By Aine Cryts
Radiologists “absolutely” should be cautious about artificial intelligence in radiology. That’s according to Eliot Siegel, MD, FSIIM, professor of radiology at University of Maryland School of Medicine and chief of imaging services at VA Maryland Health Care System. He cites “numerous pitfalls” in the adoption of artificial intelligence, such as bias in the development and training of algorithms that focus on specific populations and may not be applicable across general patient populations.
Still, Siegel is enthusiastic about what the next five years hold in terms of basic algorithms that will find lung nodules or detect brain hemorrhages, in addition to artificial intelligence solutions that will help radiologists better understand multiple sclerosis or myocardial ischemia.
Siegel is a conference co-chair for the Society for Imaging Informatics in Medicine’s (SIIM) 2019 Conference on Machine Intelligence in Medical Imaging, which takes place in Austin, Texas, on September 23 and 24. AXIS Imaging News recently discussed the need for radiologists to be both cautious and enthusiastic about the use of artificial intelligence in radiology. Here’s what he had to say.
AXIS Imaging News: Should radiologists be cautious about AI in radiology?
Eliot Siegel: Absolutely. There are numerous pitfalls in the adoption of artificial intelligence in radiology. These include bias in the development and training of algorithms and models that are trained on different populations of patients or different types of scanners or different types of studies.
Claims for artificial intelligence that suggest performance at or above “expert” levels are rarely borne out in practical experience. Artificial intelligence programs tend to treat each study independently, rather than utilizing the knowledge of prior examinations. In addition, claims made by artificial intelligence vendors are difficult to validate and FDA clearance doesn’t necessarily mean that the federal agency validates the clinical efficacy or validity of the software. Even algorithms that improve accuracy in diagnosis may result in substantial disruptions in interpretation efficiency if they’re not fully integrated into the radiologist’s workflow.
Radiologists should have the opportunity to try out potential algorithms on their own data in their own setting and insist on minimizing disruptions to their workflow. The impact of artificial intelligence algorithms to improve image quality is unknown since the use of artificial intelligence in image reconstruction is in its infancy. In addition, images that look aesthetically pleasing may not necessarily be more diagnostic than older methods of image reconstruction.
AXIS: What’s next, say in the next five years, for artificial intelligence in radiology?
Siegel: The next five years will see basic algorithms that find lung nodules or do organ segmentation or detect brain hemorrhage become relative commodities, with many vendors offering similar capabilities. We’ll see the emergence of more clinically oriented packages such as a multiple sclerosis or parenchymal lung disease or brain or myocardial ischemia suite of applications that combine multiple algorithms that detect disease with clinical decision-support tools along a clinical theme.
In addition, we’ll see the emergence of an increasing number of platforms that not only make artificial intelligence applications available within an imaging workflow locally, but also allow these applications to work together and permit adjustment and dynamic fine-tuning by the radiologist.
The next five years will bring greater choice in image visualization and image analysis by radiologists.
All the major vendors will offer artificial intelligence/deep-learning software that will be integral to their image reconstruction/formation algorithms. Increasingly, organ segmentation and lesion localization will be deployed. Natural language processing/understanding will allow improved parsing out of important details from a patient’s previous imaging reports and the EHR.
AXIS: What are your biggest concerns about AI in diagnostic imaging over the next several years?
Siegel: I have two big concerns, which include the following:
- The potential for artificial intelligence companies within or outside the United States to begin to directly offer to patients “automated artificial intelligence” interpretation for their studies—for instance, brain MRI, knee MRI or chest CT. It’s critically important to have radiologists intimately involved in interpretation. In addition, artificial intelligence must continue to represent a tool for radiologists who have the background and judgement to determine whether or not to believe the findings/recommendations of artificial intelligence systems.
- Over-reliance on artificial intelligence recommendations with radiologists not questioning the validity of artificial intelligence algorithms or letting their diagnostic skills atrophy with the use of this new technology. It’s critical that radiologists understand not only the great potential but also the limitations of artificial intelligence systems that will be present in the foreseeable future.
I agree with Aine that we need to approach AI deliberately. Right know there are a plethora of findings from disparate organizations and very little to unify and even less information to show that the image training sets not only included but also created specific learned algorithms for various races, sexes, and common congenital variances.
Until the creators can prove that their AI algorithms can competently address these variance the interpreting physician should regard the devices a highly specialized computer aided diagnostic tool.