Applying a deep learning approach to echocardiography could save clinicians time while improving diagnostic accuracy, one Georgia-based cardiologist reported at the American College of Cardiology’s annual meeting in Orlando, this year.
Chief medical officer at Bay Labs, a San Francisco-based medical technology company, Randolph P. Martin, MD, said machine learning—and its self-training subset, deep learning—is often perceived as a threat to cardiologists, but the tools could aid clinical decision-making in a way humans can’t. Edwards’ CardioCare program revealed that in a study of 150,000 consecutive echocardiograms screening for aortic stenosis, 24 percent of echos were of inadequate quality. Most cases couldn't complete quantification of aortic valve parameters, Martin said, and the trial saw frequent discordance between guideline recommendations and reality.
Deep learning goes a step further than machine learning, but the pair still fall under the umbrella of artificial intelligence, Martin said. While machine learning allows statistical techniques that enable machines to continue improving at tasks, deep learning itself is composed of hierarchical layers of algorithms designed to train itself to perform tasks.
“Quality is where I think machine learning could really help,” Martin said. “But deep learning is really going to be the next way.”
The cardiologist said the technology is already ubiquitous—just look at Amazon’s speech recognition, Apple’s ability to unlock phones with a selfie or Google’s self-driving cars. It has the ability to stretch diagnostic abilities further when it comes to aortic stenosis by capturing important measurements sonographers might miss and calculating the likelihood of comorbidities.
And now is prime time for deep learning, Martin said. The industry has never had so much computer power, such extensive labeled datasets or such advanced mathematical algorithms.
“These things have got so much power that they learn from data,” he said. “We’re not talking about replacing you or super-sophisticated echocardiography machines, but we are talking about improving quality, improving acquisition and improving the ability to interpret studies.”
AI is already at work in ultrasound labs, but where machine learning could really benefit echo is in acquisition and interpretation of results, Martin explained. With deep learning algorithms, clinicians would be able to achieve the best images with the smallest learning curve and interpret the probability of a normal or differential diagnosis.
He said that although cardiologists may perceive the advanced technology as a threat to their profession, there’s nothing to be worried about. Deep learning can improve workflow and the completeness of scientific studies, he said, and, as humans, doctors are the ones who will interpret that data and make management decisions.
“This is not going to put you out of business,” he said. “It’s going to make the system better.”