aamc.org does not support this web browser.
  • AAMCNews

    Will artificial intelligence replace doctors?

    Several new studies have shown that computers can outperform doctors in cancer screenings and disease diagnoses. What does that mean for newly trained radiologists and pathologists?

    medical doctors viewing scan

    “Do you think we’re gonna be replaced?”

    A young Johns Hopkins University fellow recently asked that question while chatting with Elliot Fishman, MD, about artificial intelligence (AI). The two men were on the opposite ends of the career spectrum: Fishman has been a professor of radiology and oncology at Johns Hopkins Medicine since 1980; the fellow was preparing for his first job as a radiologist.

    “I said, ‘Well, I think it’s going to change what we do, but the good news is, at least you’re not a pathologist,’” Fishman recalls. “And he goes, ‘My wife is just graduating and she’s a pathologist.’ So I said, ‘Put away as much money as you can really fast.’”

    Fishman laughs when he tells the story, but he understands the concern. Over the past few years, many AI proponents and medical professionals have branded radiology and pathology as dinosaur professions, doomed for extinction. In 2016, a New England Journal of Medicine article predicted that “machine learning will displace much of the work of radiologists and anatomical pathologists,” adding that “it will soon exceed human accuracy.” That same year, Geoffrey Hinton, PhD, a professor emeritus at the University of Toronto who also designs machine learning algorithms for Google (and who received the Association for Computing Machinery’s A.M. Turing Award, often called the Nobel Prize of computing, in 2019), declared, “We should stop training radiologists now."

    The reason for the predictions? AI’s tantalizing power to identify patterns and anomalies and to examine “pathologies that look certain ways,” says Fishman, who is among the enthusiasts: He’s studying the use of AI for early detection of pancreatic cancer.

    “The hope is that if we could pick up early tumors that are missed, we would have better outcomes,” he says.

    An array of studies have offered glimpses of AI’s enormous potential. In a study published by Nature Medicine in May 2019, a Google algorithm outperformed six radiologists to determine if patients had lung cancer. The algorithm, which was developed using 42,000 patient scans from a National Institutes of Health clinical trial, detected 5% more cancers than its human counterparts and reduced false positives by 11%. False positives are a particular problem with lung cancer: A study in JAMA Internal Medicine of 2,100 patients found a false positive rate of 97.5%.

    Furthermore, AI performed comparably to breast screening radiologists in a study in the March 2019 Journal of the National Cancer Institute. At Stanford University, computer scientists developed an algorithm for diagnosing skin cancer, using a database of nearly 130,000 skin disease images. In diagnostic tests, the algorithm’s success rate was almost identical to that of 21 dermatologists, according to a study published in Nature in 2017. In another skin cancer study, AI surpassed the performance of 58 international dermatologists. The algorithm not only missed fewer melanomas, but it was less likely to misdiagnose benign moles as malignant, the European Society for Medical Oncology found.

    “Machine intelligence presents us with an opportunity to significantly improve the delivery of health care, particularly in high-disease or low-resource settings.”

    Sameer Antani, PhD
    PhD National Library of Medicine

    The possibilities extend beyond cancer. Recent studies have shown how AI can detect rare hereditary diseases in children, genetic diseases in infants, cholesterol-raising genetic diseases, and neurodegenerative diseases, as well as predict the cognitive decline that leads to Alzheimer’s disease. A variety of companies are offering FDA-approved AI products, such as iCAD’s ProFound AI for digital breast tomosynthesis. Israeli-based Aidoc has received three FDA approvals for AI products, the latest occurring in June 2019 for triage of cervical spine fractures. In 2018, the FDA approved Imagen’s OsteoDetect, an AI algorithm that helps detect wrist fractures.

    So when will AI become an everyday tool for diagnosis? It might not happen as fast as people think, Fishman believes, but it’s definitely coming.

    “If you think it’s just some passing fancy, you’re making a mistake,” he says. “In radiology and pathology, it’s going to affect everything you do.”

    The upsides of AI

    AI has huge potential but it’s still in its infancy as a diagnostic tool. “There are a lot of rapid advances being made and a rush for monetizing them,” says Sameer Antani, PhD, acting branch chief for the National Library of Medicine’s Communications Engineering Branch and Computer Science Branch, who leads a research team on AI in such areas as cervical cancer screening and rare diseases. Initially, AI will be most effective for dealing with specific problems. “A program that does the whole CT of the abdomen is going to take a while,” says Fishman. “There are so many organs and there’s so much variability. But for specific tests that examine individual organs — the liver, the kidneys, the pancreas, the lungs, the heart — I think that’s where AI is going to be strong. And it’s happening.”

    AI could be particularly beneficial in places with limited access to health care. “Machine intelligence presents us with an opportunity to significantly improve the delivery of health care, particularly in high-disease or low-resource settings,” says Antani. Consider a study published in March 2019 by the American Academy of Ophthalmology, which found that a Google algorithm improved doctors’ ability to accurately diagnose diabetic retinopathy. The algorithm has been tested in India, which is the type of country that could benefit from AI screenings, since it suffers from a shortage of doctors and ophthalmologists. At Stanford, researchers believe their skin cancer algorithm could work on a smartphone, allowing people to screen themselves.

    AI could also help reduce radiology’s 30% error rate. “Our goal is to be perfect, but that’s unreasonable. People are busier than ever, they’re reading more studies, they’re working harder — errors increase with faster speed,” says Fishman. And algorithms can handle a larger workload than humans. The Google skin cancer program looked at about 130,000 images. A dermatologist looks at about 12,000 in his or her lifetime, Fishman notes.

    “Computers can look at cases 24/7 and they can keep learning,” he says. “It has the opportunity to be the best helper you could ever imagine.”

    Data bias and lost jobs

    For all of its upsides, scientists such as the late Stephen Hawking have warned that artificial intelligence could destroy mankind. At Harvard Medical School’s 2019 Precision Medicine conference, Harvard Law School professor Jonathan Zittrain compared AI to asbestos: “It turns out that it’s all over the place, even though at no point did you explicitly install it, and it has possibly some latent bad effects that you might regret later, after it’s already too hard to get it all out.” He also noted that AI can be tricked, according to a story from Stat, citing a Google algorithm that correctly identified a tabby cat. When some pixels were changed, the algorithm thought the kitty was — no joke — guacamole.

    The newness of AI applications presents a challenge for regulatory agencies to measure and validate its performance in medical diagnostics, Antani says. In April 2019, the FDA announced plans to develop regulations focusing on medical AI products that adapt based on new data.

    Another issue is the data itself. The promise of AI depends on “the availability, quality, and completeness of training data and design of the AI framework/algorithm,” says Antani. Factors that affect data strength include social, geographic, or economic biases, as well as simply acquiring data. Computer scientists are developing AI architectures that produce compelling results with less data, Antani says, but “while these are exciting technological advances, they still do not address the shortcomings.”

    Efforts to improve data diversity include Count Me In, a cancer project involving more than 6,000 patients so far. Count Me In collects medical information — from tumor samples to blood and saliva — and compiles it in databases for researchers. Launched in 2018, Count Me In is a collaboration between the Broad Institute of Massachusetts Institute of Technology and Harvard, the Biden Cancer Initiative, the Dana Farber Cancer Institute, and the Emerson Collective (founded by Laurene Powell Jobs, the widow of Apple co-founder Steve Jobs). Its efforts include the Metastatic Breast Cancer Project, which has received tumor samples and medical information from over 4,800 patients.

    Count Me In is not only providing its data to researchers for free, but it’s collecting it from across the United States. “If you’re at Hopkins, or you’re at Mass General, your population is Boston or Baltimore. There could be some baked in biases in your population, for good or bad,” says Fishman. “So you need wide data and this is a way of potentially collecting data that’s unbiased.”

    “I’m not exactly worried that they’re going to put us out of business right away. Way back when, pathologists looked at every pap smear. Now a machine looks at them and once in a while it kicks something out. Things change.”

    Elliott Fishman, MD
    Johns Hopkins Medicine

    But does better data simply hasten the loss of jobs for health care professionals? Will radiologists and pathologists be replaced by smartphone apps or a medical voice assistant? AI has the potential not only to be more accurate, but to work faster than humans.

    “I’m not exactly worried that they’re going to put us out of business right away,” says Fishman. “Way back when, pathologists looked at every pap smear. Now a machine looks at them and once in a while it kicks something out. Things change.”

    That change, while potentially profound, does not mean that radiologists and pathologists are destined for the unemployment line. Experts also predicted the demise of radiologists when MRI machines were introduced, wrote Curtis P. Langlotz, professor of radiology at the Stanford University Medical Center, in a May 2019 editorial for the journal Radiology. “Radiologists are being trained to recognize AI’s shortcomings and capitalize on its strengths,” he wrote, adding that most comparisons between algorithms and radiologists are too simplistic. “An AI algorithm that diagnoses common chest conditions at the level of a subspecialty thoracic radiologist is a major step forward, an incredible asset to underserved regions, and could serve as a valued assistant for a subspecialty radiologist.” But radiologists are also trained to detect less-common diseases such as rheumatoid arthritis and sickle cell disease. “AI is impressive in identifying horses,” he wrote, “but is a long way from recognizing zebras.”

    How will the job change? Over the next 5 to 10 years, the most successful radiologists and pathologists will be those “who are well equipped and eager to participate in data management and integrated diagnoses,” says Frank J. Rybicki, MD, PhD, vice chair of operations and quality with the University of Cincinnati Department of Radiology and medical director of Imagia Cybernetics Inc. When an algorithm produces unexpected results, radiologists will need to understand why. Fishman believes this will lead to a deeper role in patient care — which is one reason why he’s more excited than threatened.

    “If you ask me who will benefit from AI, it’s the patients,” says Fishman. “That’s why I’m so excited. Better care for our patients. What can be better than that?”