Concept-Enhanced Multimodal RAG: Towards Interpretable and Accurate Radiology Report Generation

Interpretability and accuracy are complementary, not competing goals: The research demonstrates that making AI systems more transparent through interpretable clinical concepts actually improves diagnostic accuracy rather than reducing it, challenging the common assumption that explainability comes at a performance cost.

AI hallucinations in medical imaging are a solvable problem: By combining concept-based explanations with retrieval-augmented generation (RAG), the system significantly reduces instances where the AI generates findings that don't match the actual imaging evidence—a critical safety concern for clinical adoption.

The framework improves performance across different AI models and settings: Testing on multiple Vision-Language Models and training configurations showed consistent improvements in both clinical accuracy metrics and standard performance measures, suggesting the approach is robust and generalizable rather than dependent on specific technical setups.

Modular design enables practical clinical implementation: The system's decomposition into separate components (visual transparency and language model conditioning) makes it easier to understand which part of the AI's reasoning led to specific conclusions, supporting clinician trust and oversight in real-world medical settings.

Reducing documentation burden without sacrificing safety: By automating radiology report generation while maintaining interpretability and factual grounding, this approach addresses a key barrier to AI adoption in healthcare—the need to balance efficiency gains with the clinical accuracy and transparency that physicians require.