Could a robot read a CT scan better than your doctor?

By the time Dr. Ty Vachon started his radiology training in 2010, Apple had already turned the humble cell phone into a powerful personal computer. The tech giant was also just months away from introducing Siri, its voice recognition technology with the ability to understand questions and respond to commands. But Vachon, who did his residency with the U.S. Navy, was still reading imaging scans on a clunky machine that wasn’t connected to other hospital systems.

“Two days into my residency, I realized that my phone was smarter than my big radiology computer,” he said.

The realization soon led him to start learning programming languages like Python and Java, if only to learn how programmers think. He now works as a consultant to medical informatics companies trying to break into the space.

These days, the healthcare industry is playing catch-up and finally starting to experiment with machine learning and artificial intelligence as tools to help doctors make treatment decisions. It’s undeniable that these technologies excel at pattern recognition, making them well-equipped for use in, say, identifying suspicious moles or lesions.

That means that fields like dermatology and radiology will likely be the first to see wide adoption of these platforms. While some doctors are excited about the possibilities, others are concerned about what the future holds.

Many of the conversations around the future of work are focused on automating low-skill tasks like supermarket cashiers. But even highly-skilled workers are wondering if they too could one day be replaced by machines.

As radiologists gathered this week in Chicago at the annual meeting of the Radiological Society of North America, the topic of artificial intelligence made an appearance more than two dozen times in the program.

We asked doctors for their thoughts on this topic and you can read their responses below.

Vachon said he has tracked about 170 companies working on artificial intelligence in radiology and has had conversations with more than 60 of them.

The biggest issue that radiologists point to in integrating AI is about workflow. Many manufacturers require that radiologists use a separate computer for reading imaging scans, creating a barrier to startups that can’t integrate with their platforms. The giants in this space, like Siemens, Philips and GE, often have multi-year contracts with the hospitals that purchase their imaging equipment.

The result: Much of the work around AI and machine learning has been happening in academic and research settings.

But this workflow issue could be resolved as more of the heavyweights begin to unveil AI platforms. Philips, for instance, says that 60 percent of its research and development team is working on predictive analytics as well as AI and machine learning.

The low-hanging fruit for the company is advanced pattern recognition. For instance, the company’s Illumeo platform can sort through the large number of images created during a single CT scan, and identify which lesions should be measured first.

“It learns how the radiologist works,” said Dr. Roy Smythe, the company’s chief medical officer for medical informatics, who claims that the system makes radiologists 30 percent more efficient. “Clinicians want this because it speeds them up and it helps them narrow down the differential diagnosis much more quickly.”

He likens the technology to an interior decorator taking measurements of a room before the real work begins. It’s that kind of busywork that the company is trying to reduce, he says.

But even on the diagnostic side, the company is looking at whether its software can make simple screening decisions. The company recently conducted a study in Singapore looking at whether the technology could help speed up the screening process that all immigrants must undergo to check for tuberculosis. Southeast Asian countries are facing a dire shortage of radiologists.

“We’re on the cusp of such an impact on efficiency,” Smythe said. “Once we begin to use AI to predict and prevent illness … then the cost savings are going to be tremendous.”

AI might be able to replace specific aspects of a radiologist’s job, but what it can’t do, many physicians argue, is replace the years of knowledge doctors develop from reading scans that don’t fit the typical pattern.

Making a diagnosis and coming up with a treatment plan is often more art than science, and highly dependent on a patient’s individual medical history.

“AI is really well-suited to image recognition,” said Dr. Michael Muelly, a Stanford radiologist who also works on the medical imaging team at Google Brain. “But it’s a misconception that people have with the radiology field. All of these companies don’t understand what radiologists actually do. … Medicine, to a large extent, isn’t a data-driven field.”

Still, he’s had enough medical students ask him whether they should consider a career in radiology. For the foreseeable future, he assures them, they’re safe.

“The problem right now is that we don’t have enough of a dataset to encode a lot of stuff that we know from general medicine,” Muelly said. “And there are questions about whether we can even get enough of a dataset.”

That’s a big reason why Dr. Anand Patel, a California-based radiologist and a medical and technology consultant at Maverick Capital, has not recommended any AI companies as investments.

“So far it’s been a lot of marketing and very little, if any, evidence,” he said. Startups need access to vast amounts of data that isn’t easy to get, and their algorithm for one part of the body (like a chest X-ray) likely won’t work for another (like mammography.) “It becomes very hard to scale.”

Still, Muelly and others believe that the advances in machine learning have the potential to make the radiology field more data-driven, or help answer questions that currently can’t be solved.

One of the biggest advantages of AI is that it can detect sub-visual features on images, or those that a human can’t see, said Dr. Raym Geis, vice chair of the informatics commission at the American College of Radiology. And that will open up more opportunities for radiologists, rather than fewer.

“Suppose it could look at brain scans of young people with head injuries, and predict which ones will get CTE?” he said, referring to chronic traumatic encephalopathy, a degenerative disease seen in people who get frequent head injuries. “Or which patients with fatty liver will go on to develop cirrhosis or hepatoma? I think AI ... will provide a whole new world of information and it’s going to require a lot more radiologists to help manage and communicate these entirely new procedures.”

In their own words: Doctors weigh in

Quest for Efficiency: How Artificial Intelligence Would Drive Image Interpretation in Private Practice Radiology by Dr. Brian Yue

Six Challenges To Tackle Before Artificial Intelligence Redesigns Healthcare by Dr. Bertalan Mesko

RSNA and ML: 3 Big Questions by Dr. Ty Vachon