Using AI technologies when imaging isn’t recommended for detection offers little benefit, according to an industry expert.
Artificial intelligence (AI) tools touted to help radiologists detect COVID-19 on imaging scans are popping up in the market almost as quickly as the virus continues to spread. And, while having a tool that could help you identify possible infections on the chest scans you do end up reading could sound enticing, you should view these technologies – and proceed – with caution, says one industry expert.
“The most important thing to ask about any AI tool is if there is a need,” said Ella Kazerooni, M.D., professor of radiology and internal medicine in the cardiothoracic radiology division at Michigan Medicine. “Just because you can doesn’t mean you should, and that can be applied to the use of thoracic imaging in the setting of COVID-19 pneumonia, in general.”
In addition to many new online, cloud-based tools, some existing AI technologies designed to detect other infections, such as tuberculosis or pneumonia, are now being offered as potential tools to help with COVID-19 detection.
Are AI Tools Needed During COVID-19?
Whether you even need an AI tool is up for debate, she said, and it all depends on the context in which you’re gathering the images.
Current guidance from both the American College of Radiology and the Centers for Disease Control & Prevention recommends against using any chest imaging for COVID-19 detection. That decision has been underscored by the growing body of research that indicates chest CTs are only roughly 60-percent accurate in correctly identifying which patients have a COVID-19 infection.
“If a negative CT doesn’t mean you don’t have the virus, and an abnormal scan doesn’t necessarily mean that you do, then what is the legitimate application of the AI tool?” she posited.
There are some reasonable instances where you should proceed with imaging in COVID-19 patients, she said, such as conducting a chest CT or chest X-Ray on an inpatient due to acute decline or other complications. In these cases, however, the patients have already been diagnosed, and the imaging is being done as part of disease management only.
Why Are AI Tools Flooding the Market?
Even though imaging and the associated AI tools should not be used for viral detection, Kazerooni said, she’s not surprised by the appearance of existing, as well as new, technologies coming to market to, potentially, help radiologists with identification. A driving force, she said, could be the current lack of access to high-quality testing as more hot spots appear as the pandemic continues to develop.
For detection and diagnosis, a nasal pharyngeal swab and reverse transcription-polymerase chain reaction (RT-PCR) are the recommended tests. However, due to the rapid spread of the virus and high demand for test kits, not every hospital or health system has enough resources to test every patient who presents with a possible infection.
Consequently, she said, AI tools that are presented as possible aids in detecting the virus can be attractive.
“If you’re faced with a clinical situation where you don’t have access to testing, and the testing you have is not good quality, you can see how people might think, “We should do chest X-Rays or CTs and apply a tool set,’” she said.
The Inability to Deliver on the Promise
One of the benefits of AI tools, overall, has been their specificity in identifying various disease characteristics or measuring particular values. To achieve that level of accuracy, the technologies are trained on a particular data set. And, it’s that specific training that will also make it difficult to apply the tools to COVID-19 detection, Kazerooni said.
“I don’t think it’s necessarily legitimate to claim that these AI tools that have been trained to detect tuberculosis or pneumonia can now be applied to COVID-19,” she said. “It depends on what you’ve tested and validated the tool on. Community-acquired bacterial pneumonia is very different from COVID-19 pneumonia.”
The question also arises of whether these tools were initially designed for detection or for characterization of findings. Mis-applying a tool can be critical, she said, and it could lead to potential clinical problems.
“False negatives are a concern,” she said. “Everyone needs to be cautious about these claims being made and what the test sets are for these tools.”
To side-step any potential problems with AI technologies, she advised you to only use tools that have been approved for use in the United States, as well as ones that have been cleared by the U.S. Food & Drug Administration (FDA). To date, the FDA hasn’t approved any AI tools to be used with imaging for COVID-19 detection.
Challenges of Implementation
There are obstacles to launching a new AI tool under normal conditions, she said. Doing so during a global pandemic when hospitals and health systems are re-aligning resources could be nearly impossible.
Before any tool can be used, all security and privacy concerns would need to be addressed. Doing so would require pulling a facility’s information technology (IT) resources away during a time when their expertise is needed elsewhere throughout the facility.
“I’d be concerned about the deflection of resources down a rabbit hole to implement something, particularly if it’s not going to add significant value,” she said.
She pointed to Michigan Medicine as an example. The health system has closed many outpatient facilities and has removed their computers, deploying them to field hospitals designated for COVID-19 services. Getting those new locations operational takes full IT dedication, and pulling them away could negative HIPAA consequences.
“Asking IT folks to do what needs to be done to implement any new AI tools during any type of declaration of emergency has no true value,” she said.