Daily Bulletin 2017

The Reality of Deep Learning/Artificial Intelligence in Radiology: They Will Redefine the Specialty

Monday, Nov. 27, 2017

While there has been a lot of hype — and even fear — about the role deep learning (DL) and artificial intelligence (AI) play in radiology, the reality is that they are both potentially useful technologies that will add value to the specialty in a number of ways.

"Deep Learning is not going to replace us," said Paul Chang, MD, of the University of the Chicago School of Medicine, during a Sunday session on DL and AI in radiology. "But it will redefine us."

And radiology will need this technology more than ever due to the increasing demands on clinical imaging. Data sets are getting more complex and there is an increasing need to correlate images with other clinical information in order to implement practices such as radiogenomics, Dr. Chang said.

"So deep learning will help us because we are going to need something — we need some tool — some mechanism — to meet these new imaging challenges," Dr. Chang said. "We are going to need some kind of cybernetic help to get through a day's work and help us maintain and improve quality."

Infrastructure Remains a Challenge

But these are early days when it comes to incorporating DL and AI into the practice of radiology, and numerous challenges still exist.

For example, how can radiology confidently validate the performance of these new technologies?

"Deep learning is a great name for it because it has two meanings," Dr. Chang said. "It can mean 'very capable' or deep as in 'deep waters' or 'obscure,' and that's the problem.

"There are very deep layers to deep learning systems and it's very difficult to understand why they work."

Comprehending DL requires the use of cases and tons of data.

But radiologists really can't get compelling use cases unless they have the necessary data and infrastructure, Dr. Chang said.

Which brings up another challenge. Radiology doesn't have the infrastructure to either feed, train or consume these systems.

Paul J. Chang, MD


"Other industries have really revved up for cloud computing and big data and are ready to consume deep learning, because deep learning loves that kind of environment," Dr. Chang said. "Radiology is still struggling with electronic medical records (EMRs) and PACS and we generally don't have a true IT infrastructure that can feed and consume these systems."

The specialty should first pursue a "hedge strategy" by building infrastructures necessary to prepare for the cloud and big data, registries and advanced analytics, as well as DL, he said.

"The bottom line is that deep learning won't replace people — it will enhance them," Dr. Chang said. "We should be looking for the minimally heuristic use case sweet spot like workflow optimization. Something that isn't sexy, but is an easy win, saves money, and improves lives."

For those still unsure how DL fits into the healthcare landscape, Dr. Chang offers another comparison: "The analogy I use is the gold rush," he said. "Everyone went out west to dig for gold. Most miners either failed or died, but there were people who thrived — the people selling the miners the shovels. You needed to build an infrastructure."

Tip of the day:

Increasing the SID for upright radiographic exposures from 40" to 72" improves image quality through reduced focal spot blur and reduces patient dose.

The RSNA 2017 Daily Bulletin is owned and published by the Radiological Society of North America, Inc., 820 Jorie Blvd., Oak Brook, IL 60523.