As new artificial intelligence (AI) tools roll out for cancer detection, oncologists are trying to figure out how the technology will affect their practices and if it will improve care.

One third of oncologists said that they found AI helpful in diagnosing or staging cancer, according to a recent survey of 140 oncologists by Medscape. Slightly more (35%) reported that they did not consider the technology helpful and 32% were on the fence.

There were wide variations in the types of cancer where AI tools were viewed as helpful. Among respondents who found AI useful, the most cited cancer types where AI adds value were breast, lung, and prostate cancers, while brain and kidney cancers were at the bottom of the list.

The findings line up with the current uptake of AI tools across oncology, where most of the focus has been on image-based modalities, whether that is assistance in reading mammograms or lung CT scans or assessing laboratory results in prostate cancer screening or colonoscopy images for colon cancer, said Ravi Parikh, MD, associate professor of hematology and medical oncology at Emory University and the medical director of Data Science at Emory Winship Cancer Institute, Atlanta.

The current tools are mainly aimed at improving the efficiency of the physician and are employed as secondary to the clinician’s role, he said.

“The challenge has largely been that very few AI-based technologies have been prospectively validated and studied to see their impact long-term, not just rates of diagnosis, but clinical outcomes,” Parikh said. “It takes a long time to run cancer screening trials; you’re usually measuring outcomes over years and decades.”

Studies will need to demonstrate whether AI tools can produce the same or better results than clinicians and if they are missing or over diagnosing cancers. “We just need evidence to say whether that’s happening or not before doctors are going to fully rush to using or relying on these tests,” he said.

Acceptance and Caution

Opinions on the value of AI are still divided among clinicians, said Kun-Hsing Yu, MD, PhD, an associate professor in the Department of Biomedical Informatics at Harvard Medical School, Boston. 

Yu said he finds AI helpful for detecting subtle signals that might be missed during a first look at a patient. But he said he understands why others remain skeptical. “None of these AI tools have perfect performance and especially for life-and-death clinical decisions, we shouldn’t trust AI completely before it has undergone sufficient validation,” he said.

In the current clinical workflow, Yu said, it is essential for physicians to review the AI output and explain the findings to individual patients. 

Tufia C. Haddad, MD, a medical oncologist and co-leader of the Office of Platform and Digital Innovation at the Mayo Clinic Comprehensive Cancer Center in Rochester, Minnesota, said she sees clinician acceptance of AI technology increasing, in part because of an understanding that simply ignoring it is not an option. 

“I think there’s a recognition that we have to learn and understand the limitations,” she said.

Acceptance has also been aided by increased transparency about the data used in AI models and how the datasets were developed and validated. However, Haddad said she continues to have concerns about ensuring the completeness and diversity of data used to train AI models to ensure that the feedback from AI tools is applicable to diverse populations of patients. Data governance — who can access the data and the potential for re-identification of personal medical information — is another significant concern as the technology is further developed, she said.

What will ultimately drive adoption as AI tools move into the cancer treatment space are pragmatic trials where AI models are evaluated in clinical practice, Haddad said. “People want to see the evidence,” she said. “Is this model really reducing cost of care or time saving or better at predicting toxicity? People want to see the evidence that this model is actually functioning as suggested.”

Deskilling

Another concern about the roll out of AI tools for cancer diagnosis is that they could lead to an erosion of skills by clinicians. One recent study, published in The Lancet Gastroenterology & Hepatology, found evidence of “deskilling” among endoscopists who performed some of their procedures with AI assistance. The retrospective study reviewed data from four endoscopy centers in Poland that had introduced AI tools for polyp detection in 2021. The study found that the rate of adenoma detection during standard, non-AI assisted colonoscopy fell significantly from 28.4% before AI exposure to 22.4% after AI exposure.

Yu acknowledged that deskilling is a legitimate concern, but he is optimistic that in the near future AI tools can make up for this loss by making significant advances in medical diagnosis and treatment beyond what is currently possible.

He offered a calculator analogy. When calculators were first introduced, many people raised concerns that students would lose their ability to do simple arithmetic by hand. While many people did lose the ability to do simple math calculations quickly, the invention also freed people to focus on higher-level problem solving.

“I view the role of AI in cancer diagnosis and staging as that of an advanced calculator,” Yu said.

In the future, medical students and junior clinicians who have always had AI tools to assist with a diagnosis could become dependent on these systems. However, emerging AI tools are poised to serve as standard copilots in routine diagnostic tasks, while simultaneously harnessing large datasets to advance biomedical discovery and uncover new treatment strategies, Yu said.

“In the future, it will not be a matter of AI replacing clinicians in diagnostic tasks, but rather that clinicians unable to harness AI will be replaced by those who understand and can use it effectively,” Yu predicted.

Future Applications

Haddad said she is most excited about efforts to use AI to make sense of multimodal data, bringing together imaging and pathology data with electronic health record information and genomics data, for instance. As more patient data is collected, through wearables and other sources, and harnessed by AI, there’s the potential to make personalized predictions and recommendations, she said.

Understanding an individual’s risk, rather than just the population risk, could lead to earlier cancer detection and more proactive care.

“It’s going to allow us to hyper personalize medicine,” Haddad said. “We say we’re in this era of precision medicine, but we are just dipping our toe in the water with the opportunity that’s forthcoming.”

New machine learning models are increasingly capable of identifying patterns too subtle for clinicians and researchers to recognize, Yu said.

Yu and his colleagues at Harvard Medical School recently developed a machine learning model that can predict a tumor’s molecular profile based on cellular features seen on histopathology images. In a recent study published in Nature, the team demonstrated that the model could also predict a patient’s overall survival and disease-free survival after treatment using only samples received at the time of initial diagnosis.

Using AI to better understand how patients will respond to treatment has the potential to spare patients with cancer from unnecessary toxicity, he said.

“We believe this could represent the next wave of AI, augmenting clinical practice rather than just replicating what humans already know,” Yu said.

Yu reported receiving research funding from the National Institutes of Health, the Department of Defense, and the American Cancer Society, and consulting for Takeda and Curatio DL. He reported being the inventor of US Patent 10,832,406 (assigned to Harvard University). Haddad and Parikh reported no relevant financial relationships.