RadFusion Enables Threshold-Controllable Radiology Report Generation

Researchers have developed RadFusion, a novel framework addressing a critical limitation in automated radiology report generation: the lack of control over the sensitivity-specificity trade-off of diagnostic content. Unlike perception models, existing report generators offer a fixed output, which is unsuitable for diverse clinical scenarios where, for example, emergency triage prioritizes sensitivity while confirmatory interpretation emphasizes specificity. This fixed nature also hinders regulatory clearance, which often requires ROC-based validation. RadFusion tackles this by fusing a multi-label classifier, which provides per-disease confidence scores, with a VQA-based report generator that describes medical findings in detail. An LLM then rewrites the generated report to ensure its stated diagnoses align with the classifier's decisions at a chosen threshold, while remaining grounded in the generator's descriptions. This innovative approach allows the performance of RadFusion to conform to the classifier's ROC curve, making generated reports quantitatively evaluable and clinically adaptable. The framework not only enables operating-point selection to match clinical context but also significantly improves diagnostic accuracy. It showed sensitivity increases by 6.9% at matched specificity and specificity increases by 20.7% at matched sensitivity on the MIMIC-CXR dataset. This makes automated report generation more reliable, adaptable, and verifiable for clinical use.