RadFusion: AI Framework Adds Threshold Control to Radiology Reports
A new AI framework called RadFusion aims to give radiologists control over the sensitivity-specificity trade-off in automated radiology report generation. The system, described in a preprint on arXiv (2608.10505), combines a multi-label classifier that provides per-disease confidence scores with a VQA-based report generator that details medical findings. An LLM then rewrites the report to align stated diagnoses with a chosen threshold. This addresses a critical gap: existing generation models lack the ability to adapt to different clinical scenarios, such as emergency triage (which prioritizes sensitivity) versus confirmatory interpretation (which emphasizes specificity). The framework also supports ROC-based validation, which is often required for regulatory clearance. The research responds to the global shortage of radiologists and the need for more adaptable AI tools in medical imaging.
Key facts
- RadFusion is a framework for threshold-controllable radiology report generation.
- It fuses a multi-label classifier with a VQA-based report generator.
- An LLM rewrites reports to match the desired sensitivity-specificity trade-off.
- The method addresses the lack of control in existing generation models.
- It supports ROC-based validation for regulatory clearance.
- The research is motivated by the shortage of radiologists.
- The preprint is available on arXiv with ID 2608.10505.
- The framework can adapt to emergency triage and confirmatory interpretation scenarios.
Entities
Institutions
- arXiv