By Aine Cryts

What do Isaac Asimov’s “I, Robot,”—a series of stories published in 1950 about the development of robots—and Robert Heinlein’s “The Moon Is a Harsh Mistress,” a 1966 book about a lunar colony’s revolt against Earth, have in common? They’re two books 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, points to as sparking his interest in artificial intelligence in radiology.

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 topic of artificial intelligence in radiology with Siegel, as well as what he believes are the top misconceptions surrounding the field. Here’s what he had to say.

AXIS Imaging News: How did you get interested in artificial intelligence within radiology?

Eliot Siegel: I grew up an avid fan of science fiction, reading Isaac Asimov’s “I, Robot” and Robert Heinlein’s “The Moon Is a Harsh Mistress.” Those two books are among many others in which artificial intelligence played a prominent role. In addition, I minored in computer science and worked my way through medical school doing computer programming at the National Institutes of Health and for professors at University of Maryland School of Medicine.

Eliot Siegel

Eliot Siegel

Having the opportunity immediately after my radiology residency to design and create the world’s first fully digital radiology department at the enterprise level at the Baltimore VA Medical Center, it seemed natural to want to accomplish the following two goals:

  1. Provide access to images everywhere to everyone who needed them. That seems to be a given today, but it was considered a far-off dream 30 years ago.
  2. Make the images available in a digital format for routine analysis, detection, and diagnosis using computers to help make diagnostic imaging more efficient, accurate, timely, and relevant. This would allow radiologists to concentrate more on judgement and applying our medical knowledge to interpret the relevance of findings.

In various roles with the International Society for Optics and Photonics Medical Imaging Conference, SIIM, and the Radiological Society of North America’s Radiology Informatics Committee, it’s been gratifying to see and participate in gradual and then exponential progress in artificial intelligence applications within radiology and nuclear medicine.

AXIS: Why should radiologists be enthusiastic about the opportunities artificial intelligence in radiology presents? 

Siegel: Radiologists are struggling with the challenge of interpreting greater volumes of studies of increasing complexity. In addition, many of the tasks associated with retrieving, organizing, and reviewing those studies are repetitive and tedious, resulting in fatigue and stress.

Artificial intelligence offers the potential to provide improved workflow and intelligent protocoling and presentation of imaging studies; improved access and summarization of clinical information; better follow-up of recommendations and important findings; and assistance in making pertinent observations in images, as well as in determining the implications of those.

Artificial intelligence also offers the potential to achieve higher-quality images with less radiation, lower amounts of contrast, and shorter imaging times. In short, artificial intelligence offers many of the advantages of a radiology “fellow,” who can assist in acquisition of background information associated with a study; image organization; and preparation, interpretation, and communication.

In a manner analogous to Apple’s App Store and the Google Play Store for the Android, best-of-breed artificial intelligence applications will be offered in emerging artificial intelligence “stores” or “platforms.” As is the case with smartphone apps not being limited to a single provider, such as Apple or Google, radiologists and other healthcare providers will be able to pick and choose the applications that work best for them.

Going beyond the App Store/Google Play Store concept, the best artificial intelligence platforms will be able to support multiple applications working together. These combinations or “ensembles” of applications will work cooperatively to provide multiple perspectives on the same task or multiple components of a more complex task.

AXIS: What are some misconceptions radiologists might have about artificial intelligence? How would you address those misconceptions? 

Siegel: An early misconception that was perpetuated by non-radiologists, including machine learning experts who didn’t understand the nuances of image interpretation, was that artificial intelligence would quickly replace radiologists. However, we’ve learned that at least initial artificial intelligence algorithms have numerous limitations, such as:

  • Being relatively narrow in their scope
  • Being relatively “brittle” in their lack of generalizability to different brands/models of scanners, different patient populations and environments, and regulatory and legal restrictions
  • The high cost and time requirements to annotate datasets with difficulties keeping up with ever-evolving technologies

Another fundamental challenge was the incorrect assumption by machine learning experts—who were unfamiliar with the radiology domain—that the main challenge was to determine what was “in” an image, rather than what’s wrong or has changed about an image. In short, major successes in everyday images of animals and plants and other object identification did not generalize to determining what’s wrong with complex medical images.

Radiologists [also falsely assume that artificial intelligence will mostly impact systems used] to make findings on images and subsequent diagnoses. However, in the short term, artificial intelligence will have its greatest impact on improving image quality and time of acquisition for cross-sectional modalities, such as MRI and CT; in addition, there are opportunities for dose reduction and artifact reduction.

Artificial intelligence applications will be utilized for improved follow-up and communication and workflow improvement—and these will emerge at a faster pace than artificial intelligence applications for making findings in the marketplace.