by John W. Mitchell
, Senior Correspondent | August 18, 2016
In the next ten years, computers could be reading the majority of routine diagnostic imaging tests such as mammograms and chest X-rays.
This could allow radiologists to spend their time sorting out abnormal findings, conducting invasive procedures, and spending more time with patients.
That’s the prediction of a physician expert who presented a webinar titled “Deep Learning: How It Will Change Everything”, organized by the Society for Imaging Informatics in Medicine (SIIM) and attended by nearly 300 people on Wednesday.
“Deep learning is the hot area,” Dr. Bradley J. Erickson, M.D., Ph.D., professor of radiology and associate chair for research at the Mayo Clinic told HCB News. “Physicians may say they have information that may not be computable ... but deep learning allows imaging reading by computers that see more than we see.”
Numed, a well established company in business since 1975 provides a wide range of service options including time & material service, PM only contracts, full service contracts, labor only contracts & system relocation. Call 800 96 Numed for more info.
Erickson explained that deep learning is the next step after traditional machine [computer] learning. In machine learning an algorithm finds features – or things that are measured – and then learns a correct answer.
Deep learning automates both of these steps to make the findings faster and with greater accuracy. He noted that deep learning research has improved machine interpretation applications from a degree of error of around 25 percent in 2011 to less than four percent in 2015.
He also said that a criticism of deep learning is that it required millions of samples to be accurate. But Erickson “busted that myth” by presenting several research cases in which as little as tens of imaging samples were used to accurately make clinical findings.
“Machine learning is objective,” he stressed. “Its decisions are based on data.” Erickson added that machine learning is also pervasive, meaning it can be applied to images 24 hours a day (including holidays and weekends) from underserved areas as easily as in major metropolitan cities.
Deep learning still has a way to go and would be subject to FDA approval as a new medical device.
“Its use today is still limited,” Erickson said. “But I think that reports showing it can find textures and patterns in MRI images that reflect genomic properties that are not visible shows the power and promise of applying machine learning in medicine.”
Advancements are accelerating, especially in the past two years. He added that advancements in the next two years will be even more exciting. He attributed this to the exponential growth in computing and algorithmic capability, much of it developed by the gaming industry.
“Bill Gates said we always overestimate what we can do in two years, and underestimate what we can do in ten,” Erickson said. “I think that will be the case for machine intelligence applied to medical images.”
He also said that for deep learning to be successful, radiologists must embrace this new technology.
“Physician engagement has to part of this revolution,” he added. “We want the computers to be the underlords (serving physicians and patients), not the overlords (ruling physicians and patients).”