Researchers built an AI system that evaluated 8311 breast ultrasound scans across seven countries

A research team led by Seoul National University and Kookmin University developed an artificial intelligence model called B-RAD to interpret breast ultrasound scans, publishing their multinational validation in npj Digital Medicine.

Interpreting breast ultrasound is notoriously difficult. Radiologists rely on the standard BI-RADS scoring system to assess lesions, but different doctors often reach conflicting conclusions when looking at the same scan. Standard deep learning systems can predict cancer, yet their internal logic remains a black box that clinicians hesitate to trust.

B-RAD tackles this by copying the exact workflow of an experienced radiologist. Instead of jumping directly from an image to a verdict, the system searches an archive of reference cases to find similar examples. It then highlights the suspected lesion, checks the edges and sound reflections, and outlines the area to decide the final category.

The researchers evaluated the framework on 8311 scans collected across eleven patient groups in seven countries. When applied to an independent hospital group without any fine-tuning, B-RAD achieved a biopsy triage AUROC score of 0.952, beating standard vision-language models on the same task.

The real test came in a trial with four human readers. Working with the AI raised each reader's accuracy, brought their overall agreement levels from moderate up to substantial, and lowered the number of overlooked cancers across the board.

Could this sort of step-by-step reasoning help smaller regional clinics handle cancer screening with the consistency of major teaching hospitals?

Despite the strong numbers, the study has boundaries. The reader test evaluated only four practitioners, and the software has not yet been demonstrated in live, real-time scanning rooms during routine patient checkups.