Spatially Grounded Concept Bottleneck Models for Trustworthy Breast Ultrasound Diagnosis

Concept Bottleneck Models can improve AI trustworthiness in medical imaging by making diagnoses interpretable: Rather than providing a "black box" prediction, these models explain their decisions through human-understandable medical concepts, which is particularly important for clinical adoption and accountability in healthcare settings.

Spatial grounding—ensuring AI explanations point to the right anatomical regions—is critical for trustworthy AI: The paper found that without proper spatial constraints, concept activations could be "driven by irrelevant regions," meaning the model might identify the correct diagnosis for the wrong reasons. By using lesion delineations as guidance, the researchers improved "spatial alignment of concept evidence," making explanations more clinically faithful.

Data quality directly impacts both diagnostic accuracy and trustworthiness: The annotation-quality stress test revealed that supervision quality affects not just diagnosis performance but also the spatial faithfulness of explanations. This suggests that responsible AI deployment requires careful attention to training data quality, not just model architecture.

Clinical domain knowledge should inform AI system design: The researchers incorporated two clinically motivated anatomical zones (in-lesion regions for morphology and posterior acoustic bands for acoustic phenomena) into their model, demonstrating that trustworthy healthcare AI benefits from embedding expert medical understanding into the system's structure rather than treating it as a generic prediction problem.