Clinical-context guidance and hyperbolic prototype learning for long-tailed radiology report generation

Automated radiology report generation (RRG) can reduce the reporting burden, but existing methods remain limited in their ability to integrate clinical context effectively at the feature level. They also struggle with long-tailed disease distributions, which can hinder recognition of infrequent findings. To address these issues, CSMH-RRG, a clinically guided framework that integrates clinical semantic modulation (CSM) and hyperbolic prototype-based classification (HPC), is proposed. The CSM module conditions postencoder visual representations on patient-specific clinical context through semantic injection and channel recalibration. The HPC module represents multiple disease states as learnable prototypes on the Poincaré ball and converts predicted states into disease-state tokens for report generation. Under the primary MIMIC-CXR protocol, CSMH-RRG achieved a BLEU-4 score of 0.262 and a diseasewise macro CheXbert F1 score of 0.564, achieving state-of-the-art performance on clinical factuality metrics while maintaining comparable BLEU-4 performance to the best-performing baseline. When transferred from MIMIC-CXR to IU X-ray without fine tuning, it achieved a BLEU-4 score of 0.118 and a CE F1 score of 0.289. Specifically, in the MIMIC-CXR frequency-stratified classifier analysis, CSMH-RRG increased tail-group macro F1 from 35.5% to 41.4% relative to the linear baseline. Head-group macro F1 also increased from 72.1% to 73.5%. These results indicate that CSMH-RRG supports report generation and improves frequency-sensitive disease-state prediction within the evaluated benchmarks. Code is available at https://github.com/tinkerbellqwq/CSMH-RRG.