Case2Knowledge: Structured Knowledge-Infused Prompting for Faithful Radiology Report Generation

Le résumé fourni par la source

Radiology report generation (RRG) aims to automatically describe imaging studies, reducing radiologists’ workload and improving diagnostic reliability. Despite recent advances, current methods still suffer from hallucinations and semantic inconsistencies, limiting clinical adoption. To address these issues, we present Case2Knowledge (C2K), a retrieval-augmented and knowledge-infused prompting framework for RRG. Specifically, C2K first performs coarse-to-fine retrieval to identify exemplar reports that are both visually and semantically aligned with the query image. The reports are then converted into structured RadGraph-style triplets, filtered by a medical large multimodal model for clinical relevance, correctness, and utility. Finally, these triplets are used to guide the report generation process, producing accurate and faithful descriptions. Experiments on the MIMIC-CXR dataset demonstrate that C2K achieves superior factual accuracy and semantic consistency, highlighting its potential for reliable automated radiology reporting.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.