Learning a Clinically-Relevant Concept Bottleneck for Lesion Detection in Breast Ultrasound

AI explanations use medical standards: The system explains its decisions using BI-RADS, an established clinical lexicon from the American College of Radiology, making AI predictions more meaningful to radiologists rather than presenting unexplainable "black box" outputs.

Radiologists can correct AI in real time: The concept bottleneck design allows doctors to review and manually adjust the AI's intermediate predictions about specific lesion features, which then improves the final cancer classification (performance increased from 0.876 to 0.885 AUC when corrections were made).

Addresses healthcare equity: The research targets breast ultrasound analysis specifically because it's a promising tool for cancer detection in regions with limited access to mammography, potentially extending AI benefits to underserved populations.

Strong performance on real clinical data: The model was trained and tested on 8,854 images from 994 women with expert annotations and confirmed cancer diagnoses, outperforming existing lesion detection systems with 48.9 average precision on held-out test data.

Transparency enables clinical adoption: By making AI reasoning interpretable and modifiable, this approach addresses a key barrier to responsible AI deployment in healthcare—radiologists can understand why the system made a recommendation and maintain professional oversight rather than blindly trusting automated decisions.