AtomiMed: Hierarchical Atomic Fact-Checking for Universal Clinical-Aware Medical Report Evaluation

AtomiMed proposes a modality-agnostic framework for evaluating medical report generation by decomposing reports into hierarchical Atomic Clinical Facts. The hierarchy covers disease-level entities and attributes such as location, morphology, and severity. An agentic cross-verification loop compares ground-truth and predicted reports to assess diagnostic detection separately from descriptive accuracy. The work also introduces MRGEvalKit, an open-source toolkit for hierarchical extraction, and OmniMRG-Bench, a multimodal benchmark spanning X-ray, CT, MRI, and ultrasound. The abstract reports stronger correlation with radiologist judgments, but provides no detailed numerical results here.