Breast biopsy AI study reports fewer unrelated claims in generated reports

Researchers report that CorePath, a breast pathology AI model, outperformed its base model on several biopsy classification tasks and reduced unrelated, non-breast statements in generated reports from 30.1% to 2.8%. They trained it using 7901 paired biopsy images and diagnostic reports from two centers, then evaluated it across six biopsy cohorts and two public benchmarks. A version with statistical filtering withheld reports when predictions did not meet its thresholds and produced zero non-breast hallucinations among the reports it released. The findings are reported in a preprint and describe research evaluations, not use in routine diagnosis.