Vision-Language Models Encode Clinical Guidelines for Concept-Based Medical Reasoning

A research team from Queen's University and other institutions has introduced MedCBR, a novel AI framework designed to make medical imaging diagnosis both highly accurate and deeply interpretable. The system addresses a critical flaw in existing Concept Bottleneck Models (CBMs), which, while transparent, often ignore broader clinical context like diagnostic guidelines. MedCBR integrates these guidelines directly into its reasoning process. It works by transforming labeled clinical descriptors into guideline-conformant text, then training a model with a multitask objective that performs multimodal contrastive alignment, concept supervision, and diagnostic classification simultaneously. This grounds the AI's understanding in established medical knowledge.

The result is a system that doesn't just output a prediction but generates a structured clinical narrative explaining the diagnosis, effectively emulating expert reasoning. In performance tests, MedCBR demonstrated exceptional accuracy, achieving an Area Under the Receiver Operating Characteristic curve (AUROC) of 94.2% on ultrasound datasets and 84.0% on mammography. Further validation on non-medical datasets showed 86.1% accuracy, proving the generalizability of the concept-based approach. The framework represents a significant step toward reliable, explainable AI in healthcare by creating an end-to-end bridge from raw image analysis to auditable clinical decision-making.