Projections of radiology’s demise at the hands of algorithms have been greatly exaggerated. In fact, not only will machine learning not take rads’ jobs: It will become a routine component of their clinical practice, making their work more efficient, accurate, satisfying and valued.
So predict Michael Recht, MD, of New York University and R. Nick Bryan, MD, PhD, of the University of Texas in an article published online Aug. 19 in the Journal of the American College of Radiology.
Recht and Bryan point out that most algorithms now being developed or refined provide computer-assisted diagnosis and detection (CADD) of discrete radiologic findings like pulmonary nodules, intracranial bleeding, mammographic lesions and colorectal polyps. (Note the additional “D” for diagnosis in the familiar acronym.)
They expect such capabilities to supply increasingly reliable and useful pre-human-read services.
“[I]t is possible that no medical imaging study will be reviewed by a radiologist until it has been preanalyzed by a ‘machine,’” they write. “Clinically validated machine-learning technology integrated into radiology workflow will allow radiologists to more efficiently produce better quality reports. CADD will make us better radiologists, not replace us.”
With AI-driven gains in efficiency will come greater productivity, the authors suggest.
When machine learning is widely adopted, radiologists will be able to spend more time “integrating patients’ clinical and imaging information, having more professional interactions, becoming more visible to patients and playing a vital role in integrated clinical teams to improve patient care,” they write.
And with greater visibility to patients and contributions to care teams will come greater job satisfaction.
“It is unlikely that any radiologist entered the field with a desire to spend a large amount of time measuring lesions and doing segmentation,” Recht and Bryan write. “There is little doubt that machine-learning algorithms will be able to replace the majority, if not all, of the quantification tasks currently performed by radiologists as well as accomplish data mining of the electronic medical records in a more efficient manner.”
Recht and Bryan do foresee the day when AI-enabled efficiencies are so significant that fewer radiologists are needed. However, the headcount adjustments will be modest in scope, they expect, as the change will affect future residency programs rather than existing clinical practices. What’s more, it will ultimately redound to the overall health of the profession.
“As competition for residency positions has been decreasing over the past several years, this [reduction] will allow a better match between supply and demand for residency positions,” they write.
Meanwhile, among the formidable obstacles machine-learning CADD would have to surmount if it were ever to totally transform rather than judiciously improve radiology, Recht and Bryan see three big ones: FDA premarket approval of algorithms aimed at providing autopilot-level diagnostics, patient acceptance of machines as go-it-alone diagnosticians and the threat of lawsuits to AI machine makers (aka software developers).
All of which point to machine learning’s likely future as a technology that will contribute much to—rather than take anything away from—radiology, as Recht and Bryan see it.
Commenting on the legal picture, they observe it’s likely that algorithm creators “will be held responsible if machine-learning algorithms, used by themselves without supervision by radiologists, miss significant pathology. It is far less likely that manufacturers will be held liable if these algorithms are only used to aid radiologists, not replace them.”