Assessing AI
By Beth W. Orenstein
Radiology Today
Vol. 22 No. 8 P. 12

Experts say AI will not replace radiologists, but it can augment their workflow and promote better patient care.

AI may have gotten off on the wrong foot with some radiologists. In 2016, Geoffrey Hinton, PhD, known as the godfather of deep learning, said it was time to stop training radiologists because, within five years, AI would be doing their jobs better than humans. (That same year the FDA cleared the first model, an algorithm exposed to training data, for medical imaging.)

“We had a lot of residents asking us if this were true, because no one knew back then,” says Hari Trivedi, MD, an assistant professor of radiology and biomedical informatics and codirector of the HITI Lab at Emory University in Atlanta. Hinton has since retracted his controversial prediction, Trivedi says, but it has never been totally forgotten and, as a result, some radiologists still look askance at AI.

By now, most radiologists recognize that AI is not going to replace them—or almost any physician, says Bradley J. Erickson, MD, PhD, a radiologist at the Mayo Clinic in Rochester, Minnesota, and former chair of the American Board of Imaging Informatics. And, while some radiologists are making use of AI in their clinical practices, it’s more for the parts of their job they like least than for headline-grabbing tasks.

In April 2021, the ACR published the results of a member survey that was designed to understand how radiologists are using AI in clinical practice. From its more than 1,800 responses, the ACR found that 30% of radiologists currently use AI as part of their practice. Large practices were more likely to use AI than smaller practices. And, of those that use it, most do so for specific tasks such as the detection of intracranial hemorrhage, pulmonary emboli, and mammographic abnormalities.

Still, AI in radiology gets a lot of attention in the lay media and medical journals. It makes for big news whenever a study involving imaging finds that AI outperforms radiologists, Trivedi says. Over the past several years, “you see lots of headlines: ‘AI outperforms radiologists for such-and-such a task.’ There are more and more examples here and there.”

The headlines, Trivedi says, are somewhat misleading. They may give the impression that AI is better than radiologists at reading images and diagnosing patients, but all these studies really show is that AI may be slightly better at one particular task than radiologists reading the same images performing the same task. “The headlines are not technically wrong,” Trivedi says. “If you take an AI model trained on one specific task on one type of image, there are many models out there that perform as well, if not slightly better, than an aggregate opinion of a group of radiologists.”

What gets missed is that any given imaging test has dozens, if not hundreds, of potential diagnoses, Trivedi says. An AI model can be trained to look for a particular diagnosis, but radiologists don’t just focus on one particular diagnosis. “As a radiologist, if you’re given a chest X-ray, you’re not just looking for pneumonia, but you’re also diagnosing air trapping and pulmonary edema and fractures—all these kinds of things—that the AI model may not,” he says. That’s the same thing he would tell the residents who asked whether Hinton’s prediction was true. “We’d tell them that the job of a radiologist doesn’t end with a specific diagnosis on an image. In fact, it doesn’t end with the image alone.”

Also, Trivedi notes, in most of these studies, the difference in outperformance between AI and radiologists tends to be rather small. “Most of the studies that are grabbing headlines that I’ve seen are within a few points,” he says. “The study may show radiologists at 85%, where the AI models are 88% or 90%. That’s not a huge difference.”

Specific Use Cases
Tessa S. Cook, MD, PhD, CIIP, an assistant professor of radiology at the Perelman School of Medicine at the University of Pennsylvania and a member of Radiology Today’s Editorial Advisory Board, agrees that there has been a great deal of hype about AI in radiology in recent years. Clearly, AI has shown the potential to help radiologists do their job, she says; however, Cook seconds the opinion offered by Trivedi: “AI in radiology is best in very specific use cases.” For example, it can take a chest CT and find pulmonary nodules or pneumothorax or pulmonary emboli. But, she says, that’s not the same thing as saying, “‘OK, AI: Read that chest CT and interpret the whole thing.’ We’re not there yet and far from it.”

Most of the use cases are in chest CT and brain imaging looking for stroke, hemorrhage, and large vessel occlusion, which is one that actually has some funding, ie, reimbursement, behind it, Cook says. “The use cases that have been developed and implemented are the ones where there was enough training data accessible to the developers,” she notes.

In retrospect, Trivedi says, it is not surprising that AI has been shown to be better at certain tasks than radiologists. “It stands to reason, if you took a machine and trained it for three months or a year just to diagnose pneumonia on X-rays and it has thousands and thousands of X-rays at its disposal, it would probably perform better at pneumonia diagnosis than most radiologists,” he says. But radiologists can do something machines can’t: integrate the findings with the patient’s demographics and medical history.

“There are no models out there that can adequately synthesize all the information in a patient’s chart plus all the imaging to come up with an appropriate diagnosis to the point where we don’t need the radiologist, or any physician for that matter,” Trivedi says. “There’s a lot more nuance that goes into the job, and we’re really far away from being able to do something that is that comprehensive with machines at this point.”

“Making the correct diagnosis in most of radiology still requires the integration of nonpixel data such as the age, gender, symptoms, and medical history,” Erickson notes. And that’s something that radiologists still can do better than AI models, he adds.

Besides, surgeons who plan to operate on patients, based at least in part on radiologists’ diagnoses, are going to feel more confident going in if they can question the radiologists about their opinion directly, Trivedi says. “No model is 100% accurate. No person is 100% accurate,” he explains. “As a surgeon, you have to assume that the model or the human is going to be incorrect in some cases. If you don’t have any way to interrogate the model results when you disagree, it really erodes your confidence. As a surgeon, even if you know the model is 99% accurate, you still have to think, ‘How do I know this isn’t the 1% of cases where it may be wrong?’” A surgeon can discuss the findings in depth with the radiologist who read the images, but not the model, he says.

Additional Roles
AI can and does play important roles in radiology other than diagnosis, Erickson says. For example, it can be highly useful in managing workflow—known as intelligent process automation (IPA). At the very least, Erickson believes radiologists ought to embrace AI for this part of their job.

“Many of the steps that humans do today could be done better and faster if computers assisted their orchestration,” Erickson says. “Human tasks can be integrated into these workflows, and the workflow system can recognize when a human task has been waiting too long or when there are too many or too few entries on a work queue. AI tools could improve efficiency and quality at the same time.” IPA also can assist in the reliable collection of data both for training AI tools and for collecting the results of the prediction “so that monitoring performance becomes much less effort.”

Trivedi agrees that AI tools “can help radiologists to free up their time for things that are more challenging or more important.” He, too, expects that AI’s role in patient management will continue to grow. In the ACR survey, another 20% of those radiologists who are not currently using AI said they plan to buy and use AI tools in the next one to five years.

Eliminating Bias
One problem with AI that the radiologists would like to see addressed, if not solved, is bias in the data used to build models. “We know that our data are biased because the care received by patients is biased,” Trivedi says. “But we need to ensure that our data are unbiased and that the models we’re using perform equally well across stratifications of the data, whether it be race or age or gender or ZIP code. That’s a really important area, and we are just beginning to scratch the surface.”

For example, “we know that African American women across the country and in our practice are diagnosed with breast cancer at a later stage than white women, which means they are likely to have a worse outcome,” Trivedi says. Possible reasons are that Black women are not getting mammograms as often as they should or, for some reason, Black women have more aggressive cancer due to genetic factors or they have decreased access to care. Whatever the reason, the data used to train AI models to look for breast cancer must reflect these intrinsic disparities, Trivedi says. “It’s important to uncover these disparities so that when we train the model, we can try to balance out the data such that the disparity is decreased.”

Supposing one-half of his patients are Black and one-half are white, if Trivedi were to load their information into an AI model and train it, based on race alone it would have a higher likelihood of diagnosing breast cancer in his Black patients because, statistically speaking, they have a higher likelihood of breast cancer. “But that’s not the right way to do it,” he says. The model would rely on the statistical prevalence of the disease and find more cancers in the Black women, when it should be relying on the imaging. To remove the bias, the data used to build the model need to be adjusted. 

“We would have to remove some white patients with low-grade disease and add patients with higher-stage disease, as well as remove some African American patients with higher-stage disease and augment with ones with lower-stage disease, then use those data to train the model,” Trivedi continues. “This way, what the model sees is not that the African American patients have worse cancer, and that bias is removed from the data.”

Continual Training
Cook says it’s also important that AI models be retrained periodically. “You might get new imaging equipment or do imaging studies with different protocols; your patient population may change,” she says, adding that any of these things could impact the performance of the model. “You need to watch and pick up those instances and adjust and adapt.” A model’s performance, once it’s trained, is inevitably going to degrade because of these factors, Cook says. “And we radiologists, or any clinician user, need to be sensitive to that.” 

“A big concern that is growing now is that AI tools are very specific to the data used in training them,” Erickson says. “We hoped that by collecting data from many institutions and diverse populations that we would not have this ‘generalization’ problem. It appears now that might have been overly optimistic.” In fact, he says, one vendor of AI, though not one in the imaging space, recommends that users retrain their AI every six months on their data, to avoid the data-drift problem.

AI will continue to expand and provide more detection tasks, Erickson says. At the same time, it will help with doing more quantitation. “I also see more applications where images are used for predicting molecular properties, which I find very exciting,” he says. “If we can predict molecular properties of tumors with high accuracy, we will become central players in precision medicine. It also will be used to predict the risk of other diseases, such as cardiovascular disease and stroke.”

Cook believes most radiologists are well beyond Hinton’s prediction about AI replacing radiologists. “At the risk of sounding like a teenager,” she says, “that’s so 2016. I really think that most radiologists, I hope in most cases, have gotten past seeing AI as a threat to their jobs and moved on to seeing it as another tool we can use for patient care, one we can be very involved in getting and developing for better patient care.”

— Beth W. Orenstein of Northampton, Pennsylvania, is a freelance medical writer and regular contributor to Radiology Today.