Paul Chang: 3 things to know about AI, deep learning in 2018

Paul Chang, MD. Photo courtesy of The University of Chicago.

Artificial intelligence (AI) and deep learning technology continue to be hot topics in healthcare. The topics dominated RSNA 2017, generating plenty of interest as they become an integral part of health imaging. However, while many in healthcare are excited about its potential to change every workflow and improve accuracy, others remain skeptical.  

As the first month of 2018 slowly comes to an end, Health Imaging spoke with enterprise imaging and health informatics expert Paul Chang, MD, professor of radiology and vice chair, radiology informatics at University of Chicago Medicine, about what practitioners and healthcare technology leaders should keep in mind regarding AI and deep learning in the coming year.  

It will take longer to incorporate and consume AI and deep learning than to adopt them.

Referencing the Gartner hype cycle, which represents the stages a technology goes through from conception to maturity and widespread adoption, Chang explains that healthcare is no stranger to buying into the hype of new technology very early.  

"We're just beginning to learn how to consume, many of us don't even have score cards or dashboards or even human consumable analytics yet," Chang explained to Health Imaging. "There's nothing particularly special about AI or deep learning that makes people think that we're going to go through this hype cycle; it's just the nature of how we do it."   

Chang doesn't doubt that healthcare industries will consume AI and deep learning effectively. It will only take a matter of time and during that time, or "early stage," he explained that the opportunity to socialize, change workflow and impact workplace culture will be feasible and necessary.   

"AI and deep learning doesn't replace us. It frees us to do more valuable work," he said.   

There will be challenges, but take advantage of the early stages.  

Chang asserts that deep learning has two meanings that go hand in hand: Deep learning is both capable but also is obscure—and with obscurity comes the overarching challenge of adopting AI and deep learning into the healthcare without having a preset underlying model.  

Specifically, Chang outlines four challenges that AI and deep learning will present to any industry, including healthcare:  

  • We don't have the data needed to train and validate deep learning.

"These systems are incredibly greedy; they need a lot of vetted data because in a lot of ways deep learning is a lazy approach, it's a brute force approach. The clever approach is machine learning like CAD that uses a preconceived model," he said. "Deep learning gets to that model, but by using a brute force method, it requires you to feed it lots of vetted data. You don't know why it [deep learning] is looking for patterns, which is why you need incredible amounts of data to train it and need a lot of data to validate it."

  • Healthcare industries don't have the proper IT infrastructure to feed and consume AI and deep learning data.

"It's time now to build up our IT infrastructures, go beyond our PACS-centric perspective and build a new IT infrastructure to consume these systems," Chang said.  

  • Lack of available data sets and cases.

"These youth stages are nice to have," Chang said. "AI and deep learning are driven by data availability, not by youth stages. We don't have available data yet and that's why the youth stages are pretty boring."  

  • Weak data infrastructure and inefficient workflow automation.

"The real utility of deep learning is the minimally heuristic cases to improve the business of radiology, not diagnosis," according to Chang. "When you look at other industries—Google, Amazon—the way they use deep learning is not to so much replace the knowledge worker, but to augment and improve the efficiency, to reduce the error, reduce the variability of the workflow."  

Investments must be made for data security and privacy.  

"I think deep learning is a version of what I call data-centric decision support tools. Big data, deep learning, analytics, they are all types of data driven support tools. As such, these technologies are going to require a degree of data access and interoperability that will require us to spend and invest a lot more in data security and privacy," Chang said.