Three out of four breast biopsies in the suspicious BI-RADS 4 category come back benign. A retrospective study in Breast Cancer Research and Treatment suggests an FDA-cleared AI tool could shift a meaningful share of those patients from tissue sampling to surveillance.
Fourteen radiologists at two academic medical centers documented 271 lesions in 253 women between March and May 2025, all rated BI-RADS 4 and all biopsied, aspirated or excised. Pathology later found 63 malignant and 208 benign.
Holding the safety line on cancers
The Koios DS decision-support tool flagged 90.5% of malignant lesions as BI-RADS 4, agreeing that sampling was warranted. It missed six cancers, sorting them into BI-RADS 2 or 3, which is precisely why the authors insist any clinical use must preserve timely diagnosis.
On the benign side the split was wider. The algorithm labeled 41.8% of benign lesions BI-RADS 2 and a further 18.8% BI-RADS 3, which put 60.6% into tiers that usually mean follow-up rather than immediate biopsy.
Why this is a signal, not a mandate
Design matters here. This was no prospective trial in which AI output changed care. The radiologists had already made their BI-RADS 4 calls before the tool’s readings were logged, and the analysis asked what might have followed if those readings had driven triage. Enrollment was narrow too, since only lesions that experienced breast imagers at academic centers had flagged as suspicious were included, so the evidence addresses one high-stakes choice rather than screening at large. Women in the study ranged from 24 to 92 years old.
The finding lands as radiology confronts an uncomfortable arithmetic. Catching every possible cancer has meant accepting a high rate of biopsies that find nothing, and that trade-off sits in the path of every AI triage tool entering breast imaging workflows.
