Teaching a machine to see: Italian doctors turn to Chinese A.I. to diagnose COVID-19

Each day, overworked Italian medical technicians upload detailed images of Italians’ lungs to a computer server that sits in an underground radiology lab in Rome. As Italy’s coronavirus crisis has worsened, the volume of scans has climbed into the thousands. 

In a perverse way, the rise in numbers is just what the Roman lab is looking for. That’s because inside the server is an A.I.-powered diagnostic tool that, its developers say, can detect blazingly fast if a person has coronavirus—in less than a minute. The more images the server gets, the more help it can ultimately be to Italian doctors on the front lines of the deadly outbreak.

The technology, called “Infer-Read,” was developed by InferVISION, a Chinese startup that generated international attention at the height of the Wuhan coronavirus outbreak this winter. Rome’s 400-bed Polyclinic Bio-Medical University Hospital, where the lab is located, is the first in Europe to license and put the Chinese technology to work to help fight the contagion.

Italy’s experience offers a glimpse of the potential—and limitations—of A.I.’s role in fighting the coronavirus outbreak. While the Chinese system has failed to live up to some early promises of enabling diagnosis without human intervention, it has accelerated the diagnosis process in potentially life-saving ways.

“This new kind of system doesn’t do the job on its own, but it can serve as a very important tool to help a skilled radiologist,” said Bruno Beomonte Zobel, coordinator of the imaging center at Polyclinic Bio-Medical University Hospital. “If we get the images for, let’s say, 500 suspected cases of coronavirus, the system will be very good at determining that only 200 of them merit close inspection by a radiologist. That kind of time savings is extraordinary.”

There’s a lot on the line. As of Friday morning, COVID-19, the respiratory disease caused by the coronavirus, had killed nearly 14,000 Italians, the highest death toll of any country. Meanwhile, total confirmed cases across Italy have topped 115,000, and health officials warn the actual number could be far higher as undiagnosed cases continue to go untested.

Coronavirus, or something more ordinary?

Earlier this year, Chinese doctors started using InferVISION’s technology to help overwhelmed hospital staff more rapidly determine whether a patient arriving at the emergency ward with a nasty cough was actually suffering from the highly contagious coronavirus rather than, say, a more ordinary case of pneumonia or bronchitis.

The InferVISION system is based on machine learning. It relies heavily on image-recognition algorithms that examine roughly 600 digital images of a single patient’s lungs to make such a diagnosis. Doctors feed it a standard CT, or computerized tomography, scan—the kind available in most any medical clinic in an industrialized country. Outside China, medical experts caution against relying too heavily on CT chest scans for coronavirus diagnoses until further efficacy studies are conducted. The Chinese though are bullish on the practice— as they are on the role of A.I. to successfully flag the disease.

The InferVISION software is designed to read the chest scans and flag which patients appear to be infected by the coronavirus. It makes its determination within 20 seconds, as opposed to hours or days for conventional testing. A big selling point for the Italians was that the system could make a diagnosis directly from the scans—no need for a technician to be on hand. (Ultimately, the Italians decided it was best to give the technology human supervision.)

In Italy, where more than 60 doctors and medical professionals have caught the coronavirus and died, any time-saving advancements could be a life-saver. The usual testing method relies on a swab in the nose or throat. It can produce a diagnosis in about a day if the testing is done on-site, but a potential excruciating three-day wait is more typical. 

I learned that from experience. I was in Milan, a hotbed of infections, just before Italy issued its nationwide lockdown. A few days later I started to run a fever and had developed a light cough. A virologist I interviewed persuaded me to get a test—telling me how to get it done quickly. What followed were three sleepless nights. 

The test came back negative. But I would have gladly opted for anything faster had such an option been available then.

Training to see

According to Beomonte Zobel, the Chinese technology has limits. For starters, the diagnostic tool had to be trained on lung scans from thousands of infected Italians in order to build a reliable diagnostic model for the larger Italian population. That would suggest such a tool would only be effective in countries where they’re testing aggressively and the infection numbers are high.

And, it’s not an off-the-shelf solution. What the system learned in Hubei Province about infected Chinese doesn’t completely apply in Italy. Only after the tool analyzed repeated uploaded images of suspected Italian cases could it spot with any reliability which of the chest scans showed signs of coronavirus infection. Furthermore, what it learns in Italy might not give doctors in other countries much of a head start in diagnosing coronavirus-infected lungs there.

“We couldn’t rely on what the system ‘learned’ in China, because Italians have different [physical] characteristics, and are on average at least a decade older than the Chinese,” Beomonte Zobel told Fortune, explaining those differences were substantial enough for the Italian doctors to conclude they needed to train the tool on local cases to ensure a higher degree of accuracy in diagnoses. “The virus also mutates, and so it is slightly different than the one in China.”

It took the better part of a month to overcome that learning curve, and now Beomonte Zobel’s team is quickly scaling up. He envisions the lab will be able to handle “a few hundred” tests per day from anywhere in the country. It’s currently fielding about 150 daily. That somewhat low number is more the fault of Italian red tape associated with data protection and privacy rules, he said, than any glitch in the technology.

“Sometimes the Italian bureaucracy can be as complicated as the coronavirus,” Beomonte Zobel joked.

Technology under the spotlight

The rapid spread of the coronavirus outbreak is generating vigorous debate about the proper tools needed to fight the contagion. A series of studies, for example, have emerged in recent weeks questioning the effectiveness of using CT scans alone to make an accurate coronavirus diagnosis. The American College of Radiology, for example, warned its members that chest scans of Covid-19 “are not specific, and overlap with other infections” while acknowledging that, nonetheless, “some medical practices are requesting chest CT to inform decisions on whether to test a patient for COVID-19, admit a patient or provide other treatment.” They conclude: “The ACR strongly urges caution in taking this approach.” 

Beomonte Zobel is aware of these doubts. That’s why, he says, the final call must still be made by a trained radiologist. However, he noted, the Italian experience using the “Infer-Read” system has already sped up diagnoses tremendously.

More tests needed

The race to develop reliable and speedy Covid-19 testing systems has taken on new urgency in recent weeks as the number of cases surpasses 1 million worldwide. The U.S. Food and Drug Administration fast-tracked a new system developed by Abbott Laboratories last week that promises to detect positive cases within five minutes, and negative cases within 13 minutes. The Abbott technology is a portable device that works differently from the InferVISION apparatus—it doesn’t rely on chest scans. Rather, it’s a molecular testing kit that seeks to detect the presence of a fragment of the coronavirus genome in the patient from a nasolaryngeal swab. Another, developed by Henry Schein Inc., relies on blood samples and promises results within 15 minutes.

Frontline doctors are relieved to see any kind of innovation to speed up detection if it means saving lives.

Back in Rome, Beomonte Zobel said nobody is asking a diagnostic tool such as the A.I.-based InferVISION system to become the sole line of defense against the outbreak in Italy, which is still seeing more than 2,000 new cases a day.

“This isn’t going to turn the tide,” he said. “It will speed up the diagnosis for the patients it sees. But the main benefit isn’t for this moment. It is as a step in a process that will leave us better prepared for the next pandemic.”

More must-read coronavirus coverage from Fortune:

—Everything you need to know about the coronavirus stimulus checks
—Social distancing creates $8 trillion in economic benefits, study says
—Can’t afford to buy an apartment? This startup wants to sell you a share of one
—The lessons learned from the past 3 bear markets
—Listen to Leadership Next, a Fortune podcast examining the evolving role of CEO
—WATCH: U.S. tax deadline moved from April 15 to July 15

Subscribe to Fortune’s Bull Sheet for no-nonsense finance news and analysis daily.