For years, many radiologists have feared that advances in data science could put them out of a job, as artificial intelligence (AI) tools become increasingly adept at reading diagnostic medical images.  But Eliot Siegel, M.D., Professor of Radiology and Vice Chair of Research Information Systems at the University of Maryland School of Medicine in Baltimore, doesn’t believe AI technology will replace radiologists any time soon.

 

 

For years, many radiologists have feared that advances in data science could put them out of a job, as artificial intelligence (AI) tools become increasingly adept at reading diagnostic medical images.  But Eliot Siegel, M.D., Professor of Radiology and Vice Chair of Research Information Systems at the University of Maryland School of Medicine in Baltimore, doesn’t believe AI technology will replace radiologists any time soon.

“Early predictions of AI’s advancements have underestimated the challenges of applying data science and deep learning in medical images,” said Dr. Siegel in his November 10 webinar hosted by the University of Miami Institute for Data Science and Computing (IDSC) and Clinical and Translational Science Institute (CTSI).”  “However, he added, AI can provide valuable assistance to radiologists as well as potential benefits to patients.”

A radiologist who has been at the forefront of medical imaging for decades, Dr. Siegel gave an insightful talk on “AI and the Presumed Demise of Radiologists,” in IDSC’s Data Citizens: A Distinguished Lecture Series.

Yelena Yesha, Ph.D., Knight Foundation Chair of Data Science and AI, IDSC Innovation Officer and Head of International Relations, and Founding Director of the National Science Foundation’s Center for Accelerated Real Time Analytics (CARTA), introduced Dr. Siegel. “He launched the first filmless radiology department in the country in the 1990s and is now leading the way on AI,” she said.

Challenges with AI

Dr. Siegel said AI systems trained by machine learning can distinguish between cats, dogs, horses, and other types of images. Convolutional neural networks (CNN) work as filters and are able to classify images with incredibly high accuracy, he said. However, interpreting a patient’s radiology scan is a much more complex task, particularly with large data sets that include 3D images, and involve assessments of changes over time.

“The task of determining what is wrong with an image is fundamentally harder than simply recognizing an object in an image,” he said. For example, he cited the children’s puzzle, “What’s wrong with this picture?” that might show a TV on the roof of a house or a snowman in the yard on a sunny day. “No supercomputer can beat a five-year-old at identifying the errors,” he said. “You just can’t teach a computer the general knowledge about how the world works.”

Validating AI algorithms

Another challenge facing radiologists is assessing the accuracy of the thousands of AI algorithms being developed. “It’s hard to know what’s best for different data sets,” said Dr. Siegel. “You can create algorithms in just a few days but validating them still takes months or years. Regulatory approval takes time as well, making validation a major bottleneck.”

Dr. Siegel added that physicians are still struggling with how to test human trainees for competence in radiology, let alone testing a new algorithm. There is also the “black box” nature of deep learning, which makes it difficult to tell how AI is making its image determinations. For instance, one AI system that claimed a high accuracy in identifying TB images was actually just reading the TB Clinic’s labels at the bottom of patient disease images.

For the foreseeable future, there needs to be a human in the equation to be sure there are not common-sense things that AI would miss, said Dr. Siegel, adding that patients value their physician’s empathy and compassion along with medical skills.

Benefits of AI

Looking ahead to the next four years, Dr. Siegel said, AI offers a number of benefits for radiologists and their patients. For instance, AI systems can organize and prepare patient images, allowing physicians to spend more time interpreting scans.  AI systems can also incorporate deep learning to improve the quality of low-resolution images. As a result, scanning times in some cases could be reduced from several minutes to just a few seconds, reducing patient exposures. Those AI-generated image enhancements can also give radiologists a better view of certain anatomic structures.

Dr. Siegel’s wish list for the next generation of clinical AI in radiology includes new capabilities such as measuring liver or pulmonary tissue textures. Other advancements include incorporating laboratory reports or patient-family histories to assist radiologists in assessing the probability of disease.

“AI will have a major positive impact on radiology and healthcare,” concluded Dr. Siegel. “I counsel our residents that radiology will become more interesting specialty, as AI becomes their partner rather than their replacement.”

 

STORY by Richard Westlund

 

This Data Citizens: A Distinguished Lecture series talk took place on Wednesday, November 10, 2021 (4:00-5:00 PM ET via Zoom).

TALK TITLE: “AI and the Presumed Demise of Radiologists”  Paraphrasing Mark Twain’s famous quote, “The reports of my death are greatly exaggerated,” predictions made during the past half a dozen years about the demise of “Radiologist” as a profession have been made with alarming frequency. Geoffrey Hinton, one of the “fathers” of Deep Learning stated that “They should stop training Radiologists now.” Andrew Ng, Professor of Computer Science at Stanford, declared that it would be easier to replace a Radiologist than his personal assistant. Ezekiel Emmanuel proclaimed in the New England Journal of Medicine that, Radiologists, as a profession, could cease to exist in the next few years and that AI was the greatest threat to the practice. A half a dozen years later, it has become obvious that those predictions made by some of the most acclaimed “experts” in Computer Science and Medicine got it completely wrong. This talk will explore the fascinating, daunting, and yet tantalizing challenges associated with the implementation of AI in clinical practice for Diagnostic Imaging and will discuss the hype, reality, and future of Deep Learning for Diagnostic Imaging.