With Triage System, Companies See Toehold for A.I. in Radiology
This story is part of an ongoing Xconomy series on A.I. in healthcare. Other stories cover big-company efforts, a genomics hackathon, the impact on doctors and patients, and business models of A.I. in healthcare.
Years ago, radiology underwent a radical transformation with the shift from film to widespread use of digital image displays. That set the stage for the kinds of artificial intelligence technologies that promise—or threaten, depending on your perspective—to upend the profession again today.
As with any new technology that would impact the healthcare industry, even a relatively small corner of it, a major challenge persists: How can it find a way through the medical bureaucracies and into the harried routines of practicing healthcare providers?
That’s where a company called Clario sees its opening. Hundreds of radiologists already use the Seattle firm’s software to manage work in their increasingly large practice groups. Now, Clario is teaming with CuraCloud, an under-the-radar Seattle startup building image analysis, genomics, and natural language processing technologies, to develop an A.I. tool for radiologists that would automatically identify life-or-death radiographic exams and move them to the top of the radiologist’s queue.
“What our A.I. can do is escalate the priority to ‘stat’: look at this exam next,” says Ed Butler, vice president of marketing at CuraCloud. “It shortens the time between [taking the scan] and the radiologist actually looking at it, based on the severity of the situation.”
The companies’ partnership illustrates a possible pathway to market for the multitude of startups and corporations looking to bring A.I. to the front lines of healthcare.
Many companies in A.I. are working on “really intellectually interesting problems,” Butler says. “But getting it actually into the workflow at the right time has been one of the toughest nuts to crack.”
Clario, founded in 2008, solves a problem that emerged after the physical X-ray films that used to be carried by hand—their folders covered with notes and special instructions—were replaced with digital image files. Hospitals adopted picture, archive, and communications systems (PACS), untethering radiologists from the physical location of the scanning equipment and changing old ways of working.
Clario’s software enables some of the workflows from the old world to continue in a digital format, allowing radiology practice groups that might be working with several hospitals and health systems to manage and prioritize incoming scans, assign them to specific radiologists, and communicate with the referring physicians.
As private radiology practices have grown larger, the ability to efficiently manage incoming work has become essential. “They have compensation that’s centered around how productive they are,” says Clario co-founder and CEO Chris Wood. “So we were able to sell to that market initially because we had a direct impact on the bottom line of the practice.”
Clario counts 25 private practice customers with some 1,200 radiologists using the company’s software to manage their daily work.
Today, each practice using Clario’s software sets its own rules to determine which incoming scan should be read next. But the A.I. technology CuraCloud is developing would automate that prioritization for certain scans that could contain life-or-death information about a patient.
A cranial scan with a possible brain hemorrhage would usually be sent “stat” by a referring physician to a radiologist, but it doesn’t always happen. And in some busy practices and hospitals, that life-critical scan could fall through the cracks and languish for hours or even a day on the radiologist’s digital pile of scans to be read.
For scans that meet certain criteria, the Clario system would call CuraCloud’s A.I. module, triggering an automated evaluation of the incoming image file. (The module would likely reside within the radiology practice, rather than in the cloud, because of the large file sizes and need for rapid evaluation, though it could also be deployed in the cloud, Butler says. The major public cloud providers are in a race to add more capable machine learning applications to their menu of services.)
Butler says CuraCloud’s team of computer scientists and diagnostic imaging equipment experts have built a deep neural network—a type of machine-learning algorithm—that can classify radiographic images. Each A.I. module is tuned to detect a specific condition—intracranial hemorrhage, for example.
CuraCloud has had no trouble obtaining data to train its triage algorithms, Butler says. There are public-domain datasets of radiologic images that can be used, as well as “any number of organizations that are very interested in collaborating with us and sharing their data in a controlled way, of course,” he says.
That said, there is a data land-grab going on in the business of A.I. for healthcare, with lots of questions swirling around.
“It comes down to who has access to it, who owns it, and who has rights to it—and can you use it, under what conditions can you use it?” Wood says, adding that definitive answers are few.
Companies like Clario that are already working with care providers—its customers process 1.3 million exams a month, creating a trove of tagged training data—would seem to be at an advantage. One challenge for the industry is establishing the right legal and business permissions to use the data, but once that’s resolved, the data will flow, allowing machine learning applications to continually improve, Wood says.
Another challenge is regulatory compliance. Digital innovation in healthcare, particularly something that impacts the use of regulated medical devices such as radiographic scanners, must first pass muster with the FDA. But Clario and CuraCloud think their first application in triage—despite having a major potential impact on patients—will have “a relatively low regulatory hurdle,” Wood says.
“It’s not like you’re really changing anything other than how fast it’s read. But then again, if you’re that person with a pulmonary embolism, the difference between five hours and 15 minutes—it could be life or death,” Wood says.
Applications in which the creator of an A.I. system makes a specific, numerical claim about a system’s efficacy or performance—it will automatically detect a certain percentage of tumors, say—will likely trigger at least the FDA’s 510(k) clearance process to determine whether a medical device is equivalent to something else already on the market, or the more-rigorous premarket approval (PMA) process for medical devices, potentially including clinical trials. That would be a significantly more expensive regulatory pathway.
“The companies that have the biggest budgets will probably go for the PMA so they can get the strongest claim,” Wood says. “A lot of the startups out there will not spend the money to get a giant, strong claim.”
He expects a scattershot approach until the FDA provides guidance on how it will regulate A.I. technologies in healthcare, as part of a broader focus on digital health.
The regulatory requirements—and therefore the costs—of developing A.I. systems for healthcare raise another question: Who’s going to pay for them?
In the U.S., there’s no billing code for triaging radiology scans, Wood notes. “The reason people will buy this is to just provide better patient care,” he says.
And that could be the key: Large, private radiology practices compete for hospital contracts. “The one that has A.I. might win the contract because they can talk about prioritization,” he says. “And they can talk about how it results in better patient care. So all other things being equal, that will help a private practice win business. … [Radiologists] won’t get paid a dime more for using this technology, so it’s only this competitive nature of the practices that really makes it possible for us to even go to market right now.”
CuraCloud, for its part, has more ambitious plans down the road. The company was formed in late 2015 and includes a team of longtime collaborators—many originally from China who earned computer science PhDs from U.S. universities—led by CEO Qi Song, Butler says. The company’s road map includes pre-populating radiology reports and computer-aided detection and diagnosis—an area where Seattle has something of a pedigree, adding to its core strengths in cloud computing, A.I., biomedical research, and medical devices.
Seattle-based Confirma, for example, was a leader in the development of computer-aided detection software in applications including mammography, and Wood was previously its vice president of research and development. Merge Healthcare acquired Confirma in 2009 for about $22 million. Merge was scooped up by IBM in 2015 for $1 billion and incorporated into the IBM Watson Health business.
Clario is also looking at applications in pathology, which is undergoing its own digital transformation akin to radiology’s transition away from film. In April the FDA permitted Philips to market a system that enables pathologists to review and interpret digital surgical pathology slides from biopsied tissue—a first for the industry.
“Because the system digitizes slides that would otherwise be stored in physical files, it also provides a streamlined slide storage and retrieval system that may ultimately help make critical health information available to pathologists, other health care professionals, and patients faster,” said the FDA’s Alberto Gutierrez in a news release announcing the approval.
Of course, it’s still a long road from the triage application CuraCloud and Clario are working on now to an A.I. radiologist that can diagnose and write reports.
Wood and Butler diverge on the timeline by which such a system could be realized, with Butler being more bullish.
And, if that system does arrive, will insurers or the Centers for Medicaid and Medicare Services, which sets reimbursement policies for a large share of U.S. healthcare spending, pay a computer for a service that used to be performed by a human—assuming that it was at least equal to the quality provided by a human? “My guess is they won’t,” Wood says.
But if an A.I. radiologist could competently read routine chest X-rays as well as or better than a human for half the fee—admittedly, a big if—“then obviously the payers are going to love it,” he adds.
That would suggest a bleak future for radiology jobs. Add it to the list of occupations across the economy from truck drivers to journalists to accountants threatened by A.I. Not so fast, say Butler and Wood.
Demand for radiology is growing, driven in large part by demographics, they argue. “People are getting older, especially the baby boomers, and you get a lot of images later in life,” Wood says. There’s also demand from outside the U.S.; many countries have far fewer resources to train sub-specialists in radiology and related fields, Butler says.
Data compiled by the Harvey L. Neiman Health Policy Institute show the number of radiologists in the U.S. increasing significantly on a per-capita basis, from 10.6 per 100,000 people in 1995 to 12.3 per 100,000 in 2013.
Wood thinks automated evaluation of routine radiographic scans would not only increase efficiency, it would also improve job satisfaction among radiologists who are sometimes “a little bored by the easy stuff.”
Butler points out that there’s no radiologist at most dentists’ offices, but dental X-rays are routinely performed and interpreted there.
He says radiologists would just as soon pass the quotidian broken arms and chest X-rays off to a competent machine because “they don’t get paid much for it, and they’re not practicing at the top of their license to look at some of these routine things that even a computer could figure out.”
CuraCloud is not out to replace radiologists, he adds. “Our goal is to help them work faster, more efficiently, with higher quality,” he says. “We’re going to go for those things that have the highest value to the radiologist and to the broader healthcare delivery system.”