AI in Breast Imaging: Hype, Hope, or Hazard?
I spend a lot of my professional time reading mammograms. It’s a deeply human task requiring both visual pattern recognition and clinical judgment. And it’s one area where AI is drawing serious attention—and serious investor dollars.
At the recent Society of Breast Imaging annual conference, the AI buzz was less “healthy skepticism” and more “full-blown hype.” Booths were packed. Claims were loud. Evidence? Often harder to find.
One vendor proudly told me their AI could have avoided hundreds of biopsies on calcifications—if only radiologists had “trusted the system”. Intrigued, I asked for the source of this remarkable claim.
As it turned out, this claim stemmed from a single breast radiologist who looked back at her own cases and concluded—informally—that the AI might have helped avoid biopsies. I asked the usual questions:
Was this published?
Peer-reviewed?
Presented anywhere?
No. No. And no.
Then came the kicker: “We probably shouldn’t have told you that.”
Yikes.
Across the show floor, vendor booths flashed bold claims: “Up to 50% reduction in interval cancers!” and “Detect cancers 9 months earlier!” Dig into the fine print, and you’ll often find a single-institution, retrospective study with words like “potential” doing a lot of heavy lifting.
But what really concerns me isn’t the marketing fluff—it’s the growing idea that radiologists can relax on mammograms AI flags as normal.
“Fly through the negatives!” they say.
“Read faster!” they promise.
Let’s ground this in reality. About 90% of screening mammograms are negative. If we start zoning out on those because an AI gave the green light, we’re putting real lives at risk.
A little math:
43 million screening mammograms per year in the U.S.
5 cancers per 1,000 = ~215,000 cancers annually
At 97% AI sensitivity? That’s 6,450 missed cancers
At 99%? Still 2,150 missed
And if radiologists let their guard down on “safe” cases, many of these will slip through
These aren’t just statistics. They’re people—mothers, daughters, sisters.
So, where do we go from here?
AI has tremendous potential. It could help many radiologists, serve as a second reader, and maybe someday even write a clean, coherent report or fix our bloated EMRs. But today? It’s not the revolution we’re being sold.
If AI is going to be a meaningful partner in breast imaging—or in radiology more broadly—it must come with:
Peer-reviewed, reproducible, real-world data—not anecdotes, case studies, single-institution retrospective reviews, or non-peer-reviewed “white papers”
Clear disclosures of limitations—not marketing spin
Grounded expectations—not inflated promises
And radiologists? We need to hold AI to the highest bar—because if we don’t, who will?
The hype is intruiguing. But I’m still looking at every single image.
To read about a sea change in supplemental screening for dense breast tissue at SBI 2025 click here.
