Journal: Nature communications
This study reports the development and validation of BL4AS, an AI system designed to improve MRI-based diagnosis of BI-RADS 4 breast lesions, where false positives and variability are major clinical problems.
Key points:
- Modality and input: Uses dynamic contrast-enhanced breast MRI, exploiting rich spatiotemporal imaging features via foundation-model-based AI.
- Population and data: Trained and evaluated on a multicenter cohort of 2,803 lesions from 2,686 women.
- Diagnostic performance:
- • Achieved AUCs between 0.892 and 0.930 for distinguishing malignant from benign lesions.
- • Showed substantially higher specificity than radiologists (0.889 vs 0.491), indicating a large reduction in false positives at comparable sensitivity.
- Impact on radiologists:
- • When used as an assistive tool, BL4AS improved diagnostic accuracy for both junior and senior readers.
- • Reduced inter-reader variability by 24.5%.
- • Decreased overall false-positive rates by 27.3%.
- Risk stratification:
- • Beyond binary classification, BL4AS assigns lesions to BI-RADS 4A, 4B, and 4C subcategories, enabling more granular risk assessment and potentially better biopsy decision-making.
Clinical implication: An AI system leveraging MRI spatiotemporal data can meaningfully reduce unnecessary biopsies and inter-reader variability in BI-RADS 4 lesions, supporting more precise, standardized breast cancer management.