Breast ultrasound diagnosis is increasingly aided by artificial intelligence, but most current systems are driven primarily by image features and pay limited attention to the standardized knowledge encoded in the Breast Imaging Reporting and Data System (BI-RADS). A new preprint on arXiv proposes a concept-informed framework that explicitly brings BI-RADS knowledge into the AI pipeline.
The framework is described as interpretable, meaning it aims to align model predictions with clinical concepts rather than relying solely on opaque image patterns. By grounding the model in BI-RADS, the authors hope to make the system more useful to radiologists, who routinely use this reporting standard in practice.
The abstract notes that existing systems have limited incorporation of BI-RADS knowledge, suggesting the new approach is a direct response to that gap. The paper is available as arXiv:2610.16128.