Image & Video

New AI Reads Mammograms and Shows Its Work — But It's Still Early

⚡A step toward AI that assists radiologists, rather than replacing them.

Deep Dive

A researcher has built an experimental AI system called MammoClaw that looks at mammograms — the X-ray images used to screen for breast cancer — and tries to describe what it sees. The unusual part is that it needs no retraining. It plugs into an existing "multimodal" AI (one that can look at pictures and read text at the same time) and simply hands it a set of small, purpose-built tools. The work was accepted at a breast imaging workshop tied to MICCAI 2026, a major medical imaging conference.

The tools mimic what a human radiologist actually does. One zooms into a specific region of interest instead of squinting at the whole image. Another compares multiple views of the same breast, taken from different angles. A third compares the other breast — a classic trick, since breasts are usually roughly symmetrical and differences stand out. On top of that, MammoClaw has a "skill evolution" feature: when it fails a case, that failure gets turned into written guidance it can reuse next time. Think of it as an AI that keeps a growing checklist of its own lessons learned.

The results are honest and mixed. The team tested the system on two real tasks: BI-RADS assessment (the standard 0-to-5 risk rating radiologists assign) and breast density estimation (how much dense tissue a breast has, which affects both cancer risk and how well screening works). Strikingly, the tools by themselves did not reliably make the AI better. The accumulated skills helped in some situations, but not consistently. The paper is upfront that this is an exploratory study, not a finished product.

What may matter most is transparency. MammoClaw lets people inspect how it gathered evidence, which tools it used, and where it went wrong — something most AI systems hide. That makes it auditable, which matters enormously in medicine. The catch: this is early research, not a clinically validated tool, and it is nowhere near replacing a radiologist. But it points toward a future where AI offers a transparent second look at your scan.

Key Points
  • MammoClaw analyzes mammograms without any retraining — it reuses an existing AI model plus small add-on tools
  • It copies radiologist habits: zooming in on spots, comparing views, and checking the other breast
  • Tools alone didn't reliably improve accuracy; learning from failed cases helped somewhat, and the AI shows its reasoning

Why It Matters

Could one day give radiologists a transparent second opinion on mammograms — early research, not clinic-ready.

📬 Get the top 10 AI stories daily