The FDA is aware of the pace of release of AI solutions and the need to accommodate clearances to this pace but still stay vigilant with how clearances are being provided,” she says.
AI adoption will come but not without challenges. Chang compares the process to driving a race car; AI is the car, data are the “gas” that fuels the car, and the “road” is improved workflow orchestration.
“We need to start drilling for gas and build better roads to support AI,” Chang says. “Our current EMR-centric IT models are inadequate to give us the gas machine learning requires.
“We need better data interoperability,” he adds. “AI applications are currently being driven more by data set availability rather than compelling use case. We will require a significant investment in improving our IT infrastructure to support our appetite for AI and other big data applications.”
When it comes to workflow orchestration, Chang contends that existing PACS and EMR workflow models need to become more “AI aware in real time” in order to support true advances in the use of machine learning in the clinical setting.
“We’re hesitant to change how our PACS work,” Chang says. “Most AI vendors are startups, and PACS and EMR vendors are hesitant to drastically change how they work in order to accommodate these AI offerings. This reluctance results in relatively primitive AI workflow orchestration. We need to go beyond the simple recapitulation of the CAD model. We need deeper AI integration and an AI-aware reengineering of PACS and EMR.”
Users also want to see workflow that operates more like an integrated highway and less like a dead-end street. Glenn Kaplan, MD, vice president of radiology with Nashville-based Envision Physician Services, believes interoperability is a key to success when it comes to AI adoption.
“No one wants to operate with separate workflows,” he says. “That’s being done now in many instances, and it’s just not good for productivity.”
Envision Physician Services employs more than 850 radiologists across the country. In his role with the company, Kaplan travels to sites to help establish best practices and leverage distributed radiology to overcome operational limitations. He has a particular focus on providing new opportunities at smaller facilities and health care systems in rural areas. Kaplan advises on restructuring processes to align technology platforms and integrate unified operating systems. With that, he sees cost as another potential deterrent to adoption, particularly with reimbursements dropping.
“You need to make investments to drive quality, but the technology also needs to be affordable for practices and medical facilities to use,” he says.
Erickson also points to the financial aspect of incorporating AI tools as a potential challenge to widespread adoption.
“How we get paid for this is still a question,” he says. “If using AI reduces acquisition times, and reduces dose exposure for patients, that’s the payback. However, if the radiologist doesn’t get paid for AI, it won’t be adopted.”
Near Future Goals
Kaplan sees collaboration as a key to successful AI adoption down the road. He wonders whether developers will “hoard” data sets to create their own solutions or share data and ideas for further advancements in algorithm development. With involvement from industry organizations such as the ACR and RSNA, he believes data collection will focus more on patient care and less on individual gain.
“With organizations like ACR acting as a repository for data sets, acting as a central broker, the algorithms won’t become commodities,” Kaplan says. “We want everyone to have access to this data.”
Having a central data repository could also help standardize algorithm development. With existing available data being used to train AI algorithms, there’s a risk of underrepresenting certain demographic groups. There can also be an impact on data from confounding medical conditions, as well as technical inconsistencies that can result from different scanning techniques.
The future may include using AI to predict the onset of certain conditions or malignancy of lesions, Shoham says. She cites research showing that Aidoc’s decision-support software can shorten length of stay in inpatient and ED settings.
“Imagine the value shortening length of stay in the ED could bring—obviously to patients with acute conditions, but also to patients who have nothing wrong with them,” she says. “Who hasn’t experienced, personally or through loved ones, the hours spent waiting to be notified that there is actually nothing wrong with you and you can go home? Imagine the patient satisfaction with a facility that can show such improvements.”
She also sees AI tools as useful in getting radiologists out of the darkroom.
“Radiologists want to get back to consulting and determining paths of treatment,” Shoham says. “AI tools that flag and prioritize will help them do that, by increasing their overall efficiency and freeing them up to provide additional value beyond the detection.”
Regardless of the perspective, those involved in the development and use of AI in the clinical arena predict that an aging process is underway, as solutions become more specific and enhance the tool kit already available to radiologists.
“We’re seeing a separation between new random algorithms and AI solution providers,” Shoham says. “It’s been like the ‘Wild West,’ with so many algorithms out there being shared and built upon. It is time to show that there are mature AI solutions out there that should be providing value both to physicians and their patients.”
— Kathy Hardy is a freelance writer based in Phoenixville, Pennsylvania. She is a frequent contributor to Radiology Today.