SG-CBM Model Boosts Trustworthy AI for Breast Ultrasound Diagnosis

Concept Bottleneck Models (CBMs) offer interpretable-by-design predictions by routing decisions through human-understandable concepts—but in medical imaging, their trustworthiness often falters because concept activations can latch onto irrelevant regions. The new SG-CBM (Spatially Grounded Concept Bottleneck Model) tackles this head-on by using coarse lesion delineations as weak supervision. For breast ultrasound, it defines two clinically motivated zones from each lesion mask: an in-lesion region for morphology-related concepts and a posterior acoustic band for posterior phenomena. A grouped spatial grounding objective trains concept maps to fire only within these anatomically relevant areas, while a linear bottleneck classifier preserves semantic faithfulness.

Across five-fold cross-validation, SG-CBM improves both diagnostic AUROC and concept macro-AUROC while markedly increasing the spatial alignment of concept evidence. The team also designed a Train-corrupt/Test-clean annotation-quality stress test to reveal how supervision quality affects diagnosis and spatial faithfulness. The results underscore a critical lesson: deployable healthcare AI needs not just interpretability, but spatially faithful explanations that clinicians can trust. Accepted at the IEEE/ACM CHASE 2026 workshop on data-quality-aware AI, this work pushes toward safer, more reliable AI-assisted ultrasound analysis.