Radiologists, Pick Your Replacement: Watson or Pigeons
Forget turf battles with non-radiologist providers. The new bogeyman looking to replace radiologists isn't even human. Should we worry?
Welcome to another Radiology Firing Line podcast recorded at RSNA 2015. Host Saurabh "Harry" Jha, MBBS, poses the following question to his collaborator C. Matthew Hawkins, MD, first-time podcast guests Andrew DeLaO (patient advocate), Adam Powell, PhD (adjunct professor of informatics at Northeastern University), and returning guest Jason N. Itri, MD, PhD (vice chair of operations at the University of Cincinnati): Will Watson replace radiologists? As you might imagine, they don't all agree.
Behold the power of binary code: The IBM supercomputer Watson. Watson represents artificial intelligence more robust than existing machine learning. Originally designed for Jeopardy!, Watson won when it played on the game show in 2011. But it also displayed a few quirks, such as incorrectly answering a question about cities that have multiple airports: when comparing choices, it could not eliminate a nonviable choice. In other words, Watson was guilty of essentially concluding "cannot exclude" (a radiology report pet peeve).
Having recently acquired Merge Healthcare (medical image handling and processing technology), IBM Watson Health announced that it is developing cloud-based analysis of medical images. If the job of radiologists is to merely combine pattern recognition and statistical probabilities to make diagnoses when viewing images, then a limitless capacity to learn and retrieve knowledge, to apply complex rules and algorithms, and to modify those rules and algorithms by feedback loops, combined with zero fatigue concerns, favors Watson over physicians.
But perhaps, Watson is better suited to enhance radiologic diagnosis as a trusty sidekick, relieving radiologists of mundane, repetitive duties, allowing us to expend our energies on high-order thinking. After all, IBM called it Watson, not Sherlock (though, full disclosure, the supercomputer was named after Thomas Watson, IBM's first CEO, but the imagery of Dr. Watson and Mr. Holmes is more captivating).
"You know my methods, Watson."- Sherlock Holmes in The Crooked Man
Look at computer aided detection (CAD) systems for niche applications, like finding pulmonary nodules on CT. The goal of CAD is to find all potentially abnormal findings, i.e., to be highly sensitive. Although CAD may identify many candidate abnormalities, the radiologist decides which CAD-selected abnormalities represent pathology in need of further surveillance, additional testing, or other management recommendations. Thus, the radiologist interprets the findings to provide specificity.
"It is of the highest importance in the art of detection to be able- Sherlock Holmes in The Reigate Puzzle
to recognize, out of a number of facts, which are incidental and which vital. Otherwise
your energy and attention must be dissipated instead of being concentrated."
Perhaps this explains the role Watson may best serve: eliminating missed diagnoses. Watson can leave the assessment of the relative importance of findings, including which to ignore, up to the radiologist. Although Watson can assist the radiologist by data mining the electronic medical record to alert the radiologist to the statistical probability of meaningful pathology, radiologists' expertise actually lies in their judgment regarding imaging findings.
Watson may face its most formidable foe not in the potential form of radiologist resistance, but in the form of government bureaucracy. IBM will have to demonstrate safety and efficacy as part of the FDA's lengthy premarket approval process. But unlike CAD's fixed software rules, Watson's deep machine learning is designed to intelligently modify its algorithms based on continual feedback and data analysis. The ultimate purpose of this artificial intelligence is image interpretation (i.e., diagnosis), not merely detecting potential abnormalities. Based on the size and scope of IBM's ambitions, the scrutiny for Watson as a broad-based diagnostic image interpreter must be much greater than for any one niche radiology CAD application. It will likely take many years to obtain FDA approval.
Until then, a flock of pigeons is waiting in the wings (pun intended). Ever since recent published reports about pigeons trained to identify specific findings on mammograms, the pigeon-as-mammographer has become an amusing social media meme.
In an era of patient-centered care, analyzing findings in the context of the whole patient and displaying sensitivity to individual patient values have taken on greater importance in medicine, including radiology. But Watson doesn't care one iota about you or anyone else.
The pigeons are indifferent to your health and well-being too, but at least they'll seem interested in you as long as you keep tossing them stale bread chunks. And, unlike IBM, pigeons aren't interested in making a profit and don't answer to shareholders.
In the 2014 medical thriller Cell by Robin Cook, a radiology resident discovers that a hot new smartphone app with diagnostic and treatment capabilities is malfunctioning with horrific unforeseen consequences. Is the culprit a software engineering glitch or a medical-industrial complex conspiracy? You'll get no spoilers from me, but the story does foreshadow current concerns related to ceding physician control to an autonomous supercomputer. Which makes me wonder: What might happen in the future if an autonomous Watson, FDA-approved to both interpret and manage imaging findings, goes rogue? Maybe we'll wish we had partnered with the pigeons.