Study: How Adding AI May Improve Mammography Screenings

While breast cancer rates, particularly in younger women, have been increasing since 2012, the number of radiologists in the United States is on the decline, and getting quick and accurate mammography reports is essential to optimal outcomes. Two information systems associate professors at The University of Texas at Dallas set out to evaluate if artificial intelligence (AI) might be an economically feasible way to address this gap.

Along with their colleagues from other universities, Drs. Mehmet Ayvaci and Radha Mookerjee determined that AI can be a useful tool to flag high-risk breast cancer cases quickly for a radiologist to review, saving up to 30% in health care costs, Ayvaci said, compared to relying solely on human expertise. The researchers, who are both on the faculty of the Naveen Jindal School of Management, published their findings March 7 in Nature Communications.

Ayvaci cautioned against eliminating humans from the equation because AI is not as accurate as a radiologist’s diagnosis.

“The workflow and most cost-effective strategy are the question,” Ayvaci said. “I’m talking about how to replace a task, not replace decision-making. Could we design a workflow where AI is playing a triage role?”

“The workflow and most cost-effective strategy are the question. I’m talking about how to replace a task, not replace decision-making. Could we design a workflow where AI is playing a triage role?”

Dr. Mehmet Ayvaci, associate professor of information systems

From allowing AI to read all mammograms, to integrating the technology to help radiologists with cases, to continuing the current practice of doctors reviewing the images, the researchers determined that AI could effectively be used to delegate concerning results to radiologists. They built their models using actual mammography images from the Digital Mammography DREAM Challenge, a repository of over 640,000 anonymous images available to scientists to solve problems. Another of the study’s co-authors, Dr. Gustavo Stolovitzky, is founder, chair emeritus and currently one of the directors of the DREAM Challenges, a crowdsourcing effort that uses collaborative science to solve translation medicine problems through data analysis.

“Our approach enabled us to precisely identify which patients could safely be evaluated using AI alone while determining which ones would need to be referred to a human,” Mookerjee said. “Every patient’s data, including the mammography images, is input to an AI algorithm, which outputs a measure of risk. You can think of this measure as a probability that the patient has breast cancer. This risk value is then used to determine whether the patient should be further referred to a human.”

Integrating AI in mammogram screenings could help reduce false-positive diagnoses, which carry additional costs and can subject patients to unnecessary procedures such as biopsies, Ayvaci said. Conversely, a false negative result from AI could trigger questions about liability.

“If AI makes a wrong decision, who should be held accountable? All of these are open questions,” he said.

Ayvaci began studying AI in mammography screenings as a doctoral student.

“I worked with a radiologist who specialized in mammography and developed these models to help make better decisions,” he said. “After I joined UTD, I began asking questions that relate to economics and societal impacts of decision-making using AI.”

AI’s effectiveness as a diagnostic tool has not been thoroughly tested in real-world settings, said Ayvaci, who is also interested in investigating ways to help with administrative burdens in health care as well as cost-cutting solutions. How AI can be integrated and adapted to different environments, such as rural settings, also intrigues him.

“The new generation of AI systems are capable of providing context and improving in a much faster timeline,” Ayvaci said. “All of this contributes to a brighter future of using these AI systems.”

For Mookerjee, whose mother is a breast cancer survivor, AI’s applications in health care inspired her participation in the research.

“When Dr. Ayvaci told me about his idea to use AI in screening mammography images, it seemed like the perfect merger of these two areas,” she said.

Ayvaci plans to continue to examine AI applications in health care, and he and Mehmet Eren Ahsen PhD’15, the study’s corresponding author at the University of Illinois Urbana-Champaign, are examining how they can use such a system to help schedule same-day procedures or follow-ups after a diagnosis instead of requiring the patient to wait.

“The newer generation of AI is going to have more influence in business and health care,” Ayvaci predicted. “This is one tiny step, and future research can address what we are trying to improve upon.”

Ultimately, a collaboration between humans and AI may prove most effective in treating patients.

“We shouldn’t take the human out of the loop,” Ayvaci said. “And we should create systems that take the patient into account first, not second.”