AI mammography catches cancer doctors missed but overdiagnosis concerns remain

AI is increasingly being used to analyze mammograms and help radiologists detect breast cancer, with some software programs catching cancers that doctors might have missed. While proponents highlight cases like Deirdre Hall’s, whose Stage 1 cancer was identified only because AI flagged a suspicious area her radiologist initially deemed normal, experts caution that more research is needed to prove AI actually saves lives and doesn’t lead to unnecessary overdiagnosis.

How AI mammography works: Machine learning algorithms are trained on hundreds of thousands or millions of mammogram images to distinguish between malignant and benign tissue patterns.

• Some AI programs identify suspicious areas by drawing circles around concerning spots, while others predict a woman’s likelihood of developing breast cancer.
• The FDA has authorized multiple AI mammography programs with varying accuracy rates.
• Lunit, the software used in Hall’s case, accurately identified cancers 88.6% of the time in a 2024 study of over 8,800 Swedish women.

Why dense breast tissue matters: About 40% of U.S. women have dense breasts, which makes mammograms significantly harder to interpret and increases cancer risk.

• “It’s like trying to find a snowball in a blizzard,” said Dr. Otis Brawley, a professor of oncology and epidemiology at Johns Hopkins University.
• Dense tissue can camouflage cancerous tumors, as happened in Hall’s case where layers of tissue created a particularly complicated crisscrossing pattern.

Major medical centers adopting the technology: Leading institutions including MD Anderson Cancer Center, Mount Sinai Hospital, University of Pennsylvania’s Perelman Center, and others are integrating AI into their mammography workflows.

• FDA regulations require that a doctor must still interpret all mammograms—AI serves as an augmentation tool, not a replacement.
• At UC San Francisco, researchers found AI triage cut the average time from mammogram to biopsy by 87%, from 73 days to nine days for cancer patients.

Cost considerations: Most academic medical centers don’t charge patients extra for basic AI analysis since there’s no specific insurance billing code.

• SimonMed and RadNet offer initial AI screening at no additional cost but charge $40-50 respectively for secondary AI analysis.
• Hall paid $50 for the extra AI screening that ultimately detected her cancer.

Accuracy and limitations: While AI shows promise, it’s not perfect and raises several concerns among experts.

• AI gave false positives 7% of the time in the Swedish study, compared to about 10% for traditional mammograms.
• Dr. Sonja Hughes from Susan G. Komen, a breast cancer organization, said “we’re not there yet” regarding having enough research data for AI to become standard care.
• A $16 million, two-year study at seven medical centers is underway to better evaluate the technology’s effectiveness.

Potential benefits for underserved areas: AI could help address disparities in mammogram interpretation quality.

• Breast imaging specialists correctly identified cancers 89% of the time without AI, compared to 84% for general radiologists.
• With AI assistance, accuracy for both groups rose to about 93%, potentially reducing interpretation disparities.

Concerns about overdiagnosis: Some experts worry AI might detect too many technically cancerous but non-threatening tumors.

• “It’s cancer, but it’s not genetically programmed to grow, spread, or kill,” Brawley explained about some tumors that may not require treatment.
• There’s also concern about potential bias if AI is trained primarily on images from white women, potentially reducing accuracy for women of color.

What they’re saying: Medical professionals emphasize AI’s role as an enhancement tool rather than replacement.

• “The nice thing about AI is that it doesn’t get tired,” said Dr. Lisa Abramson from Mount Sinai. “It’s not going to replace the job or the expertise of radiologists, but I think it’s only going to enhance our ability to detect more and more breast cancers.”
• Hall, despite being generally skeptical of AI technology, said: “I don’t love all this AI stuff, but I definitely love this for me or anyone else in my position.”