DAMEC: Disease-Aware Multi-Expert Consensus Framework for Study-Level Radiology Report Generation
Abstract
Radiology report generation (RRG) translates medical images into clinical reports, yet existing systems diverge from real radiological practice in three respects. First, they are formulated at the image level, consuming a single image or a fixed pair of current/prior images, and discarding the variable visual evidence within a patient study. Second, report generation flows through a single end-to-end pipeline. Even when auxiliary classifiers or retrieved priors are incorporated, they remain auxiliary inputs rather than independent sources of clinical judgment. Third, both retrieval and validation operate without a finding-level clinical anchor: retrieval falls back on image-feature similarity, while validation relies on prompt-level heuristics, leaving generated reports weakly grounded. We propose DAMEC, a Disease-Aware Multi-Expert Consensus framework that recasts RRG as study-level, clinical-context generation. Four heterogeneous experts process a variable number of images from a patient study and emit per-expert clinical signals, which a trainable disease-aware consensus module aggregates into a unified consensus. The consensus, together with disease states, uncertainty, longitudinal changes, and attribute modifiers, is externalized into a structured Clinical Findings (CF) descriptor, serving as the finding-level anchor for both clinical-context retrieval and post-generation validation. On MIMIC-CXR, MIMIC-ABN, Two-view CXR, and CheXpert Plus, DAMEC consistently improves both natural language generation and clinical efficacy metrics over strong baselines. Code is available at: https://github.com/ku-milab/DAMEC.