Regina Barzilay, an MIT computer-science professor specializing in natural language processing, faced a stark reality in 2014 when diagnosed with breast cancer. Accustomed to working on the cutting edge of technology, she was shocked by the limitations of the U.S. healthcare system, which struggled to provide even basic data analysis on patient outcomes. Despite extensive electronic health records, she received no clear answers about how patients like her fared at her hospital.
After undergoing chemotherapy and recovering, Barzilay began considering how her expertise in artificial intelligence (AI) could benefit other patients. She realized that her first mammogram contained subtle indicators of cancer risk that went undetected until her third annual scan. Using Mirai, an open-source AI tool she co-developed at MIT, Barzilay was able to identify these early warning signs, illustrating the potential of AI to transform breast-cancer detection and screening strategies.
AI’s Potential to Transform Screening
The traditional approach to mammography often applies a one-size-fits-most model, relying on broad factors such as age and breast density. Annual breast-cancer screenings cost the U.S. approximately $11 billion, yet they fail to account for individual risk variability. Barzilay emphasizes that AI tools like Mirai could help allocate resources more effectively, enabling higher-risk patients to be monitored more closely while reducing unnecessary screenings for lower-risk individuals.
“For most women, current guidelines lump everyone into broad categories, which isn’t helpful,” Barzilay told MarketWatch. “AI allows us to see beyond those simple metrics.”
How Mirai Works
Mirai predicts breast-cancer risk using only mammogram images, identifying subtle tissue changes invisible to the human eye. Unlike traditional models, which combine factors such as age, family history, and lifestyle, Mirai achieves approximately 80% accuracy, outperforming conventional methods while maintaining consistency across diverse populations. Radiologist Hari Trivedi notes that this “fundamental paradigm shift” allows AI to provide reliable, reproducible predictions using just four images, bypassing inconsistencies in patient-reported histories.
Validation studies have confirmed that Mirai generalizes well beyond its original training dataset, outperforming traditional risk models for minority populations. Experts, however, caution that rigorous clinical evaluation remains essential to confirm real-world benefits.
Challenges and Adoption Barriers
Despite promising results, integrating AI into routine healthcare faces obstacles. Infrastructure costs, limited access for community hospitals, and the complexities of the U.S. fee-for-service model could slow widespread adoption. Some patients may resist reduced screening schedules, while insurers and healthcare providers must navigate financial incentives that may conflict with AI-driven approaches.
Nonetheless, AI-based risk assessment could lead to more personalized mammogram schedules, with high-risk patients screened more frequently and lower-risk patients monitored less often. Even a modest reduction of 5–10% in unnecessary screenings could significantly reduce healthcare costs given that 40 million women receive mammograms annually.
Beyond Breast Cancer
Barzilay has also extended her AI work to lung-cancer detection through the Sybil model, which predicts long-term risk using imaging data. By identifying patients at high risk who do not meet traditional screening criteria, Sybil aims to improve early detection and optimize screening intervals, paralleling strategies used in colon cancer prevention.
Barzilay’s vision extends to broader healthcare applications, including AI-assisted drug discovery and modernizing clinical-trial standards. “We’re not comparing against perfection,” she says. “We’re measuring against the current clinical system, which leaves too many patients at risk.”
Related topics
